Module

ANS: 1) Number Agreement

This rule says that both words must match in number (singular/plural) to refer to
the same entity.

Explanation (easy):

●​ A singular noun can only co-refer with a singular pronoun.

●​ A plural noun can only co-refer with a plural pronoun.

Examples:

●​ ✔ The girl said she is coming. → girl (singular) = she (singular)

●​ ✘ The girls said she is coming. → mismatch (plural vs singular)

So, co-reference is allowed only when the number matches.

2) Person and Case Agreement

This rule says that the pronoun must match the noun in:

●​ Person (1st, 2nd, 3rd)

●​ Case (subjective, objective, possessive)

Explanation (easy):

(a) Person Agreement

●​ 1st-person nouns refer using I / me / my

●​ 2nd-person nouns refer using you
●​ 3rd-person nouns refer using he / she / it / they

Example:

●​ ✔ Riya said she will come.

●​ ✘ Riya said I will come. → wrong person (3rd person noun vs 1st person
pronoun)

(b) Case Agreement

Pronouns must match grammatical position:

 ●​ Subject case: he, she, I, they
●​ Object case: him, her, me, them
●​ Possessive case: his, her, my, their

Examples:

●​ ✔ Rohan said he will help. → subject = subject

●​ ✔ The teacher praised Rohan because him → object = object
●​ ✘ The teacher praised Rohan because he → wrong case

ANS: Anaphora Resolution

Anaphora resolution means finding the correct noun (antecedent) for a pronoun.​
Example: Riya went home because she was tired.​
→ “she” refers to “Riya”.

Two common algorithms used:

1) Hobbs Algorithm (Syntax-based Method)

Hobbs algorithm uses parse trees to find the antecedent of a pronoun by searching
in a left-to-right, breadth-first manner.

Steps (very easy version):

1.​ Start at the NP node that dominates the pronoun.

2.​ Go up the parse tree to the nearest S (sentence) node.
3.​ Search left side first: Perform a left-to-right, breadth-first search of all NP
nodes to the left of the pronoun.
4.​ If a suitable NP is found → select it as antecedent.
5.​ If not found, move to the next higher S node and repeat.
6.​ Continue until an antecedent is found.

Idea in simple words:

The algorithm prefers earlier NPs in the sentence and checks the tree
structurally.

Example:

Sentence: Riya met Sona when she was leaving.​

Hobbs typically selects Riya as antecedent before Sona because it appears earlier
in the tree.

2) Centering Algorithm (Discourse-based Method)

Centering Theory focuses on coherence of multiple sentences.​

It finds which entity is the “center” (main topic) and uses that to resolve pronouns.

Key Concepts (easy):

●​ Cf (Forward centers): List of important NPs in the sentence.

●​ Cb (Backward center): The NP from previous sentence that the current
sentence continues to talk about.
●​ Cp (Preferred center): The most important NP in the current sentence.

Steps (simple version):

1.​ Identify all noun phrases in each sentence.

2.​ Rank them (subject > object > others).
3.​ Select the previous sentence’s main topic (Cb).
4.​ The pronoun in the next sentence usually refers to the Cb or Cp.
5.​ Choose the antecedent that maintains maximum coherence across
sentences.

Idea in simple words:

Centering chooses the antecedent that makes the story flow smoothly, keeping the
same topic.

Example:

Sentence 1: Rohan went to the market. (Rohan = center)​

Sentence 2: He bought fruits.​
→ “He” refers to Rohan because it maintains topic continuity.

What is Reference Resolution?

Reference Resolution is the task in NLP of identifying what a word or phrase

refers to in a text.​
It is needed because natural language uses many referring expressions like:

●​ Pronouns: he, she, it, they

●​ Definite Noun Phrases: the boy, that car
 ●​ Demonstratives: this, that, these
●​ Proper nouns: Riya, India
●​ Zero references (omitted subjects) in some languages

Goal:

To find the correct antecedent (the noun that the reference refers to).

Example:​
Riya saw Tanya and she smiled.​
→ “she” refers to Riya (Reference Resolution finds this link).

🧠 Why Reference Resolution is Needed?

●​ To understand who is doing what in the text
●​ For machine translation
●​ For question answering
●​ For summarization
●​ For dialogue systems
●​ For removing ambiguity

🧩 Types of References
1.​ Anaphora: Reference to something mentioned earlier
○​ Ravi bought a car. He loves it.
2.​ Cataphora: Reference to something mentioned later
○​ Before he spoke, Rohan smiled.
3.​ Coreference: Two or more expressions refer to the same entity
○​ “John”, “he”, “the boy” → same person
4.​ Bridging Reference: Indirect relation
○​ I bought a laptop. The keyboard is amazing.

💡 How Reference Resolution Works

Reference resolution typically uses syntactic, semantic, and world knowledge
constraints.

