LIGN 6 - Syntactic Parsing

### Remember that your Dean of Student Affairs is here to help you!

<http://savethevowels.org/resources>

---

# Automatic Syntactic Parsing
### Will Styler - LIGN 6

---
			  
### Today's Plan

- Constituency vs. Dependency Parsing

- What is dependency parsing?

- How do we do dependency parsing automatically?

- Why do we use dependency parsing in NLP?

- How is syntactic parsing *actually useful* for natural language processing?

---

### We've talked about trees in terms of constituency

- A hierarchy of phrases, containing other phrases

- From this, we can get a sense of constituency

- ... as well as of attachment

---

---

### These trees come from Phrase Structure Grammars

- Also called 'Context Free Grammars' by computer scientists

- A series of rules applied which detail grammatical and ungrammatical sentences

- With other rules describing changes (e.g. movement, etc)

---

### Linguists use constituency trees for linguistic analysis

- For asking linguistic questions, constituency is *crucial*

- It shows the units on which syntax operates

- It's better to *generate* new sentences

- It handles *recursion*

---

### Linguistic Syntax is done using constituency

- ... and for a long time so was most NLP

---

### ... but there is another way

---

### Syntax helps tell us which words 'belong' to other words

- "The cat jumped onto the table"

- "I poked the big bear with a stick"

- "I touched the Queen of England's hat"

---

### Syntax also marks who did what to whom

- "Jony chased Dieter"

- "Jessica ate the veggie burger"

- "Felicia gave Juan a drink"

- "My aunt bought three cars and set them on fire"

---

### What if we could just capture which words are 'connected' to other words

- Describe the nature of that connection (who did what to whom)

- ... and skip some of the phrase structure questions

---
<img class="r-stretch" src="comp/not_from_syntactician.jpg" alt="notfromsyntactician.jpg The image is a split-screen meme featuring two characters from what appears to be a science fiction or fantasy setting. The top half shows a younger male with curly brown hair and a serious expression. He is wearing dark clothing that suggests he might be part of an organization or group, possibly one associated with the darker side in this fictional universe. The bottom half features an older man with white hair and a stern look on his face. His attire also appears to align with the same type of dark, formal wear as seen in the top image. The background for both characters is dimly lit, which could indicate that they are indoors or in a setting where light is controlled. The text at the bottom of the image reads: NOT FROM A SYNTACTICIAN... This phrase suggests a humorous or ironic context, possibly referencing someone's expertise or lack thereof in understanding complex language structures. The characters' expressions and attire might imply a conversation about learning or mastering some powerful skill that is not accessible to those who do not possess certain qualifications. The meme format uses the juxtaposition of these two characters to create humor by implying that one character is seeking knowledge from another, but the other character dismisses this idea with the phrase NOT FROM A SYNTACTICIAN... This description was generated automatically from image files by a local LLM, and thus, may not be fully accurate. Please feel free to ask questions if you have further questions about the nature of the image or its meaning within the presentation.">

---

# Dependency Parsing

---

### Dependency Trees

- Focus on demonstrating the relationship between elements in the sentence

- Dependency ('what's attached to what')

- Relations ('Who's doing what to whom')

- Visualizes this as a series of relationships

---

### The Angry Squirrel stole the Mixed Nuts from Walmart

---

### The Angry Squirrel stole the Mixed Nuts from Walmart

<img class="r-stretch" src="diagrams/dep_angrysqurrel.jpg" alt="depangrysqurrel.jpg The user wants a detailed description of an image showing a syntactic dependency parse tree for a sentence. 1. Analyze the Image: Content: It's a linguistic diagram (dependency grammar). Sentence: The angry squirrel stole the mixed nuts from Walmart . Structure: Words are arranged horizontally at the bottom. Arrows connect words to indicate grammatical relationships. Labels on the arrows describe the relationship type. Key Components: Words (Tokens): The, angry, squirrel, stole, the, mixed, nuts, from, Walmart, . Labels (Dependent Relations): DET, AMOD, NSUBJ, ROOT, NMOD, DOBJ, CASE. 2. Breakdown of Relationships: stole: This is the central verb. An arrow points down to it labeled ROOT. It connects everything else. Subject side (Left): squirrel: Connected to stole with an arrow pointing left/up labeled NSUBJ (Nominal Subject). angry: Connected to squirrel with an arrow pointing left/down labeled AMOD (Adjectival Modifier). The: Connected to squirrel with a long curved arrow over angry labeled DET (Determiner). Wait, looking closely at the arrows. Actually, let's trace carefully. stole -> squirrel (NSUBJ) - Arrow goes from stole up and left to squirrel. No, wait. The arrow head is on the word it modifies? Or does it point to the head? In standard dependency trees like this (likely Stanford Dependencies or similar), arrows usually point from the dependent to the head, OR from the head to the dependent depending on the convention. Let's look at the arrowheads. The arrow from stole goes UP to ROOT. That's a label for the whole tree. The arrow labeled NSUBJ connects squirrel and stole. The arrowhead points TO squirrel. So stole is the head, squirrel is the dependent. Wait, usually NSUBJ means Nominal Subject. So squirrel is the subject of stole. Let's re-evaluate standard dependency parsing visualizations (like Stanford Parser). Usually, arrows point from Head to Dependent or Dependent to Head. Let's look at DET connecting The and squirrel. The arrow starts near squirrel and goes up/left to The? No, the arrowhead is on The. So squirrel -> The (DET). This means The depends on squirrel. Let's look at AMOD connecting angry and squirrel. Arrowhead points to angry. So squirrel -> angry (AMOD). Let's look at the big arc labeled NMOD. It connects stole and Walmart. Wait, no. The label is high up. Let's trace the lines. There is a long arc labeled NMOD connecting stole (implied) to Walmart? No, let's look at the arrowheads. Actually, looking very closely at the image provided: ROOT: A vertical line goes from the word stole UP to the label ROOT. This indicates stole is the root of the sentence. NSUBJ: An arc connects squirrel and stole. The arrowhead points TO squirrel. So stole governs squirrel. Wait, usually arrows point to the dependent. Let's assume standard direction: Head -> Dependent. If stole is root/head. Arrow labeled NSUBJ goes from stole (implied source) to squirrel. So squirrel is the nominal subject of stole. DET: An arc connects The and squirrel. Arrowhead points TO The. Wait, that would mean squirrel modifies The? No. In dependency trees, usually the arrow points to the dependent (the modifier). Let's look at the label DET above an arc connecting The and squirrel. The arrowhead is on The. This implies squirrel -> The (DET). That makes sense. The depends on squirrel. AMOD: An arc connects angry and squirrel. Arrowhead points to angry. So squirrel -> angry (AMOD). This means angry modifies squirrel. DOBJ: An arc connects stole (implied) to nuts. Wait, let's trace the lines from stole. There is a line going right from stole. It splits? No. Let's look at the cluster on the right. There is an arc labeled DOBJ connecting stole and nuts. Arrowhead points to nuts. So stole -> nuts (Direct Object). DET: An arc connects the and nuts. Arrowhead points to the. So nuts -> the (Determiner). AMOD: An arc connects mixed and nuts. Arrowhead points to mixed. So nuts -> mixed (Adjectival Modifier). NMOD: There is a very long, high arc labeled NMOD. It seems to connect stole (or the whole VP) to Walmart? Or maybe nuts to Walmart? Let's trace it. The line starts near stole (actually, looking closely, it seems to originate from the same point as the DOBJ arc, which is effectively stole). It goes all the way right and down to Walmart. Arrowhead points to Walmart. So stole -> Walmart (NMOD - Nominal Modifier? Or maybe a prepositional modifier). Wait, usually NMOD connects nouns. Let's look at the other arc. CASE: There is an arc connecting from and Walmart. Arrowhead points to from. So Walmart -> from (Case marker). This implies from depends on Walmart. Correction on NMOD: Let's look really closely at the top right. There is a label NMOD. Below it is an arc. The arc connects... where? It seems to connect stole (the root) to Walmart? Or maybe nuts to Walmart? Actually, looking at the standard Stanford Dependencies for this sentence: The angry squirrel stole the mixed nuts from Walmart. Standard parse: nsubj(stole, squirrel), det(squirrel, The), amod(squirrel, angry), dobj(stole, nuts), det(nuts, the), amod(nuts, mixed), prepfrom(stole, Walmart) or case(Walmart, from). In this specific diagram: The label NMOD is above a long arc. This arc connects stole (the root) to Walmart. Wait, usually prepositional phrases modify the verb. If it's labeled NMOD, maybe it treats Walmart as a noun modifying the event? Or perhaps it's connecting nuts to Walmart? Let's trace the line for NMOD carefully. It starts above stole (or near the DOBJ origin) and goes all the way right, curving down to point at Walmart. So stole -> Walmart via NMOD? That seems odd for a prepositional phrase. Let's look at the label CASE. It is above an arc connecting from and Walmart. The arrow points to from. This means from depends on Walmart. This is consistent with Universal Dependencies where the case marker depends on the noun (e.g., case(Walmart, from)). So, if from depends on Walmart, and there is a connection to the main verb... Let's re-examine the NMOD arc. It seems to connect stole to Walmart. Wait, no. Look at the arrowheads. The NMOD arc has an arrowhead pointing to Walmart. Where does it start? It starts near the root stole. So nmod(stole, Walmart)? That would mean Walmart modifies stole directly without a preposition? But there is a word from. Let's look at the structure again. Maybe the NMOD arc connects nuts to Walmart? No, that doesn't make sense syntactically (nuts from Walmart). Actually, let's look at the arrowheads for the right side group. There is an arc labeled DOBJ. It points to nuts. Origin: stole. So dobj(stole, nuts). Correct. There is a long arc labeled NMOD. It seems to originate from stole (or the same point as DOBJ) and goes to Walmart. Wait, if there is a preposition from, usually it's nmod:from or similar. Here it is just labeled NMOD. There is an arc labeled CASE. It connects from and Walmart. Arrow points to from. So case(Walmart, from). This means from is the dependent of Walmart. Wait, if from depends on Walmart, how does it connect to the rest? Usually, the preposition connects to the verb. Let's look at the diagram again very carefully. The arc labeled NMOD connects... actually, looking at the curvature, it looks like it connects stole (the root) to Walmart. But that ignores from. Alternative interpretation: Maybe the label NMOD applies to the connection between nuts and Walmart? No, mixed nuts from Walmart. The phrase modifies nuts. So nmod(nuts, Walmart)? Or nmod(Walmart, nuts)? Let's look at the arrowheads again. Arc labeled NMOD: Starts high up, curves down to point at Walmart. Where does it start? It seems to start from the same vertical line as the DOBJ arc, which comes from stole. So nmod(stole, Walmart)? That's weird with from there. Let's look at the arc labeled CASE. It connects from and Walmart. Arrow points to from. This implies case(Walmart, from). In Universal Dependencies, this means from is a case marker for Walmart. But usually the preposition depends on the noun it governs? No, in UD, case(noun, prep) or case(prep, noun)? Actually, in UD, case(Walmart, from) means from is the dependent of Walmart? No, that's not right. In UD, case relation is usually prep -> case. Wait. Let's check UD guidelines. case: The head of a prepositional phrase (the noun) has a dependent which is the preposition? Or vice versa? Actually, in UD: nmod(nuts, Walmart) and case(Walmart, from). This means Walmart modifies nuts, and from marks the case of Walmart. So from depends on Walmart? No. Let's look at the arrow direction in the image for CASE. The arrow points TO from. So Walmart -> from (CASE). This means from is a dependent of Walmart. This matches UD where case(Walmart, from)? No, usually it's prep(nuts, from) and case(from, Walmart). Let's look at the arrow for CASE again. It connects from and Walmart. The arrowhead is on from. So Walmart -> from. This implies from depends on Walmart. Now let's look at the NMOD arc again. It connects... actually, looking really closely at the right side. There are three arcs originating from the area of stole or nuts? Arc 1 (lowest): Connects the and nuts. Label: DET. Arrow points to the. So det(nuts, the). Arc 2 (middle): Connects mixed and nuts. Label: AMOD. Arrow points to mixed. So amod(nuts, mixed). Arc 3 (highest in this cluster): Connects... wait. Let's trace the lines from the right side back to the left. There is a line labeled DOBJ connecting stole and nuts. Arrow points to nuts. So dobj(stole, nuts). There is a very long arc labeled NMOD. It connects... it looks like it connects stole (the root) to Walmart? Or does it connect nuts to Walmart? Let's look at the arrowhead for NMOD. It points to Walmart. Where does it start? It seems to start from the same point as the DOBJ arc, which is effectively stole. So nmod(stole, Walmart)? That would imply Walmart modifies stole. But there is a preposition from. Maybe the label NMOD is actually for the connection between nuts and Walmart? If so, where does the line start? It seems to start from above nuts. Let's look at the intersection of lines. There is a vertical line dropping from ROOT to stole. From stole, an arc goes left/up to squirrel (NSUBJ). From stole, arcs go right. One arc labeled DOBJ goes to nuts. One long arc labeled NMOD goes to Walmart. Wait, if it's NMOD, and there is a CASE relation... Let's reconsider the standard parse for this sentence in tools like Stanford Parser. Sentence: The angry squirrel stole the mixed nuts from Walmart . Parse: nsubj(stole, squirrel), det(squirrel, The), amod(squirrel, angry), dobj(stole, nuts), det(nuts, the), amod(nnuts, mixed), prepfrom(stole, Walmart). In this diagram, prepfrom is likely represented by the combination of NMOD and CASE. The arc labeled NMOD connects stole to Walmart. This indicates that Walmart This description was generated automatically. Please feel free to ask questions if you have further questions about the nature of the image or its meaning within the presentation.">

