Zero Anaphora

The defining characteristic of zero anaphora is the null syntactic realization of an argument that is fully interpreted at the semantic level. Unlike overt pronouns, there is no lexical token to detect. The existence of the argument is inferred from the verb's argument structure and discourse context. For example, in Spanish, the sentence 'Llegó tarde' (Arrived late) omits the subject pronoun, yet the third-person singular subject is unambiguously understood through verbal inflection and prior discourse.

Zero anaphora is a grammatical requirement in consistent pro-drop languages such as Japanese, Chinese, Korean, Italian, and Spanish. These languages license null subjects (and sometimes null objects) as a core syntactic parameter. In contrast, non-pro-drop languages like English require overt expletive subjects ('it is raining') and generally prohibit zero anaphora in finite clauses, making cross-lingual coreference resolution fundamentally asymmetric.

In many pro-drop languages, the omitted argument is recoverable through rich verbal inflection. The verb carries morphological features—person, number, and gender—that uniquely identify the null subject. This is known as the Avoid Pronoun Principle: speakers omit the pronoun precisely because the inflection makes it redundant. NLP systems must leverage morphological analyzers to reconstruct these dropped arguments for downstream tasks like machine translation.

Resolution of zero anaphora requires cross-sentential discourse modeling. The antecedent of a zero pronoun is often found in a preceding sentence, introduced as a discourse topic. This demands that coreference systems maintain a discourse registry of active entities and their salience scores. A zero subject in Japanese typically refers to the discourse topic rather than the most recent syntactic subject, requiring topic-tracking mechanisms beyond simple recency heuristics.

In topic-prominent languages like Chinese and Korean, zero anaphora frequently interacts with topic chains. A single overt topic can control multiple subsequent clauses with zero subjects, forming a zero anaphora chain. The topic is established once and then omitted across several predicates until a new topic is introduced. Parsing these chains requires segmenting discourse into topic units and identifying topic shift boundaries where the null subject's antecedent changes.

Detecting zero anaphora is fundamentally harder than resolving overt mentions because there is no span to detect. Systems must predict the location of a null argument before resolving it. This is typically done by examining the argument structure of each predicate and identifying missing required arguments. For Japanese, the NAIST Text Corpus provides gold-standard annotations for zero pronouns, enabling supervised models to learn position prediction and antecedent selection jointly.

A linguistic typology where pronouns can be omitted when they are pragmatically or grammatically inferable. Zero anaphora is a direct consequence of this parameter. Japanese, Spanish, and Italian are classic examples.

• Topic-drop: Languages like Chinese allow the subject to be dropped if it is the discourse topic.
• Agreement morphology: In Spanish, rich verb inflection (hablo - 'I speak') makes the subject pronoun redundant.
• Computational challenge: Parsing these languages requires models to reconstruct the dropped argument to build a complete predicate-argument structure.

The broader linguistic phenomenon where an expression's interpretation depends on a preceding expression (the antecedent). Zero anaphora is a subtype where the anaphoric expression is phonologically null.

• Pronominal anaphora: 'Sara lost her wallet.' (Explicit pronoun)
• Zero anaphora: 'Sara arrived and [Ø] sat down.' (Implicit subject)
• Resolution dependency: A zero anaphor cannot be resolved without identifying the antecedent in the discourse model, making it a critical step in coreference resolution for pro-drop languages.

The prerequisite subtask of identifying all spans of text that refer to an entity. For zero anaphora, this step is uniquely difficult because the mention has no surface form.

• Standard detection: Scans for noun phrases and named entities.
• Zero detection: Requires a syntactic or semantic trigger, often a missing argument in a verb's subcategorization frame.
• Modeling approach: Sequence-to-sequence models are trained to insert a special [NULL] token at the position of the dropped argument before standard coreference resolution can proceed.

The complete ordered set of all mentions within a discourse that refer to a single entity. A zero anaphor is a link in this chain, not an isolated phenomenon.

• Chain example: ['John', 'he', Ø, 'the CEO'] — all point to the same entity.
• Transitive closure: If Ø refers to 'he', and 'he' refers to 'John', then Ø refers to 'John'. This transitive reasoning is essential for building coherent chains.
• Evaluation impact: A system that fails to resolve zero anaphors will produce fragmented, incomplete chains, severely penalizing coreference metrics like MUC, B³, and CEAF.

A syntactic theory governing the distribution of anaphors, pronominals, and referring expressions. Its principles provide structural constraints for zero anaphora resolution.

• Principle A: An anaphor must be bound in its local domain. (Not applicable to zero pronouns).
• Principle B: A pronoun must be free in its local domain. This constrains where a zero pronoun can find its antecedent.
• Empty Category Principle: In Government and Binding theory, zero anaphors are classified as pro (a null pronoun) or PRO (the subject of an infinitive), each with distinct syntactic licensing conditions.

A discourse coherence theory that models local focus to constrain pronoun interpretation, directly applicable to zero anaphora resolution.

• Forward-looking centers (Cf): A ranked list of entities in an utterance.
• Backward-looking center (Cb): The single entity that the current utterance is 'about', typically the preferred antecedent for a zero pronoun.
• Rule of continuation: If the Cb of the current utterance is the same as the highest-ranked Cf of the previous utterance, a zero pronoun is the preferred form of reference. This explains why zero anaphora signals discourse continuity.