Discourse Deixis

Unlike exophoric deixis which points to the physical world, discourse deixis targets abstract linguistic entities within the text itself. The referent is not a concrete noun phrase but an event, proposition, fact, or speech act described in a preceding clause or sentence.

• Example: 'The board rejected the merger. This surprised the shareholders.' (This = the entire rejection event)
• Example: 'She lied under oath. That is a serious crime.' (That = the act of lying)
• The demonstrative encapsulates a complex semantic unit rather than a simple entity

Discourse deixis is most commonly signaled by demonstrative pronouns (this, that, these, those) used in subject position to refer backward to clausal or sentential antecedents. The choice between 'this' and 'that' often reflects temporal or emotional distance from the referenced discourse segment.

• Proximal 'this': Suggests immediacy, current relevance, or speaker alignment with the proposition
• Distal 'that': Suggests distance, completion, or speaker dissociation from the proposition
• Plural 'these/those': Reference multiple propositions or a complex series of events

A subtype where the deictic expression refers to a multi-sentence discourse segment rather than a single clause. This requires the NLP system to identify the boundaries of the antecedent segment and construct a summary representation of its content.

• Example: [Three paragraphs describing a policy change]. 'This decision will impact thousands.'
• Requires discourse parsing to segment text into coherent units
• The referent is a macro-proposition synthesized from multiple utterances
• Critical for summarization and question-answering systems

A defining test for discourse deixis is that the demonstrative cannot be replaced by a definite noun phrase that is a synonym or hypernym of an entity in the prior text. If substitution is possible, the reference is likely standard anaphora rather than discourse deixis.

• Discourse Deixis: 'He resigned. This shocked everyone.' (Cannot replace 'this' with 'the resignation' without altering the reference to the event)
• Standard Anaphora: 'The CEO resigned. He was tired.' (He = the CEO, a concrete entity)
• This constraint helps disambiguate between entity and event reference in coreference systems

While typically anaphoric (pointing backward), discourse deixis can also function cataphorically, pointing forward to a proposition about to be introduced. This creates a suspension in discourse coherence that is resolved when the subsequent material is processed.

• Example: 'Listen to this: the entire budget has been cut.'
• Example: 'I'll tell you this much—the project is doomed.'
• The deictic expression serves as a placeholder for upcoming propositional content
• Requires incremental parsing and prediction in real-time NLP systems

Discourse deixis must be distinguished from sentential complement anaphora, where 'it' or 'that' refers to a proposition embedded as a syntactic complement. The key difference is whether the referent is a grammatically subordinated clause or an independent discourse unit.

• Sentential Complement: 'She claimed [that the earth is flat]. I don't believe it.' (It = the embedded clause)
• Discourse Deixis: 'The earth is flat. I don't believe that.' (That = the preceding independent assertion)
• The distinction affects syntactic parsing and reference resolution strategies

The fundamental linguistic mechanism where an expression's interpretation depends on a preceding expression. While discourse deixis is a subtype of anaphora, it differs critically: standard anaphora refers to a concrete noun phrase (e.g., 'Sara lost her wallet'), whereas discourse deixis refers to an abstract proposition, event, or fact (e.g., 'Sara lost her wallet. This ruined her day.'). Coreference resolvers must distinguish between these to build accurate discourse models.

A non-identity anaphoric relationship where a definite noun phrase refers to an entity inferentially linked to a previously introduced discourse referent, rather than directly coreferring with it. This is distinct from discourse deixis: bridging links a noun phrase to an implied part-whole or associative relationship (e.g., 'We entered a restaurant. The waiter took our order.'), while discourse deixis points to a propositional abstraction. Both require world knowledge and inferential reasoning beyond simple string matching.

The complete ordered set of all mentions within a discourse that refer to a single entity. When discourse deixis is present, the abstract proposition itself becomes a discourse entity with its own chain. For example:

• 'The company missed earnings. This shocked investors. It led to a sell-off.'
• The chain for the proposition includes: 'The company missed earnings' → 'This' → 'It' Modern neural coreference systems must construct chains that include these abstract referents alongside concrete entity chains.

The prerequisite subtask of identifying all spans of text that refer to an entity. Discourse deixis poses a unique challenge for mention detection because the referring expression (often demonstrative pronouns like 'this', 'that', 'these', 'those') points to a non-contiguous, abstract antecedent. Traditional mention detection models trained on noun phrases struggle to recognize that 'this' in 'This proves the hypothesis' is a valid mention whose antecedent is the entire preceding clause or sentence.

A pronoun disambiguation challenge requiring deep world knowledge and commonsense reasoning, where two sentences differ by a single word that flips the pronoun's antecedent. While Winograd schemas typically test concrete entity resolution (e.g., 'The trophy didn't fit in the suitcase because it was too big/small'), the underlying requirement for propositional understanding and causal reasoning is directly relevant to resolving discourse deixis. Both demand that systems model situational context beyond syntactic cues.

A discourse model that assigns a real-valued prominence score to each entity based on recency, grammatical role, and mention frequency to guide pronoun resolution. For discourse deixis, salience models must be extended to score abstract propositions as candidate antecedents. A proposition's salience is influenced by:

• Recency: How recently was the event stated?
• Syntactic prominence: Was it in a main clause?
• Causal connectivity: Does the discourse marker signal a result? These features help the model determine when 'this' refers to a proposition rather than a noun phrase.