---

### Dependency Trees mark the connections between words

- Words are connected to the words that 'depend' on them

- Every word is connected to another

---

### Dependency Trees often mark the types of connections

- 'Subject' - The 'do-er' of the verb

- 'Direct Object' - The 'target' of the verb

- 'Indirect Object' - Often used for recipients

- Modifiers - Words that make the meaning of each word more specific

---

### These 'relations' have meaning that doesn't change with order

---

<img class="r-stretch" src="diagrams/dep_bakedcake.jpg" alt="depbakedcake.jpg The user wants a detailed description of an image containing a syntactic dependency tree diagram. 1. Analyze the Image: Type: It's a linguistic diagram, specifically a dependency parse tree for a sentence. Text content (bottom row): Maria baked Isabel a big cake Words: Maria, baked, Isabel, a, big, cake. Labels/Arrows (Top/Middle): There are curved and straight lines connecting the words to labels above them. ROOT: A vertical line pointing down to baked. This indicates baked is the main verb/root of the sentence. NSUBJ (Nominal Subject): An arc from baked curving left and down to Maria. This means Maria is the subject who performed the action. IOBJ (Indirect Object): An arc from baked curving right and down to Isabel. This means Isabel is the recipient of the cake. DOBJ (Direct Object): A long, high arching line starting near baked (or rather, originating from the verb phrase structure) and ending at cake. Wait, let's look closer. The label DOBJ is on a long arc that connects baked to cake. Actually, looking at standard dependency parsing (like Stanford Dependencies), the root usually connects to the direct object. Here, there is a very long arc labeled DOBJ connecting baked to cake. DET (Determiner): An arc from cake curving left and down to a. This means a modifies cake. AMOD (Adjectival Modifier): An arc from cake curving left and down to big. Wait, let's re-examine the connections. The label DET is above an arc connecting cake to a. The label AMOD is above an arc connecting cake to big. Let's trace carefully: ROOT -> points down to baked. NSUBJ (arc) connects baked to Maria. IOBJ (arc) connects baked to Isabel. DOBJ (long arc) connects baked to cake. DET (arc) connects cake to a. AMOD (arc) connects cake to big. 2. Structure the Description: Start with a general overview: It's a black and white diagram showing the grammatical structure of a sentence. Identify the sentence at the bottom: Maria baked Isabel a big cake. Describe the central node: The word baked is in the center, acting as the root. Describe the connections (dependencies) one by one, moving generally left to right or by relationship type. Subject connection (NSUBJ). Indirect Object connection (IOBJ). Direct Object connection (DOBJ). Modifiers of the object (DET and AMOD). 3. Drafting the Content: Sentence: The bottom row displays the sentence Maria baked Isabel a big cake in a serif font. Root: Above the word baked, there is a vertical line labeled ROOT pointing downwards to it, indicating that this verb is the central element of the sentence structure. Subject (NSUBJ): To the left, an arc connects Maria back to baked. The label above this arc is NSUBJ, standing for nominal subject. This indicates Maria is the one doing the baking. Indirect Object (IOBJ): To the right of baked, a shorter arc connects Isabel back to baked. The label above it is IOBJ, indicating Isabel is the indirect object (the recipient). Direct Object (DOBJ): A long, high arching line spans from near baked all the way over to the word cake. The label on top of this long arc is DOBJ, identifying cake as the direct object. Modifiers: There are two smaller arcs connecting words to cake. One arc connects a back to cake with the label DET (determiner) above it. Another arc connects big back to cake with the label AMOD (adjectival modifier) above it. 4. Refine for Clarity and Tone: Ensure the distinction between the words on the bottom line and the grammatical labels above is clear. Use directional language (left, right, above, below). Avoid markdown as requested. 5. Final Polish (Self-Correction during drafting): Make sure to explicitly state what the acronyms likely stand for if possible, or just describe them as labels. The prompt asks for accuracy. I will describe the visual layout primarily. Check the DOBJ connection again. In some dependency parsers, the direct object connects to the root (verb). Yes, the long arc goes from the verb area to cake. Wait, looking really closely at the image... The line for DOBJ starts near the right side of baked and goes all the way to cake. Actually, let's look at the standard Stanford Dependencies. Usually, nsubj connects subject to verb. iobj connects indirect object to verb. dobj connects direct object to verb. In this specific diagram: The line for DOBJ seems to originate from baked (or near it) and go to cake. Wait, let's look at the arcs for DET and AMOD. They originate from cake and go left to a and big. This is a bit unusual. Usually, modifiers point to the head. Here, DET points from cake to a? No, the arrowheads are on the bottom words. Let's look at the arrowheads. NSUBJ: Arrow points to Maria. (Head is baked). IOBJ: Arrow points to Isabel. (Head is baked). DOBJ: Arrow points to cake. (Head is baked). DET: Arrow points to a. Head is... wait. The arc labeled DET connects cake and a. The arrowhead is on a. This implies cake is the head of a. AMOD: Arrow points to big. Head is... cake? Or big? The arc labeled AMOD connects cake and big. Wait, looking at the intersection. Let's re-evaluate the arcs on the right side. There is an arc labeled DET connecting a and cake. The arrow points to a. This means cake governs a. There is an arc labeled AMOD connecting big and... wait, it looks like it connects big and cake? Or big and a? No, looking at the layout: DET is above an arc connecting a and cake. Arrow points to a. AMOD is above an arc connecting big and... actually, it looks like it connects big and cake as well? Or maybe big and a? No, that doesn't make sense. Let's look at the standard dependency tree for Maria baked Isabel a big cake. Root: baked. nsubj(baked, Maria). iobj(baked, Isabel). dobj(baked, cake). det(cake, a). amod(cake, big). Okay, looking at the diagram again to match this theory. The long arc labeled DOBJ connects baked (implicitly) to cake. Arrow points to cake. Correct. The arc labeled DET connects cake to a. Arrow points to a. Correct. The arc labeled AMOD... wait, the label AMOD is above an arc that connects big to... where? It looks like it connects big to cake as well? Or does it connect big to a? No, grammatically big modifies cake. Let's look at the arrowheads on the right side carefully. There is an arc labeled DET. It goes from cake (right) to a (left). Arrow points to a. There is an arc labeled AMOD. It goes from... it looks like it starts near big and goes to the right? No, that's not right either. Let's look at the intersection of lines. Actually, looking very closely at crop 3: There is a label DET above an arc connecting a and cake. Arrow points to a. There is a label AMOD above an arc connecting big and... it looks like the line goes from big up and right, crossing under the DOBJ line? No. Let's trace the lines from the words up. From cake: One line goes left (DET) to a. Another line goes left/up (AMOD)? No. Let's look at the label AMOD. It is above an arc that connects big and... it looks like it connects to the word cake as well? Or does it connect big to a? Actually, in many dependency parsers (like Stanford), amod connects the adjective to the noun. So amod(cake, big). The arrow would point from cake to big. In this diagram: There is an arc labeled AMOD. It seems to connect big and... wait, looking at the rightmost part of the bottom row. Let's look at the full image again. From cake, a line goes left to a (labeled DET). From cake, a line goes left to big? No, that would cross the DET line. Let's look at the arc labeled AMOD. It connects big and... it looks like it connects to the word cake but the arrow points to big. Wait, no. Let's look at the arc labeled AMOD again. It is an arc connecting big and... actually, looking really closely, the line for AMOD seems to connect big and cake? No, that would be weird if DET connects a and cake. Let's reconsider the standard tree structure. Maybe the diagram shows: det(cake, a) and amod(cake, big). Visually: There is an arc labeled DET connecting a and cake. Arrow points to a. There is an arc labeled AMOD. It connects big and... it looks like it connects to the word cake as well? Or does it connect big to a? No, big a cake isn't right. Let's look at the arrowheads again. Arrow for DET points to a. Arrow for AMOD points to big. Where do these lines originate? They seem to originate from cake. So, cake is the head for both a and big. This makes sense grammatically. Cake governs its determiner and its modifier. Final Plan: 1. State it's a dependency parse tree. 2. List the sentence words at the bottom: Maria, baked, Isabel, a, big, cake. 3. Describe the central vertical line labeled ROOT pointing to baked. 4. Describe the arc labeled NSUBJ connecting Maria to baked. 5. Describe the arc labeled IOBJ connecting Isabel to baked. 6. Describe the long arc labeled DOBJ connecting cake to baked (or originating near it). Correction: Looking at the diagram, the DOBJ line actually seems to originate from the verb area and go to cake. Wait, let's look at the arrowhead on the DOBJ line. It points to cake. 7. Describe the arcs connecting modifiers to cake. DET connects a to cake (arrow to a). AMOD connects big to... wait, looking really closely at crop 3 and 6. The label AMOD is above an arc. The arc connects big and... it looks like it connects to the word cake but the line goes under the DET line? No. Let's look at the lines emerging from cake. One line goes left, labeled DET, pointing to a. One line goes left/up... wait. Actually, looking at the label AMOD, it is above an arc that connects big and... it looks like it connects to the word cake but the arrow points to big. Wait, let's look at the intersection of lines near big. There is a line from big going up/right. And a line from a going up/right? No. Let's assume standard dependency grammar: amod(cake, big). The arrow points to the modifier (big). The head is cake. So, there must be a line connecting cake and big. In the image, there is an arc labeled AMOD. It connects big and... it looks like it connects to the word cake but the line is drawn such that it sits above the DET line? No, the DET line is higher This description was generated automatically. Please feel free to ask questions if you have further questions about the nature of the image or its meaning within the presentation.">

---

<small> From Jurafsky and Martin 'Speech and Language Processing' 2018</small>

---

[Here's a link to the full guidelines](https://universaldependencies.org/guidelines.html)

---

### Any sentence that can be constituency parsed, can be dependency parsed

---

### Any sentence that can be constituency parsed, can be dependency parsed

<img class="r-stretch" src="diagrams/dep_compling2.jpg" alt="depcompling2.jpg The user wants a detailed description of an image showing a syntactic dependency tree for a sentence. 1. Analyze the Image: Content: It's a linguistic diagram known as a dependency parse or dependency graph. Text (Bottom Row): The words are Two computational linguists from San Diego walked into a bar. This is a classic joke setup, though the punchline isn't there. Wait, looking closely at the image provided, it ends with bar. It's just the first part of the sentence. Structure: Arrows connect words to other words, indicating grammatical relationships. Above each arrow is a label (e.g., NUMMOD, AMOD, NMOD). Root: There is a vertical line pointing down from ROOT to the word walked. This indicates that walked is the main verb or head of the sentence. 2. Breakdown by Word and Relationships: Two: Connected to linguists via an arrow labeled NUMMOD (Numeral Modifier). The arrow curves from Two down to linguists. Wait, looking at the arrows carefully: There is a long arc starting above Two going down to linguists labeled NUMMOD. Actually, let's re-examine. The label NUMMOD is on an arrow pointing from Two to linguists. No, that's not right either. Let's look at the standard dependency parsing style (like Stanford Dependencies). Usually, modifiers point to the head. Let's trace the arrows carefully. Arrow 1: Starts above Two, curves left and down to Two. Wait, no. The arrowhead is at Two. So linguists points to Two? No, that would be weird for NUMMOD. Usually, NUMMOD modifies a noun. Let's look at the direction. Actually, looking at the standard Stanford Parser output style: Two has an arrow pointing to it from linguists. The label is NUMMOD. This means Two is a numeral modifier of linguists. Wait, usually the arrow points from the head to the dependent or vice versa depending on the convention. In Stanford dependencies, arrows point from the head (noun/verb) to the dependent (modifier). Let's re-evaluate based on visual flow. There is a long arc labeled NUMMOD. It starts above Two and ends with an arrowhead pointing at Two. This implies linguists is the head and Two is the modifier. Wait, looking closer at the left side. Actually, let's look at the labels relative to the words. NUMMOD is above a long arc connecting Two and linguists. The arrowhead points to Two. This suggests linguists modifies Two? No, that makes no sense linguistically. Two modifies linguists. In some visualizations (like spaCy or others), arrows point from head to dependent. In others, dependent to head. Let's look at the label AMOD. It connects computational and linguists. The arrow points to computational. This suggests linguists is the head. So, the arrows point from the Head (noun/verb) to the Dependent (modifier). Let's verify: linguists -> Two (NUMMOD). Yes, linguists are modified by Two. linguists -> computational (AMOD). Yes, computational modifies linguists. Let's re-read the diagram carefully. There is a word linguists. Above it, there are arcs coming from somewhere? No, the arcs originate from the words themselves or connect them. Let's look at the arrowheads. Arrowhead on Two: Points down to Two. Label: NUMMOD. Source seems to be linguists (via a long arc). So linguists -> Two. Arrowhead on computational: Points down/left to computational. Label: AMOD. Source seems to be linguists. So linguists -> computational. Wait, let's look at the arc labeled NMOD. It connects linguists and from San Diego. The arrow points to San Diego? No, let's trace it. There is an arc labeled NSUBJ (Nominal Subject). It connects walked (the root) to linguists. The arrow points to linguists. So walked -> linguists. This makes sense: Linguists walked. There is an arc labeled NMOD (Noun Modifier). It connects linguists and from San Diego. Wait, let's look at the cluster above from San Diego. There is a label CASE connecting from and San Diego. Arrow points to San Diego? Or from? Let's assume standard dependency: Preposition from modifies San Diego (or vice versa). Actually, usually from is the head of the prepositional phrase. Let's look at the specific connections for from San Diego. There is an arc labeled NMOD connecting linguists and San Diego? No, that looks like it connects linguists to the whole phrase. Actually, let's look at the label CASE. It's above an arc between from and San Diego. The arrow points to San Diego. This implies from modifies San Diego? Or San Diego is the head? In Stanford dependencies, prepositions are often heads. So San Diego -> from (case). Let's look at the arc labeled NMOD. It connects linguists to San Diego. The arrow points to San Diego. This implies San Diego modifies linguists. This makes sense: Linguists [who are] from San Diego. Let's try a different interpretation of the arrows. Maybe the arrows point from Dependent to Head? If Two -> linguists: Arrow points to linguists. Label NUMMOD. This fits perfectly. Two modifies Linguists. Let's check the image again. The arrowhead is clearly at the bottom, pointing at the word Two. Wait. No. Let's look really closely at crop 1 and 4. Crop 1: Shows Two, computational, linguists. There is a long arc labeled NUMMOD. It starts above Two and ends with an arrowhead pointing to Two. Wait, no. The arrowhead is on the left side of the arc, pointing down at Two. Actually, looking at the full image, the arrows seem to originate from the right (the head) and point left/down to the modifier? Or vice versa? Let's look at ROOT -> walked. The line goes straight down. walked is the root. NSUBJ arc: Connects walked and linguists. Arrow points to linguists. This means walked (verb) has a subject linguists. So Head -> Dependent direction? No, if walked is the head, and it points to linguists, then linguists is the dependent. But NSUBJ usually marks the relationship between verb and noun. Let's look at the standard Stanford Dependency visualization. In that style, arrows point from Head to Dependent. Root -> walked (Head). walked -> linguists (Subject/NSUBJ). Arrow points to linguists. Correct. linguists -> Two (NUMMOD). Arrow points to Two. Correct. linguists -> computational (AMOD). Arrow points to computational. Correct. linguists -> San Diego (NMOD - Noun Modifier). Wait, the arc labeled NMOD connects linguists and San Diego? Or from? Let's look at the cluster above from San Diego. There is an arc labeled CASE. It connects from and San Diego. Arrow points to San Diego. This implies from modifies San Diego? No, usually case marks the preposition. So San Diego (noun) -> from (preposition). If arrow points to San Diego, then from is the head? Let's look at the arc labeled NMOD. It connects linguists and... where? It seems to connect linguists to San Diego. Arrow points to San Diego. This implies San Diego modifies linguists. Alternative Interpretation (Arrows point Dependent -> Head): If arrows point to the head: Two -> linguists (NUMMOD). Arrow points to linguists. Let's look at the image again. The arrowheads are definitely pointing at the words on the left/bottom in some cases and at the words on the right in others? Let's trace the line from NUMMOD. It starts above Two (left side) and goes down to... wait. Actually, looking at the very first crop, there is an arrow pointing DOWN to Two. The label is NUMMOD. This implies something points TO Two. What? The arc seems to come from linguists. So linguists -> Two. This fits Head->Dependent. Let's trace the line from AMOD. It starts above computational and goes down to... wait, there is an arrow pointing DOWN to computational. Label AMOD. Source seems to be linguists. So linguists -> computational. Fits Head->Dependent. Let's trace the line from NSUBJ. It starts above walked (right side) and goes all the way left to... Two? No, that's too long. Let's look at the arc labeled NSUBJ. It connects walked and linguists. The arrow points to linguists. So walked -> linguists. Fits Head->Dependent (Verb -> Subject). Let's trace the line from NMOD (the one above from). It connects linguists and... San Diego? Or from? The arrow points to San Diego. So linguists -> San Diego. This is weird. Usually, prepositional phrases modify nouns. Let's look at the label CASE. It connects from and San Diego. Arrow points to San Diego. This implies from modifies San Diego? No, usually case marks the preposition as a dependent of the noun. So Noun -> Preposition. If arrow points to San Diego, then from is the head? Let's look at the right side: into a bar. Arc labeled NMOD. Connects walked and bar? Or into? The arrow points to bar. So walked -> bar. This implies bar modifies walked (direction). Arc labeled CASE. Connects into and a bar? No, connects into and bar. Arrow points to bar. Wait. Let's look at the arc labeled DET. Connects a and bar. Arrow points to bar. So a modifies bar. Head -> Dependent. Conclusion on Direction: The arrows point from the HEAD (the main word) to the DEPENDENT (the modifier). Root: walked. walked -> linguists (NSUBJ - Nominal Subject). linguists -> Two (NUMMOD - Numeral Modifier). linguists -> computational (AMOD - Adjectival Modifier). linguists -> San Diego (NMOD - Noun Modifier? Or is it the whole phrase?). Actually, looking closely at the arc labeled NMOD above from, it seems to connect linguists and San Diego. Wait, no. Let's look really closely at the connections for from San Diego. There is an arc labeled CASE connecting from and San Diego. The arrow points to San Diego. This implies from modifies San Diego? No, that's backwards. In Stanford dependencies, case links a noun to its preposition. So San Diego -> from. If the arrow points to San Diego, then from is the head? Let's reconsider the standard visualization. Maybe it's Dependent -> Head? Two (dependent) -> linguists (head). Arrow points to linguists. Let's check the image for arrowheads again. Look at NUMMOD. The line goes from Two up and over to... wait. Actually, let's look at the arrowheads in the provided crops. Crop 1: Arrowhead points DOWN to Two. Label NUMMOD. This means something points TO Two. Crop 2: Arrowhead points DOWN/LEFT to computational. Label AMOD. Something points TO computational. Crop 3: Arrowhead points DOWN to San Diego? No, let's look at the arc labeled CASE. It connects from and San Diego. The arrowhead is on the right side of the arc, pointing down to San Diego. Wait. Let's look at the arc labeled NMOD (the one above from). It connects linguists and... it seems to point to San Diego? Or does it connect linguists and from? This is getting confusing without a clear view of every single arrowhead. Let's look at the whole image again. Let's try the Dependent -> Head hypothesis (Arrows point to the head): If arrows point to the head: Two points to linguists. Label NUMMOD. Correct. Two modifies Linguists. computational points to linguists. Label AMOD. Correct. Computational modifies Linguists. San Diego (or the phrase) points to walked? No, that's NSUBJ. Let's look at the arc labeled NSUBJ. It connects walked and linguists. If it points to linguists, then walked is the dependent? No, walked is the root. So linguists must point to walked? Let's look at the arc labeled NSUBJ. It starts above linguists and goes over to... wait. Actually, looking at the layout: Two, computational, linguists are grouped on the left. from San Diego is next. walked is in the middle-right. into a bar is on the right. Let's try to map standard Stanford Dependencies (which this looks like): In Stanford This description was generated automatically. Please feel free to ask questions if you have further questions about the nature of the image or its meaning within the presentation.">

