🔠Intro to Semantics and Pragmatics Unit 11 – Anaphora and Coreference

Key Concepts and Definitions

• Anaphora refers to the use of an expression that depends on another referential element for its interpretation
• Antecedent is the linguistic expression that an anaphor refers back to and derives its meaning from
• Coreference occurs when two or more expressions in a text refer to the same entity or concept
• Referential ambiguity arises when there are multiple potential antecedents for an anaphor
• Binding theory governs the syntactic constraints on the interpretation of anaphors and pronouns
  • Principle A states that an anaphor must be bound in its local domain
  • Principle B requires a pronoun to be free in its local domain
• Discourse model represents the mental representation of entities and their relationships in a text
• Salience refers to the degree of prominence or accessibility of an entity in the discourse model

Types of Anaphora

• Pronominal anaphora involves the use of pronouns (he, she, it) to refer back to a previously mentioned entity
• Nominal anaphora uses noun phrases (the book, that idea) to refer to an antecedent
• Verb phrase anaphora occurs when a verb phrase (do so, do it) refers to a previously mentioned action or event
• Zero anaphora, also known as ellipsis, involves the omission of an anaphor when it can be inferred from the context
• Bridging anaphora requires inferring the relationship between the anaphor and its antecedent based on world knowledge (the door, implying the door of a previously mentioned house)
• Cataphora is a less common type of anaphora where the anaphor precedes its antecedent in the text
• Split antecedents occur when an anaphor refers to multiple antecedents in a coordinated structure (John and Mary, they)

Coreference Explained

• Coreference chain is a sequence of expressions in a text that all refer to the same entity
• Identity coreference occurs when two expressions refer to the exact same entity (John, he)
• Near-identity coreference involves expressions that refer to almost the same entity with slight differences (the book, the novel)
• Bound variable coreference arises when a pronoun is bound by a quantified antecedent (every student, their)
• Coreference resolution is the task of determining which expressions in a text corefer
  • It involves identifying potential antecedents and selecting the most likely one based on syntactic, semantic, and pragmatic factors
• Coreference clusters are formed by grouping expressions that corefer into sets
• Singleton coreference occurs when an entity is mentioned only once in the text and does not participate in a coreference chain

Linguistic Theories and Models

• Government and Binding Theory (GB) provides a modular approach to anaphora and coreference within the framework of generative grammar
• Centering Theory focuses on the local coherence of discourse and the role of salience in anaphora resolution
  • It introduces the concepts of backward-looking center (CB) and forward-looking centers (CF) to model the attentional state of discourse
• Discourse Representation Theory (DRT) represents the semantic content of a discourse using discourse representation structures (DRS)
  • Anaphoric expressions are resolved by linking them to accessible discourse referents in the DRS
• Optimality Theory (OT) approaches anaphora resolution as a constraint satisfaction problem
  • It proposes a set of violable constraints (e.g., AGREE, DISJOINT) and selects the optimal candidate based on their ranking
• Mention-pair models treat coreference resolution as a binary classification task between pairs of mentions
• Entity-mention models consider the properties of entities and their mentions across the entire text

Anaphora Resolution Techniques

• Rule-based approaches use a set of hand-crafted rules based on linguistic knowledge to resolve anaphors
  • They often rely on syntactic patterns, gender and number agreement, and semantic compatibility
• Machine learning methods learn patterns and features from annotated data to make coreference decisions
  • Supervised learning techniques (decision trees, logistic regression) require labeled training data
  • Unsupervised learning (clustering algorithms) can discover coreference chains without explicit annotations
• Neural network models, such as LSTMs and transformers, can capture complex semantic and contextual information for anaphora resolution
• Hybrid approaches combine rule-based and machine learning techniques to leverage their strengths
• Knowledge-based methods incorporate external knowledge sources (WordNet, Wikipedia) to aid in resolving anaphors
• Discourse structure and rhetorical relations (elaboration, contrast) can provide additional cues for anaphora resolution
• Psycholinguistic factors, such as recency and frequency of mention, influence the accessibility of antecedents

Pragmatic Considerations

• Gricean maxims, particularly the maxim of quantity and manner, guide the use and interpretation of anaphors in communication
• Conversational implicature can affect the choice of anaphoric expressions and their perceived meaning
• Common ground and shared knowledge between the speaker and hearer play a role in resolving anaphors
• Perspective-taking and theory of mind enable speakers to use anaphors in a way that is easily interpretable by the hearer
• Referential intentions of the speaker, as inferred by the hearer, contribute to the resolution of anaphors
• Pragmatic presupposition, such as the existence and uniqueness of referents, is often assumed when using anaphoric expressions
• Deixis, or the use of expressions that depend on the extralinguistic context (I, you, here), interacts with anaphora resolution

Real-world Applications

• Machine translation systems need to handle anaphora and coreference to produce coherent and fluent translations across languages
• Information extraction and question answering tasks require accurate identification and resolution of anaphoric expressions
• Summarization systems aim to maintain coreference chains and avoid ambiguity when generating concise versions of texts
• Dialogue systems and conversational agents must keep track of referents and resolve anaphors to engage in natural interactions
• Sentiment analysis and opinion mining can benefit from coreference resolution to attribute sentiments to the correct entities
• Legal and medical text processing often involves complex anaphoric relations that need to be properly resolved
• Coreference annotation tools, such as brat and WebAnno, assist in creating datasets for training and evaluating anaphora resolution models

Common Challenges and Pitfalls

• Pronoun resolution can be difficult when there are multiple compatible antecedents in terms of gender and number
• Cataphoric references, where the anaphor precedes its antecedent, can be harder to resolve than anaphoric references
• Long-distance anaphora, where the antecedent is far from the anaphor in the text, poses challenges for resolution algorithms
• Ambiguous or underspecified antecedents, such as "it" referring to a complex idea or event, can be difficult to resolve
• Metaphorical or non-literal language use can complicate the interpretation of anaphoric expressions
• Cross-document coreference, resolving anaphors across multiple texts, introduces additional complexity compared to within-document resolution
• Domain-specific language and terminology can require adapted anaphora resolution approaches
• Evaluation metrics for anaphora resolution, such as MUC, B-CUBED, and CEAF, have their own strengths and limitations