TL;DR
This paper tackles the complex problem of anaphora resolution (figuring out what pronouns like "he," "she," or "it" refer to) by proposing a novel multi-strategy approach. Instead of a single, monolithic solution, they combine multiple linguistic knowledge sources (syntax, semantics, pragmatics) as distinct strategies, arguing this is more effective and cognitively plausible.
What Problem Does This Paper Address?
In natural language understanding, anaphora resolution is crucial for systems to grasp the meaning of text or dialogue. It's the process of identifying the antecedent (the noun or noun phrase) that a pronoun or other referring expression refers to. While seemingly simple for humans, computationally resolving anaphora is notoriously difficult because it requires integrating various types of knowledge:
- Syntactic knowledge: Grammatical rules and sentence structure.
- Semantic knowledge: Word meanings and how they fit together in phrases and sentences (e.g., what can be "eaten" vs. "washed").
- Pragmatic knowledge: Real-world knowledge, context, and conversational intent.
Previous computational approaches often used rudimentary methods or focused on limited types of anaphora (e.g., within a single sentence). This paper aims to address intersentential anaphora (references across sentence boundaries) by proposing a more comprehensive, yet still manageable, computational framework.
Methods
The core of the paper is the proposal of a multi-strategy approach to anaphora resolution. Instead of a single, all-encompassing algorithm, they break down the problem into several distinct strategies, each leveraging a different source of linguistic knowledge. These strategies act as constraints or preferences:
- Local Anaphor Constraints: Basic checks like number and gender agreement (e.g., "he" can't refer to "Mary").
- Case-role Semantic Constraints: Using semantic roles (like agent, patient, recipient) from case-frame semantics to filter candidates. For example, "eating" requires an edible object, while "washing" might not.
- Precondition/Postcondition Constraints: Applying real-world knowledge about actions. If an action implies an object is no longer available, it can't be the antecedent for a subsequent action requiring that object.
- Case-role Persistence Preference: A preference for an anaphor to refer to an entity that occupied the same semantic role in a previous utterance.
- Syntactic Parallelism Preference: A preference for anaphora resolution that maintains parallel syntactic structures between clauses.
- Syntactic Topicalization Preference: A preference to search topicalized parts of a sentence (like those in cleft constructions) first for potential antecedents.
- Intersentential Recency Preference: A general preference to search for antecedents in more recent sentences, working backward through the discourse.
The system first applies hard constraints to eliminate impossible antecedents. Then, preferences are used to rank the remaining candidates. If multiple strategies suggest different candidates, or if no candidate satisfies all constraints, the anaphor is considered ambiguous.
Datasets
Not discussed in the paper. The authors mention using a "test suite" of ten examples, totaling 31 sentences with 27 anaphors, for evaluation. They also refer to earlier hand simulations on 70 examples.
Key Insights and Ideas
- Decomposition is Key: The central hypothesis is that anaphora resolution is best handled by combining multiple, specialized strategies rather than a single, monolithic approach. This makes the problem more tractable and aligns with human cognitive processes.
- Cognitive Plausibility: The authors argue their multi-strategy approach mirrors human anaphora resolution. When their strategies agree, humans typically find a unique referent. When strategies conflict, humans often perceive ambiguity.
- Hierarchy of Knowledge: While syntax plays a role, semantic and pragmatic knowledge are highlighted as particularly crucial for intersentential anaphora.
- Constraints vs. Preferences: Distinguishing between hard constraints (must be met) and preferences (influence choice) is vital for a robust system.
- "Linguistic Inertia": Concepts like case-role persistence and semantic alignment capture a natural tendency in language to maintain consistency in how entities are referred to across utterances.
Key Results
The paper reports that their implemented anaphor resolver, using a subset of the proposed strategies, correctly resolves all but four anaphors in their test suite of 31 sentences. For definite noun phrases, it correctly identifies referents. The authors state that their method achieves "human-like performance" in resolving anaphora, with human judgments correlating well with the system's unique resolutions and perceived ambiguities.
Limitations and Future Work
- Limited Scope: The current implementation doesn't handle anaphora referring to actions indirectly mentioned or actions that are not explicitly part of a case frame.
- Ambiguity: In some cases, the system still produces ambiguous results, particularly when multiple candidates receive similar "votes" from the preference strategies.
- Pragmatic Topicalization: The paper acknowledges the importance of pragmatic topicalization (dialog focus) but notes the lack of a computationally adequate theory to implement it.
- Future Work: The authors suggest further research into more sophisticated discourse structure theories to improve recency preference and the integration of pragmatic topicalization. They also plan to make the resolver an integral part of their multi-lingual machine translation effort.
Summary updated