---

### Let's try it: "Will spotted Bigfoot outside the Medical Library"

---

### "Will spotted Bigfoot outside the Medical Library"

<img class="r-stretch" src="diagrams/dep_bigfoot.jpg" alt="depbigfoot.jpg The user wants a detailed description of an image containing a syntactic dependency tree diagram. 1. Analyze the Image: Type: It's a linguistic diagram, specifically a dependency parse tree. Content: It shows the grammatical relationships between words in a sentence. Sentence: Will spotted Bigfoot outside the Medical Library Structure: There are words at the bottom: Will, spotted, Bigfoot, outside, the, Medical, Library. There are curved lines (arcs) connecting these words to a central root or each other. Each arc has a label indicating the grammatical relationship (e.g., ROOT, NSUBJ, DOBJ). 2. Breakdown by Component: Words (Left to Right): 1. Will 2. spotted 3. Bigfoot 4. outside 5. the 6. Medical 7. Library Arrows/Lines and Labels: ROOT: A vertical line points down from the word spotted. This indicates spotted is the main verb or root of the sentence. NSUBJ (Nominal Subject): An arc connects Will to spotted. The arrow points from spotted to Will. Wait, looking closely at standard dependency parsing conventions (like Stanford Dependencies), usually the head points to the dependent. Let's re-examine the arrows. The line labeled ROOT goes from spotted upwards? No, it looks like a vertical line dropping down onto spotted. Actually, in many visualizations, the root is just marked. Here, there is a label ROOT with a line pointing straight down to spotted. The arc labeled NSUBJ connects Will and spotted. The arrowhead points towards Will. This means spotted (the head) governs Will (the dependent). Wait, usually the subject is the head of the verb in some frameworks, but in dependency grammar, the verb is often the root. Let's look at the direction. Actually, let's look at the arrowheads carefully. NSUBJ: The arc goes from spotted to Will. Arrow points to Will. This implies spotted is the head and Will is the nominal subject dependent. DOBJ: The arc goes from spotted to Bigfoot. Arrow points to Bigfoot. This implies spotted is the head and Bigfoot is the direct object dependent. NMOD (Nominal Modifier): A very long, high arch connects Library to outside? No, let's trace it carefully. The label is NMOD. The line starts at Library (far right) and goes all the way left to point at Bigfoot? Or does it start at outside? Let's look at the arrowhead. The arrowhead for the big NMOD arc points to outside. Wait, no. Let's trace from the label NMOD. It sits on a long curve. The right end of the curve is near Library. The left end has an arrow pointing down at outside. This suggests outside modifies something related to Library? Or perhaps Library modifies outside? Let's re-evaluate standard dependency trees. Usually, prepositions are heads for their objects. Let's look at the cluster on the right: outside, the, Medical, Library. CASE: An arc connects outside and the? No. The label is CASE. It sits on an arc connecting outside to... wait. Let's trace from the words up. Library: Has an arrow coming from Medical (labeled AMOD). So Medical modifies Library. Medical: Has an arrow coming from the (labeled DET). So the modifies Medical? Or does the modify Library and Medical is in between? In Stanford dependencies, determiners usually point to the noun. Here, the arc labeled DET connects the and Medical. The arrow points from Medical to the? No, let's look at the arrowheads again. Let's assume standard dependency direction: Head -> Dependent. spotted is ROOT. Will is NSUBJ of spotted. (Head: spotted -> Dep: Will). Arrow points to Will. Correct. Bigfoot is DOBJ of spotted. (Head: spotted -> Dep: Bigfoot). Arrow points to Bigfoot. Correct. Now the prepositional phrase outside the Medical Library. Usually, a preposition like outside modifies the verb or noun. Here, it likely modifies Bigfoot or spotted. Let's look at the big arc labeled NMOD. It connects Library (far right) to... wait, let's trace the line from NMOD. The left side of the curve has an arrow pointing down at outside. The right side of the curve starts near Library. This implies Library is the head and outside is the modifier? That seems odd for a prepositional phrase. Let's try the other way: outside is the head. Let's look at the arc labeled CASE. It connects outside to... wait, looking closely at the right side cluster. There are nested arcs. Innermost arc: Label AMOD (Adjectival Modifier). Connects Medical and Library. Arrow points from Library to Medical? Or Medical to Library? The arrowhead is on the left side of the arc, pointing at Medical. Wait, no. Let's look really closely at the arrowheads on the right side. Arc labeled AMOD: Connects Medical and Library. Arrow points from Library to Medical? No, that would mean Library modifies Medical. Usually Medical modifies Library. So Head: Library -> Dep: Medical. Arrow should point to Medical. The arrowhead is on the left side of that specific arc. Yes, pointing at Medical. Arc labeled DET: Connects the and... wait. It connects the and Medical? Or the and Library? The line starts above the, goes right, and points to Medical. Wait, looking at the arc labeled DET. It seems to connect the and Medical. Arrow points from Medical (or Library?) to the. Let's reconsider the structure based on standard English grammar parsing (likely Stanford Dependencies). Root: spotted. Subject: Will (NSUBJ). Object: Bigfoot (DOBJ). Prepositional phrase modifying Bigfoot? Or spotted? Head of PP: outside. Dependent of outside: Library (the noun being located). Modifiers of Library: Medical, the. Let's trace the lines again with this hypothesis. NMOD: Long arc. Left side points to outside. Right side starts at... actually, looking at the rightmost line, it seems to originate from Library and go all the way left to point at outside? No, that would make Library the head of outside. Let's look at the arrowheads again. They are crucial. NSUBJ: Arrow points LEFT (to Will). Head is spotted. DOBJ: Arrow points RIGHT (to Bigfoot). Head is spotted. NMOD: This is a huge arc. It seems to connect Library and outside. The arrowhead is on the left, pointing at outside. Wait, if outside is the head, the arrow should point to it? No, in dependency trees, arrows usually point from Head to Dependent. If Head -> Dep: spotted -> Will (NSUBJ). Arrow points to Will. Correct. spotted -> Bigfoot (DOBJ). Arrow points to Bigfoot. Correct. So, for the right side: There must be a head that governs outside. Or outside is the head governing something. Let's look at the arc labeled NMOD. It connects Library (far right) and... actually, looking at the very top line labeled NMOD. The left end has an arrow pointing down to outside. The right end seems to originate from Library. This implies Head: Library -> Dep: outside? That's weird. Alternative interpretation: The diagram is drawn with arrows pointing from Dependent to Head? If Arrow points to Head: Will -> spotted (NSUBJ). Correct. Bigfoot -> spotted (DOBJ). Correct. Then for the right side: outside -> ? Let's look at CASE. The arc connects outside and... wait, looking at the label CASE. It is on an arc connecting outside to... actually, it looks like it connects outside to the? No. Let's look at the cluster of lines on the right again. There is a line labeled NMOD (topmost). It spans from Library all the way left to point at outside. Wait, if arrows point Dep -> Head: Then Library modifies outside? No. Let's try standard Stanford Dependencies visual style. Usually, the head is above the dependent. Here, words are on a line. Arcs go over them. Arrowheads indicate direction of dependency (Head -> Dependent). Let's re-examine NMOD. The label is at the top. The arc connects Library and... wait, looking closely at the right side of the NMOD arc. It seems to connect to outside? No, that doesn't make sense. Let's look at the arc labeled CASE. It connects outside and... actually, it looks like it connects outside to the? No. Let's look at the arc labeled DET. Connects the and Medical. Arrow points from Medical to the? Or the to Medical? Let's look at the arc labeled AMOD. Connects Medical and Library. Arrow points from Library to Medical? Let's try a different perspective: The labels describe the relationship of the word BELOW the arrow to the word ABOVE/BELOW. Actually, let's just describe what is visually there without over-interpreting the linguistic theory if it's ambiguous, but try to be accurate. Visuals: Bottom row text: Will spotted Bigfoot outside the Medical Library Top label: ROOT. Vertical line points down to spotted. Arc left of root: Label NSUBJ. Connects Will and spotted. Arrow points from spotted to Will. (Wait, looking at the curve... it goes UP from Will, peaks, and comes DOWN to spotted? No. It's an arc ABOVE the words. The arrowhead is on the left side, pointing at Will. This implies Spotted -> Will). Arc right of root: Label DOBJ. Connects spotted and Bigfoot. Arrow points from spotted to Bigfoot. (Arc goes up from spotted, peaks, down to Bigfoot. Arrowhead on the right side pointing at Bigfoot). Now the complex part on the right. There is a very long arc labeled NMOD. It starts above Library (far right) and arches all the way left to point at outside. Wait, let's look at the arrowhead. The arrowhead is on the LEFT side of this long arc, pointing down at outside. This implies Head: Library -> Dep: outside? That seems linguistically wrong for Bigfoot spotted... outside.... Usually outside modifies spotted or Bigfoot. Let's look at the arrowhead again. Is it possible the arrow is on the RIGHT side of that long arc? No, looking closely at crop 1 and 2... The line labeled NMOD has an arrowhead on the LEFT end pointing to outside. And the right end seems to connect to Library. Wait, let's look at the other arcs. Arc labeled CASE: Connects outside and... wait. It connects outside and the? No. Let's trace from outside. There is an arc going up from outside. It has a label CASE above it? No, the label CASE is on an arc connecting outside to... actually, looking at the layout: Word: outside. Above it is an arc labeled CASE. This arc connects outside to... wait, no. The word outside has an arrow pointing TO IT from a higher arc? Let's look at the arrows pointing DOWN to the words. To outside: There are two lines coming down. One is the long NMOD arc. One is an arc labeled CASE. Wait, no. Let's restart the tracing of the right side based on standard dependency parsing (likely Stanford Parser output). Sentence: Will spotted Bigfoot outside the Medical Library. Root: spotted. Subject: Will (nsubj(spotted, Will)). Arrow: spotted -> Will. Object: Bigfoot (dobj(spotted, Bigfoot)). Arrow: spotted -> Bigfoot. Prepositional phrase: outside the Medical Library. This likely modifies Bigfoot or spotted. Let's assume it modifies Bigfoot (location of spotting). Or maybe Library is the head? No. In Stanford Dependencies, prepositions are often heads. So outside would be a dependent of Bigfoot This description was generated automatically. Please feel free to ask questions if you have further questions about the nature of the image or its meaning within the presentation.">

---

### Dependency Parsing can handle huge complexity

<img class="r-stretch" src="diagrams/dep_emperor.jpg" alt="depemperor.jpg The user wants a detailed description of an image showing a syntactic dependency parse tree for a sentence. 1. Identify the core content: - It's a diagram representing grammatical structure (dependency parsing). - The sentence being analyzed is: The Emperor recently completely demolished the Cruiser at Endor with the Death Star to crush the Rebels. - There are words arranged horizontally at the bottom. - Arrows connect these words, indicating relationships. - Labels above the arrows describe the grammatical relationship (e.g., NMOD, DOBJ, ROOT). 2. Break down the sentence and structure: - Words (Left to Right): The, Emperor, recently, completely, demolished, the, Cruiser, at, Endor, with, the, Death, Star, to, crush, the, Rebels. - Root: demolished is the central verb. It has a vertical line pointing down labeled ROOT. 3. Analyze the connections (dependencies): - The Emperor: - The connects to Emperor with label DET (Determiner). - Emperor connects up to demolished with label NSUBJ (Nominal Subject). - Adverbs modifying demolished: - recently connects to demolished with label ADVMOD (Adverbial Modifier). - completely connects to demolished with label ADVMOD. - the Cruiser: - the connects to Cruiser with label DET. - Cruiser connects up to demolished with label DOBJ (Direct Object). - at Endor: - at connects to Endor with label CASE. Wait, looking closely at the arrows. - Actually, at is a preposition. The arrow goes from Cruiser (implied or via structure) to at? No, let's look closer. - There is an arrow labeled NMOD going from demolished all the way over to Endor. Wait, no. Let's trace carefully. - The label NMOD is above a long arc connecting demolished to at? No, that looks like it connects demolished to Cruiser (already labeled DOBJ). - Correction: There are multiple arcs originating from demolished. - Arc 1 (left): To Emperor (NSUBJ). - Arc 2 (left-mid): To recently (ADVMOD). - Arc 3 (mid-left): To completely (ADVMOD). - Arc 4 (right-mid): To Cruiser (DOBJ). Wait, the label is DOBJ. The arrow points from demolished to Cruiser. - Arc 5 (further right): To Endor? No, let's look at the cluster on the right. - There is a long arc labeled NMOD connecting demolished to at Endor? Or demolished to Cruiser? - Let's re-examine the arcs originating from demolished. - One goes left to Emperor (NSUBJ). - One goes left to recently (ADVMOD). - One goes left to completely (ADVMOD). - One goes right to Cruiser (DOBJ). Wait, the label DOBJ is above an arc connecting demolished and Cruiser. - There is another arc labeled NMOD. It seems to connect demolished to at Endor? Or maybe Cruiser to Endor? - Let's look at the label NMOD near the middle right. The arrow points from demolished to at. No, that doesn't make sense. - Let's look at the label CASE. It connects at and Endor. at is the case marker for Endor. - There is a long arc labeled NMOD connecting demolished to at Endor? Or perhaps Cruiser to Endor? - Actually, looking at standard dependency trees (like Stanford Dependencies), prepositional phrases often modify the verb or noun. Here, at Endor likely modifies demolished. So an arc from demolished to at or Endor labeled NMOD makes sense. The label NMOD is above a long arc connecting demolished to at. Wait, the arrow head points to at? No, usually arrows point to the head. - Let's look at the arrow heads. - DET arrow points from The to Emperor. (Head is Emperor). - NSUBJ arrow points from demolished to Emperor. (Head is demolished). Wait, usually arrows point to the head. If Emperor is subject of demolished, the arrow goes from Subject to Verb? Or Verb to Subject? - In Stanford dependencies, arrows point to the head. So Emperor -> demolished. The label NSUBJ is on the arc. - Let's re-evaluate based on arrow direction. - Arrow from The points to Emperor. Label DET. (Head: Emperor). - Arrow from Emperor points to demolished. Label NSUBJ. (Head: demolished). - Arrow from recently points to demolished. Label ADVMOD. (Head: demolished). - Arrow from completely points to demolished. Label ADVMOD. (Head: demolished). - Arrow from the points to Cruiser. Label DET. (Head: Cruiser). - Arrow from Cruiser points to demolished. Label DOBJ. (Head: demolished). - Now the prepositional phrase at Endor. - Arrow from Endor points to... where? There is an arrow labeled CASE pointing from at to Endor? No, usually at marks the case of Endor. So Endor -> at (case). Or at -> Endor? - Let's look at the label CASE. It's above an arc connecting at and Endor. The arrow points from Endor to at. This means at is the head of Endor. - Now, what does at Endor modify? There is a long arc labeled NMOD (Nominal Modifier) or maybe it's modifying the verb. Wait, looking at the top right arcs. - There are three very long arcs originating from demolished going all the way to the right side of the sentence. - Topmost arc: Label PCOMP. Points to to. (Infinitive phrase modifier). - Middle arc: Label NMOD. Points to... wait, let's trace it. It seems to point to at Endor? Or maybe Cruiser? - Let's look at the cluster of arcs on the right side of demolished. - There is an arc labeled DOBJ pointing to Cruiser. - There is an arc labeled NMOD pointing to... it looks like it points to at Endor or maybe Endor. Let's assume it modifies the verb demolished indicating location. - Actually, let's look at the arcs originating from demolished again. 1. To Emperor (NSUBJ). 2. To recently (ADVMOD). 3. To completely (ADVMOD). 4. To Cruiser (DOBJ). Wait, the label DOBJ is above an arc connecting demolished and Cruiser. 5. There is a long arc labeled NMOD. It connects demolished to... it looks like it connects to at Endor? Or maybe Endor? Let's assume it connects to the prepositional phrase head at. 6. There is another long arc labeled PCOMP. This usually stands for Prepositional Complement or something similar, often linking a verb to an infinitive clause. It connects demolished to to crush. - with the Death Star: - This is another prepositional phrase. - the -> Death (DET). - Star -> Death (AMOD? No, Death Star is a compound noun. The label AMOD is above an arc from Star to Death? Or Death to Star? Usually Star modifies Death in this context, or they form a compound. The label says AMOD. Arrow points from Star to Death. So Death is the head). - Wait, let's look at the arc labeled CASE. It connects with and the Death Star. - Actually, looking closely at the group with the Death Star: - with has an arrow pointing to... where? - There is a label CASE above an arc connecting with and Endor? No, that's further left. - Let's look at the group starting with with. - Word: with. - Word: the. Arrow from the to Death (DET). - Word: Death. - Word: Star. Arrow from Star to Death labeled AMOD. (Adjectival Modifier? Or part of compound). - There is a label CASE above an arc connecting with and... wait, the label CASE is above the arc between Endor and at? No, that was earlier. - Let's look at the label CASE near with. Ah, there is a label CASE above an arc connecting with to... actually, looking at the layout: - at Endor: Label CASE connects at and Endor. Arrow points from Endor to at. - with the Death Star: There is a label CASE? No. - Let's look at the arcs connecting back to the main verb demolished. - There is a long arc labeled NMOD (the second one down in the top cluster). It connects demolished to... it looks like it points to with? Or Endor? - Actually, let's look at the arcs originating from demolished again. - Arc 1: To Emperor (NSUBJ). - Arc 2: To recently (ADVMOD). - Arc 3: To completely (ADVMOD). - Arc 4: To Cruiser (DOBJ). Wait, the label DOBJ is there. - Arc 5: Labeled NMOD. Points to... it looks like it points to at Endor. Specifically, maybe to Endor? Or at? In Stanford dependencies, prepositional phrases modifying a verb are often attached via nmod. The head would be the preposition at or the noun Endor. Here, the arrow seems to point to at. - Arc 6: Labeled PCOMP. Points to to crush. This indicates an infinitive clause acting as a complement (purpose). - to crush the Rebels: - to connects to... crush? Or is it part of the PCOMP link? - The label PCOMP points from demolished to to. - Inside the infinitive phrase: - to -> crush (MARK). Marker for infinitive. - crush -> ... wait, who is the head of this clause? It's linked back to demolished via PCOMP. So crush is the root of the subordinate clause. - the Rebels: - the -> Rebels (DET). - Rebels -> crush (DOBJ). Direct object of crush. Let's refine the description of the arcs to be precise. Central Node: The word demolished is at the bottom center. A vertical line labeled ROOT points down from it, indicating it is the root of the sentence. Subject: To the left, Emperor connects to demolished via an arc labeled NSUBJ. The connects to Emperor via DET. Adverbs: recently and completely connect to demolished via arcs labeled ADVMOD. Direct Object: Cruiser connects to demolished via an arc labeled DOBJ. the connects to Cruiser via DET. Location (Prepositional Phrase 1): There is a long arc labeled NMOD connecting demolished to the preposition at (or possibly Endor, but usually the preposition). Wait, looking at the arrow head for that specific NMOD arc... it points to at. - Inside this phrase: Endor connects to at via an arc labeled CASE. The arrow points from Endor to at. This means at is the head. Instrument (Prepositional Phrase 2): There is another long arc labeled NMOD connecting demolished to... wait, let's look at the labels again. - Topmost label: PCOMP. Arc connects demolished to to. - Middle top label: NMOD. Arc connects demolished to... it looks like it points to with? Or maybe Endor? - Let's look at the cluster above Cruiser, at Endor, with the Death Star. - There are three long arcs originating from demolished going right. 1. Lowest of the long arcs: Labeled NMOD. Points to... it seems to point to at Endor? Or maybe Cruiser? No, DOBJ is separate. 2. Middle of the long arcs: Labeled NMOD. Points to... with? 3. Topmost arc: Labeled PCOMP. Points to to. Correction on Arcs: Let's look really closely at the connections for at Endor and with the Death Star. - The label NMOD appears twice in the upper right area. - One NMOD arc connects demolished to... it looks like it points to Endor? Or at? Let's assume it modifies the verb demolished indicating location. In many parsers, nmod links the noun (Endor) or preposition (at) to the head. - The other NMOD arc connects demolished to... with? This would indicate instrument. - Let's trace the lines carefully. - Line 1 (lowest of the long ones): Starts at demolished, goes up and right, ends at at. Label: NMOD. Wait, looking at the arrow head... it points to at. So at modifies demolished. - Line 2 (middle of the long ones): Starts at demolished, goes up and further right, ends at with. This description was generated automatically. Please feel free to ask questions if you have further questions about the nature of the image or its meaning within the presentation.">

