When humans read a sentence like "Of all the cars Jack saw, he liked the black one the most," or participate in a conversation where someone says, "You need to look at this," we instantly resolve what "one", "you", and "this" refer to. For a machine, however, this task—known as anaphora resolution—presents a massive search space. If a system tries to find an antecedent for every single word or phrase across an entire document, it quickly becomes overwhelmed by computational complexity and linguistic noise.
The core insight behind a long line of research is that anaphora resolution is best handled not as a single monolithic task, but as a two-stage pipeline. By decoupling the detection of an anaphoric expression from its subsequent resolution, systems can filter out non-referential noise, narrow down the search space, and apply specialized models to distinct sub-problems.
The Intuition: Why Monolithic Resolution Fails
Why not just train an end-to-end model to link every pronoun or noun phrase to an antecedent? The problem lies in the linguistic diversity of natural language: many words that can be anaphoric often are not. For instance:
- The word "you" can refer to a specific person in a room, or it can be used generically ("You win some, you lose some").
- Definite noun phrases often introduce entirely new entities rather than referring back to old ones.
- Languages like Chinese frequently omit pronouns entirely ("zero pronouns" or "dropped pronouns"), meaning a system must first find where a pronoun is missing before it can figure out what it refers to.
If a coreference resolver attempts to find an antecedent for a non-anaphoric phrase, it introduces false positives and wastes computational resources. As Bean et al. demonstrated in Corpus-Based Identification of Non-Anaphoric Noun Phrases, explicitly filtering out non-anaphoric noun phrases before resolution minimizes false positive links and drastically reduces overall computational complexity. Similarly, Kim et al. leveraged this separation in Pipeline Coreference Resolution Model for Anaphoric Identity in Dialogues to reduce model complexity and training times compared to end-to-end architectures.
[ Raw Text Input ]
│
▼
┌────────────────────────────────────────────────────────┐
│ STAGE 1: Trigger Detection & Functional Classification │
│ (Filters out non-anaphoric noise, identifies gaps) │
└────────────────────────────────────────────────────────┘
│
▼ [Verified Anaphoric Triggers Only]
┌────────────────────────────────────────────────────────┐
│ STAGE 2: Targeted Antecedent Selection │
│ (Searches reduced space, ranks candidate referents) │
└────────────────────────────────────────────────────────┘
│
▼
[ Grounded Coreference Links ]
Stage 1: Trigger Detection and Functional Classification
The first stage of a decoupled pipeline acts as a gatekeeper. Its job is to answer a binary or categorical question: Is this instance actually anaphoric, and if so, what is its functional type?
This is particularly crucial for highly ambiguous words or implicit structures:
- Ambiguous Pronouns: In multi-party dialogues, Gupta et al. in Resolving "You" in Multi-Party Dialog and Frampton et al. in Who is “You”? Combining Linguistic and Gaze Features to Resolve Second-Person References in Dialogue used hierarchical classification to first determine whether "you" is generic or referential before attempting entity grounding.
- One-Anaphora: The word "one" can be a number, a generic pronoun, or an anaphoric substitute for a noun phrase sense. Ng et al. in A Machine Learning Approach to Identification and Resolution of One-Anaphora and Khullar et al. in Finding The Right One and Resolving it showed that separating functional classification from antecedent ranking prevents non-anaphoric instances from polluting the resolution phase.
- Discourse-Deictic Pronouns: Pronouns like "this" or "that" can refer to entire clauses rather than simple nouns. Recasens in Resolving Discourse-Deictic Pronouns: A Two-Stage Approach to Do It built a two-stage system where a maximum entropy classifier first determines if a pronoun is discourse-deictic before a second classifier attempts resolution.
- Omissions and Ellipses: When elements are omitted entirely, the first stage must detect the "trigger" or location of the gap. For verb phrase ellipsis, Kenyon-Dean et al. in Verb Phrase Ellipsis Resolution Using Discriminative and Margin-Infused Algorithms separated trigger detection from antecedent identification. For Chinese dropped pronouns, Ng et al. in Identification and Resolution of Chinese Zero Pronouns: A Machine Learning Approach and Zhang et al. in Neural recovery machine for Chinese dropped pronoun first identified the positions of omitted pronouns before generating or resolving them.
Stage 2: Targeted Antecedent Selection
Once an expression is confirmed to be anaphoric, the second stage focuses entirely on finding the correct referent. Because the first stage has already filtered out the noise, the second stage operates over a much more tractable search space and can employ specialized ranking or reasoning mechanisms.
This division of labor allows for different algorithmic strategies across the two stages:
- Syntactic Filtering vs. Semantic Reasoning: Early foundational work by Webber et al. in Anaphora and Logical Form: On Formal Meaning Representations for Natural Language argued that a formal syntactic stage should generate candidate referents independently of complex common-sense knowledge, leaving the final selection to a subsequent reasoning stage.
- Possible vs. Preferred Candidates: Dorrepaal in Discourse Anaphora structured the resolution stage itself into a two-step process: first identifying linguistically compatible "possible" antecedents, and then applying pragmatic, extra-linguistic knowledge to choose the "preferred" one.
- Coarse-to-Fine Search Spaces: For abstract references, Kolhatkar et al. in Resolving “This-issue” Anaphora demonstrated that identifying a broad region of potential antecedents before refining the selection makes the resolution task significantly more manageable. In technical domains, Grishman et al. in Causal and Temporal Text Analysis: The Role of the Domain Model parsed compound noun phrases using broad semantic categories first, followed by a compositional procedure to map them to specific domain entities.
- Modular Frameworks: This structured progression dates back to early systems like the one described by Studnicki et al. in The Research Project “ANAPHORA” (In Its Present State of Advancement), which mapped clauses into formal formulas before selecting and executing targeted search algorithms.
The Trade-off: Specialized Power vs. Error Propagation
The primary reason the two-stage approach works so well is feature and objective specialization. As noted by Kenyon-Dean et al., a system can use standard classification (like logistic regression) to optimize trigger detection, and margin-based ranking algorithms to optimize antecedent selection. Each model uses features tailored to its specific task—such as gaze and turn-taking features for identifying conversational targets, or structural tree features for finding syntactic antecedents.
However, this modularity comes with a well-known vulnerability: error propagation. If the first-stage classifier mistakenly labels an anaphoric pronoun as generic, the second stage never gets the chance to resolve it. Conversely, if a non-anaphoric phrase slips through the filter, the resolver is forced to make a spurious link. Despite this risk, the dramatic reduction in search space complexity and the ability to train highly specialized models have made the two-stage pipeline an enduringly powerful paradigm across decades of anaphora research.
Key Foundational Work
The Research Project “ANAPHORA” (In Its Present State of Advancement) (1982) — contributed 2 concept elements to this direction.
Anaphora and Logical Form: On Formal Meaning Representations for Natural Language (1977) — contributed 1 concept elements to this direction.