1️⃣ Syntactic Constraints

These ensure grammatical correctness.

1. Number Agreement
A pronoun must match the number of its antecedent.

●​ Singular → singular
●​ Plural → plural

Example:​
The boy lost his bag. (✔)​
The boys lost his bag. (✘)

2. Gender Agreement

Pronoun must match gender.

●​ Riya said she would come.

3. Person and Case Agreement

●​ I / you / he / she / they

●​ Subject pronouns must refer to subjects; object pronouns to objects.

Example:​
Rohan scolded Raj because he was late.​
Ambiguous → algorithms use more constraints to decide.

4. Binding Constraints

e.g., reflexive pronouns must refer to a subject:

●​ Riya hurt herself. (herself = Riya)

2️⃣ Semantic Constraints

These relate to meaning.

1. Selectional Restrictions

Some verbs accept only certain types of objects.

Example:​
The bottle ate the food. (illogical)​
→ "bottle" cannot be semantic antecedent for actions like eating.

2. World Knowledge

Humans use common sense; systems attempt to model this.

Example:​
The dog chased the cat because it was scared.​
→ cat is scared → system uses world knowledge.

⚙️ Approaches for Reference Resolution

1) Rule-Based Approaches

Use grammar + hand-crafted rules.​

Example: Hobbs Algorithm

✔ Advantages: simple, fast

✘ Limitations: fails on complex cases

2) Machine Learning Approaches

Learn features from annotated data.

Types:

●​ Supervised models (Decision tree, SVM, Neural models)

●​ Semi-supervised models
●​ Unsupervised models

Features used:

●​ Syntactic cues
●​ Grammatical roles
●​ Gender/Number
●​ Semantic compatibility
●​ Distance

3) Deep Learning Approaches

●​ Neural Coreference Models (e.g., BERT-based models)

●​ End-to-end systems

These models capture context and semantics better.

Pronoun resolution is difficult because pronouns (he, she, it, they, this, that) do not
carry complete meaning.​
 To correctly identify who or what a pronoun refers to, NLP systems study
different reference phenomena.

Below are the main reference phenomena with easy examples that illustrate how
pronoun problems are solved:

1) Anaphora

A pronoun refers to something mentioned earlier in the sentence.

Example:

Riya met Sona when she was leaving.​

→ “she” = Riya (antecedent)

Why it matters:

Systems must look backward to find the correct noun.

2) Cataphora

A pronoun refers to something that appears later.

Example:

Before he spoke, Rohan took a deep breath.​

→ “he” = Rohan (antecedent appears later)

Why it matters:

Systems must check forward in the sentence.

3) Coreference

Multiple expressions refer to the same real-world entity.

Example:

John arrived late. The boy looked tired. He apologized.​

→ John = the boy = he​
→ All refer to the same person.

Why it matters:
Helps in building coreference chains.

4) Bridging Reference

A pronoun or phrase refers to something indirectly connected to a previous noun.

Example:

I bought a car. The engine is powerful.​

→ “engine” is not explicitly mentioned but is part of the car.

Why it matters:

Uses world knowledge and associations.

5) Exophora

Pronoun refers to something outside the text (context-dependent).

Example:

Put that there.​

→ Needs physical context, not available in the sentence.

Why it matters:

Hard for systems because information is not in the text.

6) Zero Anaphora (implicit reference)

The subject or object is missing but understood (common in some languages).

Example (English):

Went to the store. Bought fruits.​

(implied “I”)

Why it matters:

The system must detect hidden pronouns.

🧠 How These Phenomena Help Solve the Pronoun Problem

By identifying the type of reference, NLP systems can:

●​ Look backward or forward in the sentence (Anaphora/Cataphora)

●​ Build entity chains (Coreference)
●​ Use semantic links (Bridging)
●​ Understand implicit meanings (Zero Anaphora)
●​ Handle pronouns referring outside text (Exophora)
●​ This allows accurate pronoun resolution in tasks like translation,
summarization, and question answering.

Reference resolution becomes difficult because not all pronouns refer to a simple,
clear noun.​
The following three types of referents make the problem more complex:

1) One-Antecedent Referents (Simple Coreference)

These are pronouns that refer to one single noun phrase, but the sentence structure
creates confusion.

Why it complicates resolution:

●​ Many nouns may match the pronoun in number/gender.

●​ The algorithm must choose the correct one.

Example:

Riya scolded Sona because she was angry.​

→ “she” could refer to Riya or Sona → ambiguous.

Problem:

Syntactic cues (gender, number) are not enough.​

Semantic understanding is also required.

2) Event or Clause Referents (Non-NP Antecedents)

Here, a pronoun refers not to a noun, but to an entire event, action, or clause.