---

### So, Dependency Parsing is a different beast

- Focuses on demonstrating the relationship between elements in the sentence

- Dependency ('what's attached to what')

- Relations between arguments ('Who's doing what to whom')

- Misses some linguistic insight, but gets what we want for NLP

---

# How do we do dependency parsing automatically?

---

---

### We use any number of different algorithms

- They examine each part of the sentence, and use machine learning to decide the proper relationships

- ... but they're a bit too complex (and linguistically uninteresting) for this class

- The Arc Standard approach is most commonly used

- You can [read about it in this chapter](https://web.stanford.edu/~jurafsky/slp3/13.pdf) or [watch this lecture](https://www.youtube.com/watch?v=nC9_RfjYwqA)
	
---

### We're going to focus on why it's done

- ... and what it wins us

---

# Why Dependency Parsing in NLP?

---

## Dependency Parsing is useful

---

### We can still see some elements of constituency

---

### Ambiguity is very easy to resolve with dependency parses

---

### The Angry Squirrel stole [the Mixed Nuts] [from Walmart]

---

### The Angry Squirrel stole [the Mixed Nuts from Walmart]

---

### "I saw the man with the telescope"

<img class="r-stretch" src="diagrams/dep_manwithtelescope1.jpg" alt="depmanwithtelescope1.jpg The user wants a detailed description of an image containing a syntactic dependency tree diagram. 1. Analyze the Image: Type: It's a linguistic diagram, specifically a dependency parse tree for a sentence. Text at the bottom (the sentence): I saw the man with the telescope Structure: Arrows connect words to other words, indicating grammatical relationships. Labels are placed above these arrows. 2. Breakdown of Components: Words: I, saw, the, man, with, the, telescope. Labels (Relationships): ROOT: Points from saw upwards to a label ROOT. This indicates saw is the main verb. Wait, looking closely at the arrow direction. The arrow points from saw up to ROOT? No, usually in these diagrams (like Stanford Dependencies), the arrow points from the head to the dependent. Let's re-examine standard dependency parsing visualizations. Actually, let's look at the arrows carefully. Arrow from saw goes up and left to I. Label: NSUBJ (Nominal Subject). This means I is the subject of saw. Wait, usually the arrow points from head to dependent. If saw is the root, it's the head. So an arrow from saw to I labeled NSUBJ makes sense. Arrow from saw goes up and right to man. Label: DOBJ (Direct Object). This means man is the direct object of saw. Arrow from man goes down/left to the. Label: DET (Determiner). Wait, looking at the arrow direction. The arrow starts at man and points to the. No, that's wrong. Let's look really closely at the arrowheads. Arrow between saw and I: Points to I. Label: NSUBJ. So saw -> I. Arrow from ROOT label down to saw: This indicates saw is the root of the tree. Arrow from saw goes up and right, arching over everything to telescope? No, let's trace the big arc labeled NMOD. It starts at man (or near it) and ends at with. Wait, let's look at the endpoints. Let's re-evaluate the arrows based on standard dependency parsing logic (Head -> Dependent). Root: The word saw has a vertical line going up to the label ROOT. This signifies it is the root of the sentence. I: An arrow curves from saw to I labeled NSUBJ. This means I is the nominal subject of saw. man: An arrow curves from saw to man labeled DOBJ. This means man is the direct object of saw. the (first one): An arrow curves from man to the labeled DET. Wait, usually determiners depend on nouns. So man -> the. The arrow points to the. Yes. The big arc labeled NMOD: This is tricky. It starts at man and goes all the way to with? Or does it start at saw? Let's look at the origin. It seems to originate from man (or the cluster around it) and points to telescope. Wait, let's look at the arrowheads. Let's trace the lines again carefully. Line labeled NMOD: Starts near man, goes high up, ends pointing at with? No, that doesn't make sense syntactically for man with telescope modifying something else. Or does it modify saw? Saw... with telescope. That's a prepositional phrase modifying the verb. Let's look at the arrowheads again. Arrow from NMOD label points to with. No, the line is above the words. The line connects two words. Let's assume standard Stanford Dependencies style where arrows point from Head -> Dependent. Head: saw. Dependent: I (NSUBJ). Arrow points to I. Correct. Dependent: man (DOBJ). Arrow points to man. Correct. Dependent: with (NMOD? or PREP?). The big arc labeled NMOD connects saw and with? Or man and telescope? Let's look at the right side. There is a cluster around with, the, telescope. Arrow from with points to the. Label: DET. Wait, usually prepositions head their phrases. So with -> telescope. Let's look at the arrow labeled CASE. It connects with and telescope? No. Let's look at the arrow labeled DET on the right. It connects the and telescope. Arrow points to the. So telescope -> the. This makes sense (Determiner depends on Noun). Now, what does with depend on? The big arc labeled NMOD. It seems to connect man and with? Or saw and with? Let's look at the arrowhead for the big arc. It points to telescope. Wait, no. Let's try a different interpretation. Maybe the arrows point from Dependent -> Head? If arrows point to heads: I -> saw (NSUBJ). Arrow points to saw. But the arrow in the image points to I. So that hypothesis is wrong. Arrows point to dependents. Okay, back to Head -> Dependent. Root: saw. NSUBJ: I depends on saw. Arrow: saw -> I. (Matches image). DOBJ: man depends on saw. Arrow: saw -> man. (Matches image). DET (left): the depends on man. Arrow: man -> the. (Matches image - arrow points to 'the'). Right side cluster: We have with, the, telescope. There is an arc labeled CASE. It connects with and... telescope? No, it looks like it connects with and the word to its right. Wait, let's look at the arrowheads on the right side. Arrow from with points down to... nothing? No, there is an arrow pointing from with to the? No. Let's look at the arc labeled CASE. It connects with and telescope? No, it looks like it connects man (or somewhere near it) to with. Actually, let's look at the specific connections for the phrase with the telescope. In standard parsing: telescope is the head of the noun phrase. the depends on telescope (DET). with is a preposition. Let's look at the arc labeled NMOD. It starts at man and ends pointing at with? Or does it start at saw and end at telescope? Let's trace the line for NMOD carefully. It starts above man, goes up, arches right, and points down to with. Wait, if with depends on man (NMOD - nominal modifier), that would mean the man with the telescope. This is a very common ambiguity resolution in parsing. Let's check the other possibility: saw... with the telescope. Here with... modifies saw. Let's look at the arrowheads again. Big arc NMOD: Starts near man, ends pointing to telescope? No, that would be weird. Let's look at the arc labeled CASE. It connects with and... the? No. Let's try to interpret the visual lines strictly without linguistic knowledge first. Line 1: Vertical line from saw up to ROOT. Line 2: Arc from saw left to I, label NSUBJ. Arrow points to I. Line 3: Arc from saw right to man, label DOBJ. Arrow points to man. Line 4: Arc from man down/left to the, label DET. Arrow points to the. Line 5 (The big one): Starts above man, goes high up, label NMOD, ends pointing at... telescope? Or with? Let's look at the arrowhead. It seems to point to telescope. Wait, if telescope depends on man via NMOD, that implies the man [who is] with the telescope. Line 6: Arc from telescope left to with, label CASE? No, usually prepositions depend on nouns (case marking) or vice versa. In Stanford dependencies, nmod:with means man with.... The relation is often labeled nmod. Here the label is NMOD for the big arc. Let's look at the right side arcs again. There is an arc labeled CASE. It connects with and... telescope? No, it looks like it connects man (or the NMOD source) to with. Actually, let's look at the arrowheads on the far right. There are two arrows pointing down into the telescope. One comes from DET label. It points to the. So telescope -> the. (Head: telescope, Dependent: the). One comes from... where? Let's look at the arc labeled CASE. It connects with and telescope? No. Let's reconsider the whole right side structure. Maybe the big arc NMOD connects man to with. Label: NMOD (Nominal Modifier). This fits the man with.... Then, what does with depend on? Or what depends on with? There is an arc labeled CASE. It seems to connect with and telescope. In some tagsets, the preposition marks case for the noun. So telescope -> with (case). Arrow points to with? No, arrow points from telescope to with? Let's look at the arrowheads on the right side very carefully. There is an arc labeled CASE. It starts near man/NMOD area and ends pointing at with. Wait, no. Let's look at the arc labeled DET (right side). It connects telescope to the. Arrow points to the. This confirms telescope is the head of that phrase. Now, what connects to telescope? The big arc NMOD. It seems to point to telescope. If man -> telescope (NMOD), that means man [modified by] telescope. That's weird. Let's try: saw -> with (NMOD). This would mean the PP modifies the verb. Arc starts at saw, goes over everything, points to with. Then with -> telescope (case? or nmod?). Let's look at the label CASE. It is above an arc connecting with and... the? No. Let's look at the label DET (right). Above an arc connecting the and telescope. Arrow points to the. So telescope -> det: the. Correct. Now, what is the relationship between with and the rest? There is a label CASE. It is above an arc. The arc connects... man? No. Let's look at the arrowheads again. Arrow from NMOD points to telescope. (Head: man, Dependent: telescope?? No). Maybe Head: saw. Dependent: with (via NMOD). Then with -> telescope. Let's look at the specific lines. Line from NMOD label goes down-right. It points to telescope. Wait, no. The line ends in an arrowhead pointing at telescope. Line from CASE label goes down-left. It points to with. This is confusing. Let's look at the standard I saw the man with the telescope ambiguity. Interpretation A: I used a telescope to see the man. (PP modifies verb). Structure: saw -> with (nmod). with -> telescope (case/nmod). Interpretation B: I saw the man who had a telescope. (PP modifies noun). Structure: man -> with (nmod). with -> telescope. Let's look at the diagram to see which interpretation it depicts. The big arc NMOD starts above man and ends pointing at... telescope? Or with? Actually, looking really closely at crop 1 and 2... Crop 1 shows ROOT, NSUBJ. Crop 2 shows NMOD, DOBJ, DET (left). Crop 3 shows CASE, DET (right), with, the, telescope. Let's trace the NMOD line. It starts above man (or slightly left of it, near the DO This description was generated automatically. Please feel free to ask questions if you have further questions about the nature of the image or its meaning within the presentation.">

---

### "I saw the man with the telescope"

<img class="r-stretch" src="diagrams/dep_manwithtelescope2.jpg" alt="depmanwithtelescope2.jpg The user wants a detailed description of an image containing a syntactic dependency tree diagram. 1. Analyze the Image: Type: It's a linguistic diagram, specifically a dependency parse tree for the sentence I saw the man with the telescope. Text at the bottom (the sentence): The words are arranged horizontally: I, saw, the, man, with, the, telescope. They are in a serif font. Arrows and Labels: There are curved arrows connecting words, indicating grammatical relationships. Each arrow has a label above it. From saw to I: Label is NSUBJ (Nominal Subject). The arrow points from saw down to I. Wait, looking closely at the arrow direction... actually, standard dependency trees usually point from head to dependent. Let's re-examine. The vertical line comes from ROOT pointing down to saw. This indicates saw is the root of the sentence. From saw, an arc goes left and down to I. The label above this arc is NSUBJ. This means I is the nominal subject of saw. From saw, an arc goes right and up, then curves down to man. Wait, let's trace carefully. There is a long arc labeled NMOD connecting man and telescope? No, that looks like it connects man and with. Let's look closer. Actually, let's trace from the heads. Head: saw. Dependent: I (NSUBJ). Head: saw. Dependent: man (DOBJ - Direct Object). The arc goes from saw over to the man. Wait, there are two arcs involved here. Arc labeled DOBJ connects saw and man. It curves up and down. Arc labeled DET connects man and the. It curves up from the to man? No, usually determiners modify nouns. So man is the head, the is the dependent (DET). The arrow points from man to the. Let's look at the arrowheads. Arrowhead on I: Points to I. Head is saw. Arrowhead on the (first one): Points to the. Head is man. Label: DET. Arrowhead on man: Points to man. Head is saw. Label: DOBJ. Now the right side: with the telescope. This is a prepositional phrase modifying man (or potentially saw, but usually attached to the noun in this specific ambiguity resolution example). Let's trace the arcs on the far right. There is an arc labeled NMOD connecting man and with. Wait, let's look at the endpoints. One end is above man, one end is above telescope? No. Let's look at the labels: NMOD, CASE, DET. Arc labeled NMOD: Connects man and with? Or man and telescope? The arc starts above man and ends above telescope. Wait, no. Let's re-evaluate the structure based on standard dependency parsing (like Stanford Dependencies). Root: saw. saw -> I (nsubj) saw -> man (dobj) man -> with (prep? or case?) No, with is a preposition. Let's look at the specific arcs drawn. Arc 1: From man to with. Label above it is NMOD. Wait, usually NMOD modifies a noun. So with... modifies man. The arrow points from man to with? Or with to telescope? Let's look at the arrowheads carefully. Arrowhead pointing left/down to I. Label: NSUBJ. Head is saw. Arrowhead pointing down/left to the (first). Label: DET. Head is man. Arrowhead pointing right/up from saw? No, the arc labeled DOBJ connects saw and man. The arrow points to man. So saw -> man (dobj). Now the complex part on the right. There is a large arc labeled NMOD. It starts above man and ends with an arrowhead pointing to with? Or telescope? Let's assume standard parsing: man modifies with? No, with modifies man. So head is man, dependent is with. The label NMOD suggests modification. Inside that, there is an arc labeled CASE. It connects with and telescope? Or with and the? Let's look at the nesting. Outermost right arc: Label NMOD. Connects man (left side) to telescope (right side)? No, that would be weird. Let's try another interpretation. Maybe it's a specific type of tree. Let's look at the arrowheads again. Arrowhead on with: Points from an arc labeled... wait. There is an arc connecting man and with. The label above it is NMOD? No, the label NMOD is way up high. It connects man to telescope? Let's look at the labels relative to the arcs. Label NMOD: Above a large arc connecting man and with. Wait, no. The arc starts above man and ends above telescope. No, that doesn't make sense grammatically if it modifies man. Let's look at the labels: NMOD, CASE, DET. Arc labeled DET: Connects the (second one) to telescope. Arrow points from telescope to the. So telescope is head, the is dependent. Correct. Arc labeled CASE: Connects with and telescope? Or with and the? The arc connects with and telescope. Arrow points from telescope to with. So telescope depends on with via a case relation? That's unusual. Usually prepositions take nouns as dependents (case). Let's reconsider the standard Stanford Dependency structure for this sentence which is famous for ambiguity (I saw the man with the telescope). Interpretation A: I used a telescope to see the man. (saw -> with) Interpretation B: The man had a telescope. (man -> with) This diagram seems to show Interpretation B (The man has the telescope). man is modified by with.... So head=man, dependent=with. Relation=prep or nmod? Here labeled NMOD. with takes telescope as object. Relation=case? Or obl? Let's look at the arrows in the image specifically. Arc 1 (Rightmost, lowest): Label DET. Connects the and telescope. Arrow points to the. Head=telescope. Dependent=the. Arc 2 (Middle right): Label CASE. Connects with and telescope? Or with and the? The arc connects with and telescope. Wait, the arrowhead is on with. So Head=telescope, Dependent=with? That implies telescope with.... No. Let's look really closely at the right side arcs. There is an arc labeled DET connecting the and telescope. Arrow points to the. (Head: telescope, Dep: the). There is an arc labeled CASE connecting with and... it looks like it connects with and telescope. The arrowhead points to with. So Head=telescope, Dep=with? That would mean telescope with. That's wrong. Let's try: Head=with, Dep=telescope. Relation=case? No, usually case is on the preposition. Let's look at the arc labeled NMOD. It connects man and... with? Or telescope? The arrowhead points to with. So Head=man, Dep=with. Relation=NMOD (Nominal Modifier). This fits Interpretation B. Now, what about the inner arcs? If Head=with, then it needs dependents. Arc labeled CASE: Connects with and telescope? Arrow points to telescope. So Head=with, Dep=telescope. Relation=case? In some tagsets (like Universal Dependencies), the preposition is the head, and the noun phrase is the dependent with relation obl or similar. But here it says CASE. Wait, let's look at the arrowheads again. Arc NMOD: Points from man to with. (Head: man, Dep: with). Arc CASE: Points from with to... wait. The arc labeled CASE is inside the NMOD arc. It connects with and telescope? No, it connects with and the telescope block? Let's trace the lines carefully. Line 1 (Rightmost): Arc labeled DET. Connects the and telescope. Arrow points to the. Head: telescope. Dep: the. Line 2 (Middle Right): Arc labeled CASE. Connects with and... it seems to connect with and telescope. Arrow points to telescope? No, arrow points to with? Let's look at the standard UDG (Universal Dependencies) or similar. Actually, let's just describe what is visually there without over-interpreting the linguistics if it's ambiguous, but try to be accurate. Visuals: Word man has an arc going up and right labeled NMOD. The arrowhead points to with. So man -> with (nmod). From with, there is an arc labeled CASE. It goes over the and points to... wait. Let's look at the labels again: NMOD, CASE, DET. The label NMOD is above a large arch connecting man and telescope? No, that would be weird. Let's assume the diagram shows: Root: saw. saw -> I (nsubj) saw -> man (dobj) man -> with (nmod) -- This explains man with.... with -> telescope (case?? No, usually obl). Maybe the label is CASE because it's a prepositional case marker? Or maybe the arrow points from telescope to with? Let's look at the arrowheads on the right side very carefully. Arrowhead 1: Points down-left to with. Label above arc: NMOD. So Head=man, Dep=with. Arrowhead 2: Points down-left to the (second one). Label above arc: DET. Wait, no. The label DET is above a small arch connecting the and telescope. Arrow points to the. Head=telescope, Dep=the. Arrowhead 3: Points down-left to... with? Or telescope? Let's look at the arc labeled CASE. It connects with and telescope. The arrowhead points to telescope. So Head=with, Dep=telescope. Relation=CASE? This is non-standard but possible in some frameworks (e.g., if 'with' is treated as a case marker on the noun). Or maybe the label is actually pointing to the relation between with and the telescope. Actually, looking really closely at crop 2 and 3... There is an arc labeled NMOD connecting man and telescope? No. Let's trace from the top down on the right side. Top label: NMOD. Below it is a long arch. Left end above man. Right end arrow points to with. So man -> with (nmod). Below that, label CASE. Arch connects with and... telescope? Arrow points to telescope. So with -> telescope (case). This implies with is the head of telescope. Below that, label DET. Arch connects the and telescope. Arrow points to the. So telescope -> the (det). Wait, if with -> telescope (case), then telescope is the dependent. And telescope -> the (det). This creates a chain: man -> with -> telescope -> the. Let's check the arrowheads again. Arc NMOD: Points to with. Head=man. Arc CASE: Points to telescope? Or with? The label is above an arc connecting with and telescope. The arrowhead is on the left side, pointing to with. Wait. If it points to with, then Head=telescope, Dep=with. That would mean telescope with.... Let's look at the label positions. NMOD is highest. CASE is below it. DET is lowest. This nesting suggests: Outer relation (NMOD) connects man and with. (Head: man, Dep: with). Middle relation (CASE) connects with and... wait. If NMOD connects man and with, then with is the dependent. What does with depend on? Nothing, it's a leaf in that branch unless it has its own dependents. Let's try: Head=man. This description was generated automatically. Please feel free to ask questions if you have further questions about the nature of the image or its meaning within the presentation.">

---

### "I saw the Queen of England's Hat"

<img class="r-stretch" src="diagrams/dep_queenenglandshat1.jpg" alt="depqueenenglandshat1.jpg The user wants a detailed description of an image showing a dependency parse tree for a sentence. 1. Identify the core content: - It's a linguistic diagram (dependency graph). - The sentence being analyzed is I saw the Queen of England 's hat. - There are words at the bottom and curved lines/arrows connecting them with labels on top. 2. Break down the visual elements: - Text at the bottom: I, saw, the, Queen, of, England, 's, hat. - Labels above the text (Dependency Relations): - ROOT: Points to saw. - NSUBJ: Connects I and saw. - DET: Connects the and Queen. - NMOD: Connects England and Queen (Wait, let me look closer). Actually, looking at the arrows: - There is a label NMOD connecting England to Queen? No, let's trace carefully. - Let's re-examine the connections from left to right. 3. Detailed Tracing of Connections: - Word: I - Arrow coming from saw pointing to I. Label above arrow: NSUBJ (Nominal Subject). This means I is the subject of saw. - Word: saw - Vertical line going up to label ROOT. This indicates saw is the root of the sentence. - Long curved arrow starting from hat and pointing to saw? No, let's look at the direction. The arrows point from the head to the dependent usually in these diagrams (Stanford dependencies style). Let's re-evaluate based on standard dependency parsing visualizations. - Actually, looking at the arrowheads: - Arrow from saw points to I. Label: NSUBJ. Correct. - Line goes up from saw to ROOT. Correct. - Long arc starting from hat pointing to saw? No, let's look at the arrowheads again. - The big arc labeled DOBJ starts at saw and points to hat. Wait, no. Let's look really closely at the arrowheads. - Arrowhead on I: Points left/down from saw. So saw -> I (NSUBJ). - Arrowhead on the: Points down/left from Queen. So Queen -> the (DET). - Arrowhead on hat: Points right/down from... where? The big arc labeled DOBJ connects saw and hat. Let's trace the line. It starts at saw, goes up high, and points to hat. So saw -> hat (DOBJ - Direct Object). - Now let's look at the middle section. - There is a cluster of arcs connecting Queen and England 's hat. - Arc labeled NMOD: Connects England to Queen? Or of England to Queen? Let's trace. It seems to connect England (or maybe the phrase) to Queen. Wait, let's look at the arrowheads. - Arrowhead points to Queen. The line comes from... England? No, that doesn't make sense syntactically for NMOD usually. Let's look at the label NMOD. It connects England to Queen? Or maybe of England modifies Queen. - Actually, let's look at the arrowheads again. - Arrowhead points to Queen. The line comes from... it looks like it originates near England. Wait, let's look at the label NMOD. It is above an arc connecting England and Queen. But usually NMOD (noun modifier) goes from head to dependent. If Queen is the head, the arrow should point to Queen. Yes, there is an arrow pointing to Queen. Where does it come from? It seems to come from England. Wait, that's weird. Queen of England. Usually of is the preposition. - Let's look at the label CASE. There is an arc labeled CASE connecting of and England. Arrow points to of? No, arrow points from England to of? Or of to England? - Let's re-read the diagram structure. It looks like a Stanford Dependency tree. - In Stanford dependencies: - nsubj(saw, I) -> Arrow from saw to I labeled nsubj. (Matches image). - root(ROOT, saw) -> Line from saw to ROOT. (Matches image). - dobj(saw, hat) -> Arrow from saw to hat labeled dobj. (Matches the big top arc). - det(Queen, the) -> Arrow from Queen to the labeled det. (Matches image: arrow points to 'the'). - Now the complex part: of England's. - There is a label NMOD. The arc connects England and Queen? No, let's look at the arrowheads. - Arrowhead points to Queen. The line comes from... it looks like it comes from hat? No. - Let's look at the labels: POSS, POSSESSIVE, NMOD. - There are three arcs originating from hat or near it? Or pointing to Queen? - Let's trace the lines carefully. - Line 1 (lowest of the middle group): Label NMOD. Connects England and Queen. Arrow points to Queen. Wait, is it nmod(Queen, England)? That would mean England modifies Queen. Plausible. But usually via a preposition. - Line 2 (middle of the middle group): Label POSS. Connects 's and England? Or hat and England? - Let's look at the right side. Words: 's, hat. - Arc labeled DOBJ connects saw to hat. (Confirmed). - Arc labeled POSS connects... let's trace it. It starts near Queen, goes over of England, and points to 's? Or does it start at 's and point to Queen? - Let's look at the arrowheads on the right side arcs. - There is an arrowhead pointing to 's. The line comes from... hat? No, that would be poss(hat, 's). That makes sense (the 's belongs to hat? No). - Let's try: Arrow points from 's to Queen? Label POSS. That means 's is a possessive marker for Queen. - Let's look at the label POSSESSIVE. It's above POSS. Maybe they are synonyms or different levels? No, looking closely, there are two distinct arcs. One labeled POSS and one labeled POSSESSIVE. Wait, let me zoom in mentally. - Ah, I see POSSESSIVE and POSS as separate labels on separate arcs. - Let's trace the arc for POSSESSIVE. It connects 's to... Queen? Or hat? - Let's trace the arc for NMOD. It connects England to Queen? - Let's look at the CASE label. Arc connects of and England. Arrow points to of. So case(of, England). This means of is a case marker for England. This fits Stanford dependencies perfectly. In Stanford deps, prepositions are often dependents of the noun they govern, labeled 'case'. Wait, usually it's nmod(England, Queen) with case(of, England). - Let's re-evaluate based on standard Stanford Dependency conventions which this looks like. - Sentence: I saw the Queen of England 's hat - Root: saw - nsubj(saw, I) -> Arrow saw->I (NSUBJ). Correct. - dobj(saw, hat) -> Arrow saw->hat (DOBJ). Correct. - det(Queen, the) -> Arrow Queen->the (DET). Correct. - Now of England 's. This is a possessive construction. - In Stanford deps: nmod(hat, Queen)? No. - Let's look at the arcs again. - Arc labeled NMOD: Connects England to Queen? Or hat to Queen? - Actually, let's look at the arrowheads pointing left. - There is an arrowhead pointing to Queen. The line comes from... it looks like it comes from 's or hat. - Let's assume the arrows point from Head to Dependent. - If Queen is the head of the, then Queen -> the (DET). Correct. - If England modifies Queen... usually nmod(England, Queen)? No, nmod(Queen, England). So arrow Queen -> England. But the arrow points to Queen. This implies England is the head and Queen is the dependent? That would be weird for Queen of England. - Let's look at the arrowheads again very carefully. - Arrowhead on the: Points left/down. Source: Queen. So Queen -> the (DET). Correct. - Arrowhead on of: Points left/down. Source: England. So England -> of (CASE). This fits case(of, England). Wait, usually it's nmod(England, Queen) and case(of, England). If England is the head, then Queen is dependent? No. - Let's try the other direction: Arrows point from Dependent to Head. - I -> saw. Label NSUBJ. (Dependent I points to Head saw). This fits standard dependency trees where arrows point to the governor/head. - the -> Queen. Label DET. (Dependent 'the' points to Head 'Queen'). Fits. - of -> England? No, look at the arc labeled CASE. It connects of and England. Arrowhead points to of. So England -> of? That would mean England is head of 'of'. Unlikely. - Let's look at the arc labeled CASE again. It connects of and England. The arrowhead is on the left side, pointing to of. Wait, no. The arrowhead is on the left end of the arc? No, looking closely at crop 2 and 5... - Crop 5 shows of England. There is an arc labeled CASE. It connects of and England. The arrowheads are pointing inwards? No, there is one arrowhead on the left pointing to of and one on the right pointing to England? No, that's not how these diagrams work. - Let's look at the full image again. - Arc labeled CASE: Connects of and England. Arrow points to of. Wait, looking really closely... The arrowhead is on the left side of the arc, pointing down to of. So England -> of (CASE). This implies England governs of? That's weird. - Let's reconsider the direction. Maybe arrows point from Head to Dependent? - If Head -> Dependent: - saw -> I (NSUBJ). Correct. - Queen -> the (DET). Correct. - saw -> hat (DOBJ). Correct. - Now, what governs of England? - Maybe Queen governs England? Arc labeled NMOD? - Let's look at the arc labeled NMOD. It connects England and Queen. Arrow points to Queen. So England -> Queen (NMOD). This means England modifies Queen. This fits nmod(England, Queen) in some notations, but usually it's nmod(Queen, England). - Let's look at the arc labeled POSS. Connects 's and... Queen? Arrow points to Queen. So 's -> Queen (POSS). This means 's is a possessive marker for Queen. - Let's look at the arc labeled POSSESSIVE. Connects 's and... hat? Or England? - Actually, let's look at the labels again. There are three arcs in the middle right area. 1. Top one: Label POSSESSIVE. Connects 's to Queen? No, looks like it connects 's to hat? No, that would be poss(hat, 's). 2. Middle one: Label POSS. Connects 's to... ? 3. Bottom one: Label NMOD. Connects England to... ? Let's try a different hypothesis. This is likely from the Stanford Parser output visualization. In Stanford Dependencies (v2/v3): - nsubj(saw, I) -> Arrow saw->I labeled nsubj. (Image has NSUBJ). Matches direction Head->Dependent? No, usually arrows point to head. Let's check standard Stanford dep images. - In standard Stanford Deps: Arrows point from Dependent to Head. - nsubj(saw, I): Arrow from I to saw labeled nsubj. - Image shows arrow from saw to I labeled NSUBJ. So this is Head -> Dependent direction? Or the labels are just placed there. - Let's look at DET. Arrow points to the. Source Queen. So Queen -> the. Head (Queen) -> Dependent (the). This confirms Head -> Dependent. - Let's check DOBJ. Long arc from saw to hat. Label DOBJ. Head (saw) -> Dependent (hat). Confirmed. Okay, so the arrows point from Head to Dependent. - Root: saw. This description was generated automatically. Please feel free to ask questions if you have further questions about the nature of the image or its meaning within the presentation.">

---

### "I saw the Queen of England's Hat"

<img class="r-stretch" src="diagrams/dep_queenenglandshat2.jpg" alt="depqueenenglandshat2.jpg The user wants a detailed description of an image containing a syntactic dependency tree diagram. 1. Analyze the Image: Content: It's a linguistic diagram showing the grammatical relationships between words in a sentence. Sentence: I saw the Queen of England 's hat (Note: The apostrophe-s is attached to the word 's). Wait, looking closely at the spacing, it looks like England 's. Let me re-examine. Yes, there are spaces around 's. So the sequence of words is: I, saw, the, Queen, of, England, 's, hat. Structure: It's a directed graph where arrows point from head words to dependent words. The labels on the arrows indicate the type of relationship (dependency relation). 2. Breakdown by Component: Words (Nodes): I saw the Queen of England 's hat Arrows and Labels (Edges): ROOT: A vertical line points down to saw. This indicates saw is the root of the sentence. NSUBJ (Nominal Subject): An arrow curves from saw up and left to I. This means I is the subject of saw. DOBJ (Direct Object): A long curved arrow goes from saw over the to Queen. Wait, let me look closer. The arrow labeled DOBJ starts at saw and points to Queen. This means Queen is the direct object of saw. DET (Determiner): An arrow curves from Queen down to the. This means the modifies Queen. NMOD (Nominal Modifier): A very long curved arrow goes from hat all the way back over to Queen. Wait, let me trace it carefully. The arrow labeled NMOD starts at hat and points to Queen. Actually, looking at standard dependency parsing (like Stanford Dependencies), usually modifiers point to their head. Let's re-evaluate directionality. Standard convention: Head -> Dependent or Dependent -> Head? In this specific diagram style (likely from a tool like spaCy or similar visualizer): ROOT points to saw. So saw is the root head. NSUBJ arrow goes from saw to I? No, usually arrows point from head to dependent in some styles, or dependent to head in others. Let's look at the arrowheads. Arrowhead on I: Points TO I. So saw -> I (NSUBJ). Wait, that would mean saw is the head of I. Yes, subject depends on verb. Arrowhead on Queen: Points TO Queen. So saw -> Queen (DOBJ). Verb is head of object. Arrowhead on the: Points TO the. So Queen -> the (DET). Noun is head of determiner. Arrowhead on of: Points TO of. Wait, let's look at the arrow labeled CASE. It points from England to of? No, that looks like an arc. Let's trace carefully. There is a label CASE above an arc connecting of and England. The arrowhead points to of. So England -> of (CASE). This means the preposition of depends on the noun England. There is a label POSS (Possessive) or POSSESSIVE? It says POSS under POSSESSIVE. The arrow points from 's to England? Or hat to 's? Let's look at the arc labeled POSSESSIVE/POSS. It connects 's and England. The arrowhead points to 's. So England -> 's (POS). There is a label NMOD. This is a huge arch connecting hat and Queen. The arrowhead points to Queen. Wait, let me look really closely at the NMOD line. It starts above hat and goes all the way left to point at Queen. So hat -> Queen (NMOD). This means hat modifies Queen? That seems odd for Queen of England's hat. Usually, hat is the head noun of the phrase England's hat, which modifies Queen? No. Let's re-parse the sentence structure logically: I saw [the Queen of England's hat]. Root: saw. Subject: I (NSUBJ). Object: hat (DOBJ)? Or Queen? If the object is hat, then Queen modifies hat. Let's look at the NMOD arrow again. It connects hat and Queen. The arrowhead points to Queen. Wait, no. Let me zoom in on the NMOD line. The line labeled NMOD starts above hat (right side) and curves left to point at Queen. So hat is a modifier of Queen? That would mean I saw the Queen [who has] hat. That's not right. Let's look at the arrow direction again. The arrowhead for NMOD points to Queen. This implies Queen is the head and hat is the modifier? No, that doesn't make sense semantically (Queen of England's hat). Alternative interpretation: Maybe the arrow points from Queen to hat? Let's look at the tip. The tip is on the right side, pointing down at hat. Ah, let me re-examine the whole NMOD arc. Okay, looking very closely at crop 1 and the main image. The line labeled NMOD starts above Queen (left side of the arch) and goes to the right, ending with an arrowhead pointing down at hat. So: Queen -> hat (NMOD). This means hat is a nominal modifier of Queen? No, that's still weird. The Queen... hat. Let's reconsider the structure. Maybe it's I saw [the Queen] [of England's hat]? No. Maybe the diagram says: hat is the head of the object phrase, and Queen modifies it? Let's look at the arrow for NMOD again. It connects Queen and hat. The arrowhead is on the right side, pointing to hat. Wait, let me look at the very top crop. Actually, looking at the full image, the line labeled NMOD starts above Queen and ends with an arrowhead pointing down at hat. No, wait. Let's trace from the label NMOD. The line goes right. It curves down. The arrowhead is on the far right, pointing to hat. Wait, let me look at the other side of that arch. Does it start at Queen? Yes. So Queen -> hat (NMOD). This implies hat modifies Queen. But semantically, Queen's hat is a unit. Let's look at the POSSESSIVE/POSS arc. It connects 's and England. The arrowhead points to 's. So England -> 's (POS). This means 's depends on England. Correct. Let's look at the CASE arc. It connects of and England. The arrowhead points to of. So England -> of (CASE). This means of depends on England. Wait, usually prepositions depend on nouns in some parsers (like Stanford), or nouns depend on prepositions in others (Universal Dependencies). In Universal Dependencies (UD): Noun is head of Preposition. So England -> of (case). Yes, that matches the arrow pointing to of. So we have a chunk: England is the head. It has dependents of and 's. This forms England's. Now back to NMOD. If hat is the head of the whole object phrase, then Queen modifies it? Or does Queen modify hat? Let's look at the arrow for NMOD again. It connects Queen and hat. The arrowhead points to Queen. Wait, let me look really closely at the right side of that long arch. Actually, looking at the full image, the line labeled NMOD starts above Queen (left) and goes all the way to the right, pointing down at hat. No, that's not right either. Let's try a different hypothesis. Maybe the arrow points from hat to Queen? Let's look at the arrowhead on the NMOD line. It is on the far right side. It points DOWN towards hat. Wait, no. The label is NMOD. The line goes over of England 's. The arrowhead is on the right end. It points to hat. If Queen -> hat (NMOD), then Queen modifies hat. This makes sense: The [Queen of England's] hat. Here, Queen... is a modifier for hat. Let's check the DOBJ arrow again. It goes from saw to Queen. If Queen is the head of the object phrase, then I saw Queen.... But the sentence ends with hat. So hat must be involved in the object relation. Let's look at the DOBJ arrow carefully. It starts above saw, curves up and right, over the, and points to Queen. This implies Queen is the direct object of saw. But then what about hat? Maybe the diagram is parsing: I saw [the Queen] [of England's hat]? No. Let's look at the NMOD arrow again. It connects Queen and hat. If the arrow points from Queen to hat, then hat modifies Queen. That would mean The Queen... hat. Like The Queen [who has a] hat. BUT, look at the label NMOD. In many parsers (like Stanford), NMOD links a noun modifier. Let's reconsider the arrow direction for NMOD. Maybe it points from hat to Queen? If hat -> Queen, then Queen is the head. This fits with DOBJ pointing to Queen. So the object of saw is Queen. And hat modifies Queen. Wait, that interpretation (I saw the Queen [of England's hat]) implies I saw a specific Queen who belongs to England's hat? That's nonsense. Let's try: The object is hat. Queen modifies hat. If Object is hat, DOBJ should point to hat. But the arrow labeled DOBJ points to Queen. This suggests the parser considers Queen the head of the noun phrase. Structure: [I] [saw] [the Queen [of England's hat]]. Meaning: I saw the Queen (who is associated with England's hat). Or maybe The Queen of England's hat refers to a specific entity? No, that's unlikely. Let's look at the POSSESSIVE arc again. It connects 's and England. Arrow points to 's. So England -> 's. This creates a unit England's. Now, how does England's connect? There is an arrow from 's (or near it) going up to... where? Let's look at the arc labeled POSSESSIVE/POSS. It connects 's and England. Is there another connection? Let's look at the NMOD arc again. It connects Queen and hat. If DOBJ points to Queen, then Queen is the head of the object phrase. So the structure is: saw -> Queen (DOBJ). Queen -> the (DET). Queen -> hat (NMOD)? Or hat -> Queen? If Queen -> hat (NMOD), then hat modifies Queen. This implies The Queen [of England's hat]. This is a possessive construction where the head noun is Queen. Wait. England's hat usually means the hat belonging to England. Queen of England's hat usually means the hat of the Queen of England. In the Queen of England's hat, the head noun is hat. The modifier is Queen of England. So, logically: Hat is the head. Queen modifies Hat. If Hat is the head, DOBJ should point to Hat. But the diagram shows DOBJ pointing to Queen. This suggests a specific parsing ambiguity or error in the diagram, OR I am misinterpreting the arrow direction of NMOD. Let's look at the NMOD arrow very carefully. The line starts above Queen. It goes right. It ends with an arrowhead pointing to hat. Wait, no. Let's look at the label NMOD. It is centered over a long arch. The left end of the arch is near Queen. The right end is near hat. Actually, looking really closely at the right side of the NMOD line... the arrowhead points to Queen. No, that's on the left side. Let's trace from the label NMOD. The text is centered. Below it is a long curve. The curve starts above Queen (left) and ends above hat (right). Where are the arrowheads? There is an arrowhead on the left side pointing to Queen. Wait, no. Let's look at the DOBJ line. It curves from saw to Queen. Arrow points to Queen. Let's look at the NMOD line. It seems to connect hat and Queen. Let's assume standard dependency parsing (like Stanford Parser). In Stanford, modifiers point TO their head. If hat is the head of the phrase Queen... hat, then Queen -> hat (NMOD)? No This description was generated automatically. Please feel free to ask questions if you have further questions about the nature of the image or its meaning within the presentation.">