Why it complicates resolution:

Algorithms typically look for noun phrases, but here the referent is not a noun.
Example:

Rohan failed the test. This surprised everyone.​

→ “This” refers to the whole event — Rohan failing, not a single word.

Problem:

Requires understanding of context and events, not just grammar.

3) Bridging (Indirect) Referents

The pronoun or noun refers to something that is not explicitly mentioned but is
implicitly related.

Why it complicates resolution:

●​ Not directly stated → requires world knowledge.

●​ The referent is connected through a part-whole or associative relationship.

Example:

I bought a car yesterday. The engine is very smooth.​

→ “engine” is not mentioned before, but is understood as part of the car.

Problem:

Need commonsense knowledge to infer such relations.

Discourse reference resolution is the task of identifying which entity, event, or

object a word or phrase refers to across multiple sentences in a text (a discourse).

It extends normal reference resolution beyond a single sentence and focuses on

maintaining coherence across an entire paragraph or conversation.

Example:

Riya went to the market. She bought apples. The fruits were fresh.​
→ “She” refers to Riya​
→ “The fruits” refers to apples

Understanding these links across sentences is called discourse reference

resolution.
⭐ Why Do We Need Discourse Reference Resolution?
It is essential for:

●​ Understanding long texts

●​ Summarization
●​ Dialogue systems
●​ Machine translation
●​ Question answering
●​ Story comprehension in NLP models

Without resolving references across sentences, machines cannot understand who or

what is being discussed.

⭐ Challenges in Discourse Reference Resolution

1.​ Long-distance references (antecedent may be far away)
2.​ Multiple possible antecedents (ambiguity)
3.​ Event references (not nouns, but actions)
4.​ Pronouns with unclear identity (“he”, “she”, “it”, “they”)
5.​ Implicit references (not explicitly mentioned)

⭐ Types of References That Occur in Discourse

1) Anaphora (Backward Reference)

Refers to something mentioned earlier.

Example:​
Ravi lost his phone. It was expensive.​
→ “It” = phone

2) Cataphora (Forward Reference)

Refers to something mentioned later.

Example:​
When he arrived, Rohan was tired.​
→ “he” = Rohan

3) Coreference

Multiple phrases refer to the same entity.

Example:​
John came late. The boy apologized. He looked tired.​
→ John = the boy = he

4) Bridging Reference

Indirect reference; not explicitly stated.

Example:​
I bought a laptop. The keyboard is amazing.​
→ “keyboard” is part of “laptop”

5) Event Reference

Pronoun refers to an entire action or event.

Example:​
Riya failed the exam. This upset her.​
→ “This” = the event of failing

⭐ Methods / Approaches for Discourse Reference Resolution

1) Rule-Based Approaches

Use grammar rules, constraints, and heuristics.

Common Rules:

●​ Number agreement → singular ↔ singular

●​ Gender agreement → he ↔ male
●​ Semantic compatibility → subject performs meaningful actions
●​ Recency preference → recent nouns preferred

2) Syntax-Based Algorithms (e.g., Hobbs Algorithm)

Works using parse-tree search to find antecedents.

●​ Searches left-to-right
●​ Prefers nearby NPs
●​ Good for single-sentence references
●​ Limited for long discourses

3) Discourse-Based Algorithms (Centering Theory)

Centering theory views discourse as a sequence of sentences discussing centers

(main topics).

Concepts:

●​ Cf (Forward centers): all important NPs in a sentence

●​ Cb (Backward center): main entity continued from previous sentence
●​ Cp (Preferred center): most prominent NP in current sentence

Idea:

Pronouns often refer to the center of the previous sentence.

4) Machine Learning Approaches

Use annotated corpora to learn patterns.

Features used:

●​ Grammatical roles
●​ Position in discourse
●​ Semantic compatibility
●​ Distance features
●​ Gender/number constraints

5) Deep Learning Models

Use embeddings, transformers, and contextual models (e.g., BERT).

Advantages:

●​ Understand semantic relations

●​ Handle long-range dependencies
●​ Build clusters of coreferent entities

Example: End-to-End Neural Coreference Model

⭐ Steps in Discourse Reference Resolution

1.​ Identify mentions​
– Nouns, pronouns, noun phrases
2.​ Extract features​
– Number, gender, sentence position, role
3.​ Link candidates​
 – Check each previous noun phrase
4.​ Score matches​
– Use rules or ML to match pronoun ↔ noun
5.​ Form coreference chains​
– Group all expressions referring to the same entity
6.​ Maintain discourse coherence​
– Ensure connections remain logically consistent

⭐ Example Demonstration
Text:​
Riya entered the room. The girl looked tired. She sat down slowly.

Steps:

●​ Identify mentions → Riya, the girl, she

●​ Determine possible links
●​ Riya = the girl (same entity)
●​ “She” refers to this entity

→ Coreference Chain: {Riya, the girl, she}