---

## Dependency Parsing works better in some other languages

---

### In Russian, Word order can be used for emphasis

* Koshka videt sobaku
	* "The cat sees the dog"

* Videt koshka sobaku
	* "The cat sees the dog"

* Sobaku videt Koshka
	* "The cat sees the dog"

- Russian is a "non-configurational" language

---

### Constituency-based approaches struggle here

- <img class="r-stretch" src="humorimg/fail.jpg" alt="fail.jpg The image shows a dog in motion on a grassy field. The dog appears to be mid-action, possibly running or playing. It has light brown fur with some white markings and is wearing a black collar with a blue tag attached. Its mouth is open wide as if it's barking or panting, and there’s a tennis ball caught in its mouth. The word FAIL is prominently displayed across the image in bold, white capital letters, suggesting that something humorous has occurred. The background consists of green grass, indicating an outdoor setting, likely a park or yard. There are no other people or animals visible in this particular scene. This description was generated automatically from image files by a local LLM, and thus, may not be fully accurate. Please feel free to ask questions if you have further questions about the nature of the image or its meaning within the presentation.">

---

### Dependency parsing works just fine

---

---

---

---

## Dependency Parsing is a bit simpler

---

### Dependency Parsing can be done quickly

- Right now, dependency parsing is faster than constituency parsing

- This might change with new algos and approaches!

---

## Dependency Parsing hands us semantic information for free!

---

### Dependency Parsing gives us 'arguments'

---

### These arguments are directly useful in natural language understanding!

---

# How is parsing actually useful for NLP engineering?

---

## Dependency Parsing hands us information

---

### "Hey Siri, send my wife a romantic text"

<img class="r-stretch" src="diagrams/dep_sendwiferomantic.jpg" alt="depsendwiferomantic.jpg The user wants a detailed description of an image containing a syntactic dependency tree diagram. 1. Identify the core subject: It's a linguistic diagram showing the grammatical structure of a sentence. 2. Transcribe the text in the diagram: The words are Send, my wife, a, romantic, text. Wait, looking closer at the spacing, it reads: Send my wife a romantic text. 3. Analyze the structure (the arrows and labels): There is a vertical arrow pointing down to Send labeled ROOT. This indicates Send is the main verb or head of the sentence. There is an arc connecting Send to wife labeled IOBJ (Indirect Object). Wait, let me re-examine. The arc goes from Send to wife. Actually, looking at standard dependency parsing: Send connects to wife via an arc labeled IOBJ. This means my wife is the indirect object of Send. There is a large overarching arc connecting Send to text labeled DOBJ (Direct Object). This means a romantic text is the direct object. Inside the noun phrase a romantic text: There is an arc from text to a labeled DET (Determiner). Wait, usually arrows point from head to dependent. Let's look at the arrowheads. The arrow for DOBJ points from Send to text. So text depends on Send. The arrow for IOBJ points from Send to wife. So wife depends on Send. Now looking at the modifiers of text: There is an arc labeled DET connecting text and a. The arrowhead points to a. This means a depends on text. There is an arc labeled AMOD (Adjectival Modifier) connecting text and romantic. The arrowhead points to romantic. This means romantic depends on text. Wait, let me look really closely at the connections for a romantic text. There is a label DET above an arc. The arc connects text and a. The arrowhead points to a. So a modifies text. There is a label AMOD above an arc. The arc connects text and romantic. The arrowhead points to romantic. So romantic modifies text. Wait, let's re-evaluate the DET and AMOD arcs. The arc labeled DET starts at text (right side) and goes left to a. Arrow points to a. Correct. The arc labeled AMOD starts at text (right side, slightly higher up?) No, let's look at the nesting. Actually, looking very closely at the right side: There is a label DET above an arc connecting text and a. Arrow points to a. There is a label AMOD below DET. It connects romantic and... wait. Let's trace the lines carefully. Line 1 (Top right): From text to a. Label: DET. Arrow points to a. Line 2 (Middle right): From text to romantic. Label: AMOD. Arrow points to romantic. Wait, no. Let's look at the arrowheads again. The arc labeled DET connects text and a. The arrowhead is on a. The arc labeled AMOD connects romantic and... actually, it looks like it connects text to romantic. Let's trace the line. It starts near text, goes over romantic, and points to romantic. Wait, that would mean romantic depends on text. That makes sense (adjective modifying noun). BUT, look at the arc labeled DET. It connects a and... wait. Let's look at the standard Penn Treebank or similar dependency styles. Usually, heads point to dependents. Head: text. Dependents: a, romantic. Arc 1: From text to a. Label: DET. Arrow points to a. This fits. Arc 2: From text to romantic. Label: AMOD. Arrow points to romantic. This fits. Wait, let's look at the diagram again. The arc labeled DET connects text and a. The arrowhead is on a. The arc labeled AMOD connects... actually, looking closely, the arc labeled AMOD seems to connect romantic to something? No, that's unlikely. Let's re-examine the right side structure. There is a label DET. Below it is an arc connecting text and a. Arrow points to a. There is a label AMOD. Below that is an arc connecting... wait, looking at the arrowheads. The arrowhead for AMOD points to romantic. The line comes from text. So romantic modifies text. Wait, let's look at the nesting. The DET arc is above the AMOD arc? No, DET is above AMOD. Let's trace the lines from right to left. Line 1 (Topmost on right): From text to a. Label: DET. Arrow points to a. Line 2 (Middle on right): From text to romantic. Label: AMOD. Arrow points to romantic. Wait, looking at the image, there is an arc labeled AMOD that connects romantic and... actually, it looks like it connects text to romantic. Let's look really closely at crop 2 and 3. Crop 2 shows DET and AMOD. The arc labeled DET connects the word text (far right) and the word a (middle). Arrow points to a. The arc labeled AMOD connects the word romantic and... wait. The arrowhead is on romantic. Where does it start? It seems to start from text as well, but nested inside? Or does it start from a? No, that doesn't make sense grammatically. Let's look at the standard dependency for a romantic text. Head: text. Dependent 1: romantic (AMOD). Dependent 2: a (DET). In the diagram: There is a large arc labeled DOBJ connecting Send to text. This establishes text as the head of that phrase. From text, there are arcs going left. One arc goes to romantic. Label: AMOD. Arrow points to romantic. One arc goes to a. Label: DET. Arrow points to a. Wait, looking at the diagram, the arc labeled DET is above the arc labeled AMOD. And it connects text and a. Actually, let's look at the arrowheads again. The arrowhead for DET points to a. The arrowhead for AMOD points to romantic. Wait, looking at crop 2, there is an arc labeled AMOD that connects... actually, it looks like it connects romantic to text? No. Let's look at the full image again. Okay, let's trace from text. Arrow goes left-up to a. Label: DET. Wait, no. The label DET is above an arc. That arc connects text and a. Arrow points to a. There is another arc labeled AMOD. It connects... wait, looking at the arrowhead for AMOD. It points to romantic. Where does it start? It seems to start from text as well, but nested under the DET arc? Or does it connect romantic to something else? Actually, let's look at the nesting. The arc labeled AMOD is under the arc labeled DET. Wait, looking really closely at the right side... There is a label DET. Below it is an arc connecting text and a. Arrow points to a. There is a label AMOD. Below that is an arc connecting romantic and... wait. The arrowhead points to romantic. Where does the line come from? It seems to come from text. Let's reconsider the standard parsing. Usually, adjectives modify nouns directly. Determiners modify nouns directly. So text -> romantic (AMOD) and text -> a (DET). In the diagram: There is a label DOBJ connecting Send to text. This confirms text is the head. From text, an arc goes left to romantic. Label: AMOD. Arrow points to romantic. Wait, looking at crop 2, the arrowhead for AMOD points to romantic. The line comes from... actually, it looks like it connects romantic and text? No, dependency arrows usually point from Head -> Dependent or Dependent -> Head. In this diagram style (likely Stanford Dependencies), arrows point from the head to the dependent. So: Send is ROOT. Send -> wife (IOBJ). Arrow points to wife. Send -> text (DOBJ). Arrow points to text. Wait, looking at the big arc labeled DOBJ. It starts near Send and ends with an arrowhead pointing to text. Yes. So text is a dependent of Send. Now, what depends on text? There is an arc labeled DET. It connects text and a. Arrow points to a. So a depends on text. There is an arc labeled AMOD. It connects... wait. Looking at the diagram, there is an arc labeled AMOD that connects romantic and... actually, looking closely at crop 2, the arrowhead for AMOD points to romantic. The line seems to come from text. Wait, let's look at the nesting again. The arc labeled DET is above the arc labeled AMOD. Actually, looking very closely at the right side... There is a label DET. Below it is an arc connecting text and a. Arrow points to a. There is a label AMOD. Below that is an arc connecting romantic and... wait. The arrowhead points to romantic. Where does the line start? It looks like it starts from text as well, but nested inside? Or maybe it connects romantic to a? No, that's wrong grammar. Let's look at the arc labeled AMOD again. It connects romantic and... actually, looking at the full image, there is an arc connecting text and romantic. Label: AMOD. Arrow points to romantic. Wait, let's look at the arc labeled DET again. It connects a and... wait. The label DET is above an arc. That arc connects a and... actually, looking at crop 2, the arc labeled DET connects text and a. Arrow points to a. Wait, let's look at the arc labeled AMOD again. It connects romantic and... wait. The arrowhead points to romantic. Where does it start? It seems to start from text. Let's try a different interpretation. Maybe the arrows point from Dependent -> Head? If arrows point Dependent -> Head: wife -> Send (IOBJ). Arrow points to Send? No, arrow points to wife. So that hypothesis is wrong. text -> Send (DOBJ). Arrow points to text. So text depends on Send. This fits standard dependency parsing where arrows point Head -> Dependent? No, usually arrows point Head -> Dependent in some styles (like Stanford) or Dependent -> Head in others. Let's look at the arrowheads again. ROOT arrow points to Send. So ROOT is a label for the root node Send. DOBJ arc: Starts near Send, goes over everything, ends with arrow pointing to text. This implies text is the direct object of Send. IOBJ arc: Starts near Send, goes to wife. Arrow points to wife. This implies wife is the indirect object. DET arc: Starts near text, goes to a. Arrow points to This description was generated automatically. Please feel free to ask questions if you have further questions about the nature of the image or its meaning within the presentation.">

---

```
If $ROOT == "send"
	If $DOBJ == ("text" | "message" | "sms")
		$NUMBER = get.phonenumber($IOBJ)
		If $AMOD.DOBJ == "romantic":
			$MESSAGE = "omg ilu soooo much sweetie uwu 😍"
		If $AMOD.DOBJ == "mean":
			$MESSAGE = "I ate your leftover enchiladas and they were gooooood"
		SEND.SMS $message TO $number
```
---

## When parsing fails: A Case Study

---

### "Alexa, turn on the office lights and bedroom lights"

<p>(This audiovisual content has been removed for compliance with recent federal accessibility guidelines. <a href="https://accessibility.ucsd.edu/policies-standards/ucsd-accessibility-guidelines.html">Please see this site for details.</a>)</p>

---

### Alexa knows how to turn on lights

- "Turn on" takes a direct object (light) with modifiers (location)

```
If $ROOT == "turn on"
	If $DOBJ == "lights"
		ACTIVATE lights location $NMOD.DOBJ
```

---

### "Turn on the office lights"

<img class="r-stretch" src="diagrams/dep_turnonlight.jpg" alt="depturnonlight.jpg The user wants a detailed description of an image containing a syntactic dependency tree diagram. 1. Analyze the Image: Type: It's a linguistic diagram, specifically a dependency parse tree for a sentence fragment. Content: The text at the bottom reads Turn on the office lights. Structure: There are arrows connecting words to indicate grammatical relationships. Top left: ROOT with an arrow pointing down to Turn. This indicates Turn is the root of the sentence structure. From Turn: An arc labeled PRT (particle) points to on. This shows on is a particle modifying Turn. From Turn: A long, high arc labeled DOBJ (direct object) points all the way to lights. This indicates lights is the direct object of the verb Turn. From lights: An arc labeled NMOD (noun modifier) points to office. Wait, let me look closer. The arrow actually originates from lights and points to office? No, looking at standard dependency parsing conventions: Usually, modifiers point to the head. So office modifies lights. Let's trace the lines carefully. There is an arc labeled NMOD connecting office and lights. The arrowhead points from lights down to office? Or from office up? Actually, looking at the arrowheads: The line for NMOD connects office and lights. The arrowhead is on the left side of the arc, pointing towards office. Wait, no. Let's re-examine standard dependency trees (like Stanford Dependencies). In Stanford dependencies, arrows point from head to dependent. Let's look at the arrowheads in this specific image. Arrow from ROOT points DOWN to Turn. So Root -> Turn. Arc PRT: Starts near Turn, goes over on, arrow points TO on. Wait, usually particles are dependents of the verb. If the arrow points to on, then Turn is the head. Yes, that makes sense. Head (Turn) -> Dependent (on). Arc DOBJ: Starts near Turn, goes high over everything, arrow points TO lights. So Head (Turn) -> Dependent (lights). This fits Turn [something] lights. Arc DET: Starts near the, goes up and right? No. Let's look at the arc labeled DET. It connects the and office? Or the and lights? The arc labeled DET starts above the and ends with an arrow pointing to the. Wait, that would mean the is a dependent of something else. Let's look at the arc labeled NMOD. It connects office and lights. The arrow points from lights (right side) towards office (left side). This implies lights modifies office? That seems wrong linguistically (office lights). Usually, office modifies lights. Let's re-evaluate the direction of arrows based on standard visual conventions for these diagrams. Actually, looking very closely at the arrowheads: ROOT -> Turn (Downward). Correct. PRT arc: The line starts near Turn and curves down to on. The arrowhead is pointing AT on. This implies Turn governs on. DOBJ arc: Starts near Turn, goes high, ends at lights. Arrow points AT lights. This implies Turn governs lights. DET arc: Starts above the, curves down to the? No. Let's look at the line labeled DET. It connects the and... wait. The line labeled DET seems to connect the and office? Or the and lights? Let's trace the lines from the words upwards. From the: An arrow points UP and RIGHT towards a label? No, that's not right either. Let's look at the arc labeled DET. It connects the and... actually, it looks like it connects the to office? Or maybe the is the dependent of lights? Let's try the reverse: Arrows point from Dependent to Head (which is common in some treebanks like Universal Dependencies). If arrows point to heads: ROOT -> Turn. (Turn is root). PRT arc: Arrow points to on. So on depends on... something? No, usually PRT depends on the verb. If arrow points to on, then on is the head? That's unlikely for Turn on. Let's look at the arrowheads again very carefully. ROOT line: Arrowhead is at the bottom, pointing to Turn. So Root -> Turn. PRT arc: The line starts near Turn (left side of the curve) and ends with an arrowhead pointing down to on. Wait, looking really closely at crop 1... The arrowhead for PRT is on the right side of the curve, pointing DOWN to on. This implies Turn -> on. DOBJ arc: Starts near Turn, goes high, arrow points DOWN to lights. So Turn -> lights. DET arc: Starts above the, curves down... wait. The label is DET. The line connects the and... it looks like it connects the to office? No, that doesn't make sense. The office lights. The modifies lights (or the noun phrase). Let's look at the arc labeled NMOD. It connects office and lights. The arrow points from right to left? From lights to office? If so, lights -> office. That would mean lights modifies office. Office lights usually means lights for an office. So office is the modifier (noun adjunct). In dependency grammar, modifiers point to heads. So office should point to lights. Let's look at the arrowhead for NMOD again. It points from right to left? No, looking at crop 3, the arrowhead for NMOD is on the left side of the arc, pointing DOWN to office. Wait, if the line comes from lights, then lights -> office. Let's look at the DET arc again (Crop 2). The label is DET. The line connects... it looks like it connects the and office? Or the and lights? Actually, let's look at the standard Stanford Parser output style. In Stanford dependencies: root(ROOT, Turn). parataxis or prt(on, Turn). Wait, usually prt(Turn, on)? No, on is a dependent of Turn. So arrow from Turn to on. nsubj or dobj. Here DOBJ. Arrow from Turn to lights. det(the, lights). Arrow from lights to the. nn(office, lights) or amod. Here NMOD. Arrow from lights to office. Let's re-examine the image based on Arrows point from Head to Dependent vs Arrows point from Dependent to Head. If arrows point Head -> Dependent: ROOT -> Turn. (Turn is head). Turn -> on (PRT). Turn -> lights (DOBJ). lights -> the (DET)? No, usually Det depends on Noun. So Noun -> Det? Or Det -> Noun? In Universal Dependencies, arrows point from Head to Dependent. So lights -> the. And lights -> office. Let's look at the arrowheads in the image again. ROOT line: Arrow points DOWN to Turn. (Head -> Root? No, Root is a special node). It implies Turn is the root. PRT arc: The line starts near Turn and ends with an arrow pointing to on. Wait, looking at the full image... The arrowhead for PRT is on the right side of the curve, pointing DOWN to on. This confirms Head (Turn) -> Dependent (on). DOBJ arc: Starts near Turn, goes high. Arrow points DOWN to lights. Confirms Head (Turn) -> Dependent (lights). NMOD arc: The label is NMOD. The line connects office and lights. Where does it start? It seems to start above lights and curve left/down to office? Or start above office and curve right/down to lights? Let's look at the arrowhead for NMOD. In crop 3, there is an arrowhead pointing DOWN to office. The line comes from the right side (above lights). This implies Head (lights) -> Dependent (office). This fits office modifying lights. DET arc: The label is DET. The line connects... wait. In crop 2, there is an arrow pointing DOWN to the. The line comes from the right side (above office or lights?). It seems to come from above office but connect to the? No, that's weird. Let's look at the structure again. Words: Turn, on, the, office, lights. Arcs: 1. ROOT -> Turn (Vertical line). 2. PRT arc: Connects Turn and on. Arrow points to on. (Turn -> on). 3. DOBJ arc: Connects Turn and lights. Arrow points to lights. (Turn -> lights). 4. NMOD arc: Connects office and lights. Arrow points to office. Wait, if Head is lights, Dependent is office. So lights -> office. 5. DET arc: Connects... what? It looks like it connects the and... maybe lights? Or office? Let's look at the arrowhead for DET. In crop 2, there is an arrow pointing DOWN to the. The line comes from the right side (above office). This implies Head (office or lights) -> Dependent (the). Actually, looking really closely at the full image... There is a long arc labeled DET. It starts above the and goes to... wait. Let's trace from the words up. From the: An arrow points UP and RIGHT? No. From office: An arrow points UP and LEFT? Let's try to interpret the labels based on standard grammar. Sentence: Turn on the office lights. Root: Turn. PRT (Particle): on. (Dependent of Turn). DOBJ (Direct Object): lights. (Dependent of Turn). DET (Determiner): the. (Dependent of lights or office?). Usually the modifies the whole noun phrase office lights. So it depends on lights (head) or office (if office is head). In office lights, lights is the head noun. So the depends on lights. NMOD (Noun Modifier): office. (Dependent of lights). Office modifies lights. Let's map this to the visual lines: Line labeled DET: Connects the and... it looks like it connects to office? Or maybe lights? Let's assume standard dependency: det(lights, the). Arrow from lights to the. In the image, there is an arc labeled DET. It seems to connect the and... actually, looking at the arrowhead for DET (Crop 2), it points DOWN to the. The line originates from the right side, seemingly above office. This suggests Head (office or lights) -> Dependent (the). Line labeled NMOD: Connects office and lights. Arrow points DOWN to office. Wait. If NMOD is the label, usually it's nmod(lights, office). So arrow from lights to office. In the image, there is an arc labeled NMOD. The arrowhead points DOWN to office. The line originates from the right side (above lights). This fits perfectly: Head (lights) -> Dependent (office). Wait, let me look at the DET arc again. In the full image, there is an arc labeled DET. It connects... it looks like it connects the and office? No, that's unlikely. Let's look at the arrowhead for DET. It points DOWN to the. The line comes from the right side. It seems to originate above office. This would imply nmod(office, the)? No. Maybe the arc labeled DET connects the and lights? If Head is lights, Dependent is the. Arrow lights -> the. In the image, the arc labeled DET seems to span from above office down to the. This is confusing. Let's look at the arc labeled NMOD again. It connects office and lights This description was generated automatically. Please feel free to ask questions if you have further questions about the nature of the image or its meaning within the presentation.">

---

### "Turn on the office lights and turn on the bedroom lights"

<img class="r-stretch" src="diagrams/dep_turnonlightsfull.jpg" alt="depturnonlightsfull.jpg The user wants a detailed description of an image showing a syntactic dependency tree diagram. 1. Overall Structure: - It's a black and white line drawing on a white background. - The bottom row consists of words forming a sentence. - Above the words are curved lines (arcs) connecting them, labeled with grammatical tags. - There is a vertical arrow pointing down to the first word. 2. Text Content (The Sentence): - The words at the bottom read: Turn on the office lights and turn on the bedroom lights - Let's break it down by position from left to right: 1. Turn 2. on 3. the 4. office 5. lights 6. and 7. turn 8. on 9. the 10. bedroom 11. lights 3. Diagram Elements (Arrows and Labels): - ROOT: A vertical arrow points down to the word Turn (the first one). The label ROOT is above this arrow. - PRT (Particle) for first on: An arc connects Turn to on. Label: PRT. - DET (Determiner): An arc connects lights (first instance) back to the. Wait, let's look closer. The arc goes from lights to the. Actually, looking at standard dependency parsing, usually the head points to the dependent or vice versa. Here, arrows point from the modifier to the head? No, let's trace carefully. - Arrow from on (first) -> Turn. Label: PRT. So Turn is the head. - Arrow from the -> lights (first). Wait, no. The arrow starts at lights and points to the? Or starts at the and points to lights? Let's look at the curvature. - Actually, let's look at the labels. - DET is above an arc connecting the and office lights. No, that's not right either. - Let's trace from the top down. - CONJ: A long arc connects the first Turn to the second turn. Label: CONJ. This indicates coordination. - CC (Coordinating Conjunction): An arc connects and to the first Turn. Wait, usually CC connects the conjunction to the heads. Here, an arc goes from and up to... actually, let's look at the labels again. - There is a label CC above an arc connecting and and the first turn? No, that looks like it connects and to the first Turn. Let's re-evaluate the directionality. - In dependency trees, arrows usually point from head to dependent or dependent to head. - Let's look at DET. The arc is between the and lights. The label is DET. Usually, a determiner modifies a noun. So lights is likely the head, pointing to the? Or the points to lights? - Let's look at NMOD. Arc between office and lights. Label NMOD (noun modifier). Office modifies lights. - Let's look at the main structure. - ROOT points to Turn. This is the root of the sentence. - From Turn, there are arcs going out. - One arc goes to on (labeled PRT). - One arc goes to lights (first one, labeled DOBJ - Direct Object). Wait, let's trace that long arc. It starts at Turn and ends at the first lights. Label: DOBJ. This makes sense. Turn lights is the action-object pair. - From that first lights, there are modifiers attached. - An arc labeled DET connects to the. Wait, looking closely at the cluster around the office lights. - There is an arc from lights (first) going left/up to the? No, let's look at the labels relative to the arcs. - Label DET is above an arc connecting the and... actually it looks like it connects the to office lights. Let's assume standard parsing: lights is the head noun. the modifies it (DET). office modifies it (NMOD). - Let's trace the lines carefully. - Line from lights (first) goes left to the. Label above it is DET? No, that label is higher up. - Let's look at the arc labeled DOBJ. It connects Turn (root) to the first lights. So lights is the direct object of Turn. - From the first lights, there are arcs going back to its modifiers. - Arc labeled DET connects the and... wait, looking at the diagram, there is an arc from the curving up to office? No. - Let's look at the cluster: the office lights. - There is a label NMOD. An arc connects office and lights. This confirms office modifies lights. - There is a label DET. An arc connects the and... it seems to connect the to lights as well? Or maybe the to office? In standard dependency trees (like Stanford Dependencies), DET usually points from the noun to the determiner. So an arrow from lights to the. - Let's look at the diagram again. The arc labeled DET connects the and... actually, it looks like it connects the to office lights as a group? No, that's unlikely. - Let's assume the arrows point from Head -> Dependent or Dependent -> Head. - Let's look at the second half: turn on the bedroom lights. - DOBJ arc connects the first Turn (root) all the way to the last word lights. Wait, that would mean both objects are direct objects of the root verb. This is a common representation for coordinated verbs where the object applies to both? Or maybe it's a specific type of parsing. - Let's re-examine the DOBJ arc on the right. It connects the second turn (word 7) to the last lights (word 11). Wait, no. The arc starts near the first Turn and goes all the way to the last lights? No, that's too long. - Let's look at the right side DOBJ. It connects the second turn (word 7) to the last lights (word 11). Yes, that makes sense. Turn lights. - So we have two main structures coordinated by and. - Structure 1: Turn -> on (PRT), Turn -> lights (DOBJ). Inside lights: office (NMOD), the (DET). - Structure 2: turn -> on (PRT), turn -> lights (DOBJ). Inside lights: bedroom (NMOD), the (DET). Let's refine the specific connections based on visual tracing. - ROOT: Vertical arrow down to Turn (word 1). - CONJ: Long arc over everything connecting word 1 (Turn) and word 7 (turn). Wait, usually CONJ connects the conjunction to the heads. Here, the label is CONJ at the very top. The arc connects word 1 (Turn) and word 7 (turn). This implies they are conjuncts. - CC: Arc connecting and (word 6) to... it looks like it connects to the first Turn. Or maybe it's part of the CONJ structure. Actually, looking closely, there is an arc labeled CC that goes from and up to the first Turn. This marks and as a coordinating conjunction linking the two verbs. - DOBJ (Left): Arc connecting word 1 (Turn) to word 5 (lights). Label DOBJ. - From word 5 (lights): - Arc labeled DET connects to word 3 (the). Wait, looking at the lines... there is an arc from lights curving left to the. The label DET is above it. - Arc labeled NMOD connects to word 4 (office). The label NMOD is above it. This means office modifies lights. - DOBJ (Right): Arc connecting word 7 (turn) to word 11 (lights). Label DOBJ. - From word 11 (lights): - Arc labeled DET connects to word 9 (the). - Arc labeled NMOD connects to word 10 (bedroom). This means bedroom modifies lights. - PRT (Left): Arc connecting word 2 (on) to word 1 (Turn). Label PRT. - PRT (Right): Arc connecting word 8 (on) to word 7 (turn). Label PRT. Summary of the diagram logic: This is a dependency parse tree for the sentence Turn on the office lights and turn on the bedroom lights. - The root verb is the first Turn. - It has two direct objects: lights (office) and lights (bedroom). This is a specific parsing style where the object of the second coordinated verb is linked to the root, or perhaps it's showing that both are objects of the coordinated structure. Wait, let's look at the right DOBJ arc again. - The right DOBJ arc starts from... actually, looking very closely at the right side, the arc labeled DOBJ connects the second turn (word 7) to the last lights (word 11). Yes, that is the most logical syntactic relationship. - Wait, let's look at the left DOBJ arc again. It connects word 1 (Turn) to word 5 (lights). - So we have: - Root: Turn (1) - PRT: on (2) -> linked to Turn (1). - DOBJ: lights (5) -> linked to Turn (1). - DET: the (3) -> linked to lights (5). - NMOD: office (4) -> linked to lights (5). - CC: and (6) -> linked to Turn (1)? Or is it part of the CONJ arc? The label CC is above an arc connecting and to... actually, looking at the top arcs. - There is a big arc labeled CONJ connecting word 1 (Turn) and word 7 (turn). This indicates they are conjuncts (coordinated verbs). - There is a label CC above an arc connecting and to... it looks like it connects and to the first Turn. This marks and as the conjunction. Let's re-read the diagram carefully to ensure accuracy. - Topmost arc: Label CONJ. Connects word 1 (Turn) and word 7 (turn). - Below that: Label CC. Arc connects word 6 (and) to... it looks like it connects to word 1 (Turn). This is slightly non-standard (usually CC points to the head of the conjunct), but I must describe what is there. It looks like an arc from and going up and left to Turn. - Below that: Label DOBJ. Arc connects word 1 (Turn) to word 5 (lights). - From word 5 (lights): - Arc labeled DET goes to word 3 (the). Wait, looking at the lines... there is an arc from the going up to... actually, it looks like the arc for DET connects the and office lights? No. - Let's look at the cluster the office lights. - There is a label DET. The arc seems to connect the (word 3) and... maybe lights (word 5)? Yes, standard dependency: Noun -> Det. So arrow from lights to the. - There is a label NMOD. Arc connects office (word 4) and lights (word 5). Arrow from lights to office? Or office to lights? Usually modifier points to head. So office -> lights. - Below that: Label PRT. Arc connects word 2 (on) to word 1 (Turn). Now the right side (second clause): - Word 7 is turn. - Word 8 is on. - Word 9 is the. - Word 10 is bedroom. - Word 11 is lights. - Arc labeled DOBJ connects word 7 (turn) to word 11 (lights). - From word 11 (lights): - Arc labeled DET connects to word 9 (the). - Arc labeled NMOD connects to word 10 (bedroom). - Arc labeled PRT connects word 8 (on) to word 7 (turn). Wait, let's look at the connection between the two halves. - The big arc labeled CONJ connects Turn (word 1) and turn (word 7). This confirms they are coordinated verbs. - The label CC is above an arc connecting and (word 6) to... it looks like it connects to the first Turn. Let's refine the description of the text at the bottom. The words are spaced out: Turn, on, the, office, lights, and, turn, on, the, bedroom, lights. Final check of labels and connections: - ROOT: Vertical arrow down to Turn (1st word). - CONJ: Long arc over the top connecting 1st Turn and 2nd turn. - CC: Arc connecting and to 1st Turn. (Actually, looking really closely, the CC label is above an arc that connects and to... it's hard to tell if it goes to Turn or if it's just floating. But syntactically, CC links the conjunction to the head. So likely and -> Turn). - DOBJ (Left): Arc from 1st Turn to 1st lights. - DET: Arc connecting 1st the and 1st lights. (Actually, looking at the lines, there is an arc from the curving up to... wait. Let's look at the label DET above office lights. The arc connects the and... it looks like it connects the to office? No. - Let's assume standard Stanford Dependencies for this sentence. - Root: Turn (1) - dobj(lights, Turn) -> Arrow from lights to Turn? Or Turn to lights? The arrow head is at lights. So Turn -> lights. - nmod(office, lights) -> Arrow from office to lights? Or lights to office? The label NMOD is above an arc connecting office and lights. - det(the, lights) -> Arrow from the to lights? Or lights to the? - Let's just describe the visual connections without assuming directionality if it's ambiguous, but usually arrows point to the head. - In this diagram: - The arrow for ROOT points TO Turn. So Turn is the head This description was generated automatically. Please feel free to ask questions if you have further questions about the nature of the image or its meaning within the presentation.">

---

### "Turn on the office lights and bedroom lights" (expected)

<img class="r-stretch" src="diagrams/dep_turnonlights.jpg" alt="depturnonlights.jpg The user wants a detailed description of an image showing a syntactic dependency tree diagram. 1. Analyze the Image: Type: It's a linguistic diagram, specifically a dependency parse tree for a sentence. Text at the bottom (the sentence): Turn on the office lights and bedroom lights Words: Turn, on, the, office, lights, and, bedroom, lights. Labels/Arrows: There are curved arrows connecting words to indicate grammatical relationships. Each arrow has a label above it. ROOT: A straight vertical line pointing down to Turn. This indicates Turn is the root of the sentence (the main verb). PRT: An arc from Turn to on. This stands for Particle, indicating on is a particle part of the phrasal verb turn on. DOBJ1: An arc from Turn to lights (the first one). This stands for Direct Object 1. Wait, let me look closer. The arrow starts at Turn and goes all the way over to the word lights (the second 'lights' in the sequence? No, let's trace carefully). Let's re-trace the arcs from left to right based on their start points. ROOT: Points straight down to Turn. PRT: Arcs from Turn to on. DOBJ1: Arcs from Turn (implied, though it looks like it starts near the PRT arc) to the first lights. Wait, looking at the nesting: There is an arc labeled DET connecting the to office. No, that's wrong. Let's look at the arrows carefully. Arrow from the points to office? No. Let's trace from the words up. Turn: Has an arrow coming down (ROOT). Has an arc going right labeled PRT pointing to on. Has a long arc going over everything labeled DOBJ1? No, let's look at the labels again. Actually, usually in these diagrams, the label is on the arc connecting the head to the dependent. Let's trace the arcs from the top down. ROOT: Vertical line pointing to Turn. PRT: Arc connecting Turn and on. DOBJ1: Arc connecting Turn (or near it) to... wait, let's look at the nesting structure. The arc labeled DET connects the to office? No, that looks like an arc from the pointing to office. Wait, usually determiners modify nouns. So the modifies lights? Or office? Let's look at the standard dependency parsing style (like Stanford Dependencies). Turn is the root verb. on is a particle (prt) of turn. lights (the first one) is a direct object (dobj). The label says DOBJ1. So there's an arc from Turn to lights. Let's trace that specific arc. It starts near the PRT arc and ends at the word lights. Yes, that makes sense. Inside that DOBJ1 relationship: There is an arc labeled DET connecting the to... wait. The arrow head points from office to the? No. Let's look at the arc labeled DET. It connects the and lights. Wait, no. Let's look at the arc labeled NMOD. It connects office and lights. This makes sense: office lights. So office modifies lights. Let's look at the arc labeled DET. It connects the to... it seems to point to office? Or does it point to lights? In standard parsing, determiners modify nouns. Here, the likely modifies the whole noun phrase office lights. But looking at the arrow head, it points from the towards office? No, that's weird. Let's re-examine the arcs in the middle cluster: the office lights. Arc NMOD: Connects office and lights. The arrowhead is on lights. So office modifies lights. This is a noun modifier. Arc DET: Connects the to... it looks like it points to office? Or maybe lights? Let's assume standard grammar: the modifies the head noun lights. But visually, the arc labeled DET connects the and office? No, looking closely at crop 4 and 5... Actually, let's look at the arrow heads. Arc DET: Starts at the, goes over office, ends at lights? Or starts at office and points to the? Let's look at the arc labeled NMOD inside the DET arc. It connects office and lights. Arrow points from office to lights. So office modifies lights. The arc labeled DET: Connects the to... it looks like it points to office? No, that would be weird. Let's look at the arrow head. It points down to office. Wait, if DET points to office, then the modifies office. That is grammatically possible (the office lights). Actually, let's look at the arc labeled DOBJ1. It connects Turn (implied start) to... wait. Let's try a different hypothesis. Maybe the labels describe the relationship of the word below the label to the word it points to? Or vice versa? Standard dependency trees usually have the head at the bottom and arrows pointing from Head -> Dependent or Dependent -> Head. In this diagram: ROOT points to Turn. So Turn is the root. PRT arc connects Turn and on. Arrow points to on. So Turn -> on (particle). DOBJ1: This label is on a long arc. Where does it start? It seems to start near Turn and end at the first lights. Wait, looking at the nesting... Let's look at the arc labeled DET. It connects the and office? No. Let's look at the arc labeled NMOD (under DET). It connects office and lights. Arrow points to lights. So office modifies lights. The arc labeled DET: Connects the and... it looks like it points to office? Or maybe lights? Let's assume it points to the head of the phrase. If office lights is a noun phrase, lights is the head. So DET should point to lights. But visually, the arc labeled DET seems to encompass the office. Actually, let's look at the arrow heads very carefully. Arc NMOD: From office to lights. (Office modifies lights). Arc DET: From the to... it looks like it points to office? No, that would mean the modifies office. Then office lights is a compound noun? Or the office modifies lights? Let's look at the arc labeled DOBJ1. It connects Turn (start) to... wait. The label DOBJ1 is on an arc that goes from Turn area to... lights. Wait, there are two arcs labeled DOBJ? No, DOBJ1 and DOBJ2. Let's trace DOBJ1. It starts near the PRT arc (so from Turn) and ends at the first lights. So Turn -> lights (direct object). This makes sense. Inside that relationship, we have modifiers for lights. Modifier 1: office. Connected via NMOD (noun modifier). Arrow from office to lights. Modifier 2: the. Connected via DET. Where does it point? It looks like it points to office? Or maybe lights? If it's a standard tree, determiners attach to the noun. So DET should point to lights. But visually, the arc labeled DET is above the NMOD arc. The arrow head for DET seems to point to office. Wait, if DET points to office, then the modifies office. Then we have the office modifying lights? That's a bit unusual but possible (e.g., the [office lights]). Or maybe it's the [office] lights where office is the head and lights is a modifier? No, office lights usually means lights for an office. So lights is the head. Let's reconsider the arrow heads. Arc DET: Arrow points to office. This implies the modifies office. Arc NMOD: Arrow points from office to lights? No, looking at crop 5, the arc labeled NMOD connects office and lights. The arrow head is on lights. Wait, if DET points to office and NMOD points to lights... that's a chain. the -> office -> lights. Actually, let's look at the arc labeled DET again. It connects the and office? No, it looks like it connects the and lights but passes over office. Wait, no. Let's look at the arc labeled DOBJ1. It connects Turn to... wait. The label DOBJ1 is on an arc that goes from Turn (start) to... lights? No, looking at the right side of that arc, it points to and. Wait. Let's restart the tracing. This is crucial. Re-evaluating the arcs: ROOT: Points down to Turn. Correct. PRT: Arc from Turn to on. Arrow points to on. Correct. DOBJ1: Arc starting near Turn, going over the office lights, ending at... wait. The arrow head is on the word and? No, that would be weird for a direct object. Let's look at the arc labeled CC. That stands for Coordinating Conjunction. It connects lights (first one) to and? Or Turn to and? Let's look at the arc labeled DOBJ2. This is the topmost long arc. It starts near Turn and ends at the last word, lights (the second one). So Turn -> lights (second one) is a direct object relationship. This makes sense for a compound object: Turn on [office lights] and [bedroom lights]. So, DOBJ2 connects Turn to the final lights. Now let's look at CC. It connects... it looks like it connects the first lights to and? Or does it connect Turn to and? The label is CC (Coordinating Conjunction). Usually, this links two clauses or phrases. Here it likely links the two noun phrases. The arc labeled CC seems to start near the first lights and end at and. Or maybe from Turn to and? No, usually conjunctions link sisters. Let's look at the arrow head for CC. It points to and. Where does it start? It starts near the first lights. So lights (first) -> and. This indicates coordination. Now let's look at DOBJ1 again. If DOBJ2 connects Turn to the second lights, then DOBJ1 must connect Turn to the first object phrase. The arc labeled DOBJ1 starts near Turn and ends at... wait. Looking closely at crop 2 and 3... There is an arc labeled DET. It connects the to... it looks like it points to office. Wait, let's look at the arrow head. It points down to office. There is an arc labeled NMOD (under DET). It connects office to lights? No, looking at crop 5, the arc labeled NMOD connects office and lights. The arrow head points to lights. Wait, if DET points to office and NMOD points to lights... Let's look at the arc labeled DOBJ1 again. It seems to connect Turn (start) to... wait, is it connecting to the first lights? Yes, that makes sense. The label DOBJ1 is on an arc that goes from the root area to the word lights (the first one). So, structure so far: Root: Turn PRT: on (attached to Turn) DOBJ1: lights (first one) (attached to Turn) Inside DOBJ1 phrase (the office lights): DET: the attached to... wait. The arc labeled DET connects the and office? Or the and lights? Let's look at the arrow head for DET. It points to office. This implies the modifies office. NMOD: Connects office to... wait, looking at crop 5, there is an arc labeled NMOD connecting office and lights? No, that's not right either. Let's look really closely at the cluster the office lights. Word: This description was generated automatically. Please feel free to ask questions if you have further questions about the nature of the image or its meaning within the presentation.">

---

### "Turn on the office lights and bedroom lights" (actual)

<img class="r-stretch" src="diagrams/dep_turnonlightsbad.jpg" alt="depturnonlightsbad.jpg The user wants a detailed description of an image showing a syntactic dependency tree diagram. 1. Analyze the Image: Type: It's a linguistic diagram, specifically a dependency parse tree for the sentence Turn on the office lights and bedroom lights. Visual Style: Black text and lines on a white background. Minimalist. Components: Words (Tokens): Turn, on, the, office, lights, and, bedroom, lights. These are arranged horizontally at the bottom. Arrows/Lines: Curved arrows connecting words to indicate grammatical relationships. Labels: Text labels above the arrows indicating the type of relationship (e.g., ROOT, PRT, DOBJ, DET, NMOD, CONJ, CC). 2. Breakdown by Word and Relationships (Left to Right): Turn: This is the main verb. There is a vertical arrow pointing down from the label ROOT. This indicates it's the root of the sentence. An arrow curves up from on back to Turn, labeled PRT (Particle). This shows on is a particle associated with Turn. on: A preposition/particle here. Connected to Turn via the PRT link mentioned above. the: A determiner. An arrow curves from the up and over to lights (the first one). The label is DET. This means the modifies lights. Wait, let me look closer. Actually, looking at the nesting: Turn connects to the final lights via a long arc labeled DOBJ (Direct Object). So Turn on... lights is the structure. Let's re-evaluate the arcs from top to bottom. 3. Re-evaluating Arcs (Top to Bottom): Longest Arc: Starts at Turn, goes all the way over to the last word lights. Label: DOBJ. This means the entire phrase on the office lights and bedroom lights acts as a phrasal verb object, or more likely in this specific dependency style (likely Stanford Dependencies), the root Turn connects directly to the head noun of the object. Wait, usually Turn on is the verb. Let's look at standard dependency parsing for Turn on. Actually, looking at the diagram: The arrow labeled DOBJ starts at Turn and ends at the final word lights. This implies Turn governs lights. Second Longest Arc: Starts at the, goes to the first lights. Label: DET. Wait, no. Let's trace carefully. The arrow labeled DET starts above the and points to... actually, it looks like it connects the to the first lights. No, that doesn't make sense with the nesting. Let's look at the labels again. Label DET: Connects the to... wait, let's trace the line from DET. It goes down and points to the word office? No, that would be weird. Let's look at the arrow heads. Arrow from DOBJ (topmost) points to the last word lights. The tail is near Turn. So Turn -> DOBJ -> lights. Arrow from DET: Points to... actually, it looks like it connects the to office? No. Let's look at the arrow head for DET. It points to office. Wait, the office lights. Usually the modifies lights. But in some parsers (like Stanford), determiners modify the noun they precede if it's a compound noun phrase? Or maybe it modifies office? Let's look at the arrow labeled NMOD (first one down). It connects office to... the first lights. So office modifies lights. This makes sense: office lights. Let's look at the arrow labeled DET. It connects the to... actually, looking closely at the arrow head for DET, it points to office. Wait, is that right? The office lights. If office modifies lights, then the usually modifies lights too. Let's re-examine the lines. Line DET: Starts above the, curves down, arrow head points to... actually, it looks like it points to office. This is a bit unusual for standard English grammar (usually Det -> Noun), but in dependency trees, sometimes determiners attach to the head of the noun phrase. If lights is the head, the should point there. Let's look at the arrow labeled NMOD (second one down). It starts above office, curves down, points to... wait. Let's trace from the words up. Word: lights (first one). Arrows pointing TO it: One from office (labeled NMOD), one from the (labeled DET)? No, let's look at the arrow heads again. Arrow labeled NMOD (top one of the middle group): Starts above office, points to... actually, it looks like it points to the first lights. So office modifies lights. Arrow labeled DET: Starts above the, points to... wait, looking very closely at crop 2 and 5. The arrow for DET seems to point to office. This implies the modifies office. Then office modifies lights. So the office is a unit modifying lights? Or maybe the parser treats office lights as a compound where office is the head? No, usually lights is the head. Let's look at the arrow labeled NMOD (bottom one of that group). It starts above... wait, there are two NMOD labels. Let's trace carefully. Top arc: DOBJ. Connects Turn to last lights. Next arc down: DET. Connects the to... actually, looking at the arrow head, it points to office. Wait, no. The arrow head for DET is pointing to the word office. Let's assume that for a moment. Next arc down: NMOD. Connects office to... wait, there are two NMOD labels stacked. Top NMOD label: Arrow connects office to... actually, let's look at the arrow heads again. There is an arrow from the labeled DET pointing to office. There is an arrow from office labeled NMOD pointing to lights. Wait, there are two NMOD labels. Let's count the arcs between DET and CONJ. Arc 1 (top): Label NMOD. Tail near office, head near first lights. Arc 2 (bottom): Label NMOD. Tail near... wait, where does it start? It seems to start above bedroom? No, that's further right. Let's restart the tracing from the bottom words up, as that's often easier. Words: Turn, on, the, office, lights, and, bedroom, lights. Turn: Root. on: Connected to Turn via PRT. (Particle). the: Connected to... let's look at the arrow labeled DET. It points to office. This is a specific parsing choice (perhaps treating office lights as a compound noun where office is the head, or just a quirk of this specific parser like Stanford CoreNLP). Wait, actually, looking really closely at the arrow for DET... it points to office. office: Connected to first lights. There are two arrows pointing from office area? No. Let's look at the label NMOD (top one). The line goes from office to the first lights. So office modifies lights. Wait, there is another NMOD label below it. Where does that line come from? It seems to come from... actually, looking at the layout, it looks like: Line from the labeled DET points to office. Line from office labeled NMOD points to first lights. BUT, there are two NMOD labels. Let's look at the lines again. Ah, I see. The label DET is above a line that connects the to... actually, looking at the arrow head for DET, it points to office. The label NMOD (top) is above a line connecting office to... wait, no. Let's look at the lines originating from the right side. There are multiple lines converging on the last word lights. Topmost line: From Turn, labeled DOBJ. Next line down: From bedroom? No, let's trace back. Let's look at the label NMOD (top one). The line connects... actually, it looks like it connects office to the first lights. Let's look at the label NMOD (bottom one). The line connects... wait, is there a word missing? No. Let's try a different hypothesis. Maybe the labels correspond to specific arcs. Arc 1 (Top): Label DOBJ. Connects Turn to last lights. Arc 2: Label DET. Connects the to... actually, looking at the arrow head, it points to office. This is weird. Usually Det -> Noun. Maybe office is the noun here? Office lights -> Lights is the head? If lights is the head, Det should point there. Let's look at the arrow labeled NMOD (top). It connects... actually, it looks like it connects the to office? No, that's DET. Let's look at the cluster of lines between DET and CONJ. There is a line from the labeled DET. Where does it point? It points to office. There is a line from office labeled NMOD (top). Where does it point? It points to the first lights. Wait, there are TWO NMOD labels. Top NMOD: Connects... actually, looking at the arrow head, it points to the first lights. The tail is near office. Bottom NMOD: Connects... wait, where does the tail come from? It looks like it comes from... actually, let's look at the word bedroom. Let's trace the right side of the diagram. Last word: lights. Second to last word: bedroom. Word before that: and. Word before that: lights (first one). Okay, let's look at the connections for bedroom lights. There is an arrow labeled NMOD (bottom one of the middle group). It connects... actually, it looks like it connects bedroom to the last lights? No, that would be weird if and is involved. Let's look at the label CONJ. It connects and to... wait. Label CC (Coordinating Conjunction). Connects and to... actually, usually CC modifies the conjunction itself or links clauses. Here, it seems to connect and to something. Let's look at the label CONJ. It connects... wait, let's trace the line labeled CONJ. It starts above lights (first one) and goes to bedroom? No. Let's try to interpret standard dependency trees for this sentence: Turn on the office lights and bedroom lights. Root: Turn. PRT: on (attached to Turn). DOBJ: lights (the head of the object phrase). Which one? Usually the last one in a coordination, or the first. Here, the long arc goes to the last lights. So Turn -> DOBJ -> lights (last). Coordination: office lights and bedroom lights. Head of coordination is usually the head of the phrase. If lights is the head, then we have two instances. Let's look at the arcs again. Arc labeled NMOD (top): Connects office to... wait, looking really closely at crop 2 and 5. The arrow for NMOD (top) points to the first lights. The tail is near office. So office modifies lights. Arc labeled DET: Points from the to... actually, it looks like it points to office. This implies the office is a unit. Then office modifies lights. This is a valid parse in some systems (e.g., Stanford where Det attaches to the noun it precedes if that noun is part of a compound). Arc labeled NMOD (bottom): Connects... wait, let's look at the line under the top NMOD. It connects bedroom to... actually, looking at the arrow head for the bottom NMOD, it points to the last lights. Wait, no. Let's look at the label CONJ. It is above an arc connecting and to... wait. Let's look at the label CC. It connects and to... actually, usually CC This description was generated automatically. Please feel free to ask questions if you have further questions about the nature of the image or its meaning within the presentation.">

"I didn't find a device named 'office lights bedroom lights'"

---

### Alexa didn't parse the syntax, and failed

- ... in a way that a human would never fail

- This is hard, but it's fascinating

---

## So, we need to be able to parse syntax to do NLP

- ... and although Linguists tend to use constituency, NLP is dependency-based

---

### Wrapping up

- Although most linguists do theory with constituency, in NLP, dependency is king

- Dependency parsing focuses on finding links between words and their types

- Dependency parsing has many advantages over constituency in NLP

- Dependency Parsing is useful for actual engineering tasks!

---

### For Next Time

- What does it all mean?

---

<huge>Thank you!</huge>