TL;DR
This paper presents a novel anaphor resolution algorithm for German that leverages syntactic constraints, particularly Chomsky's Binding Theory, to efficiently identify pronoun antecedents. It addresses the challenge of integrating these constraints into a practical system by proposing an interdependency-sensitive approach that avoids exponential complexity, achieving high accuracy.
What Problem Does This Paper Address?
Resolving anaphors (like pronouns) in text is a fundamental task in Natural Language Processing (NLP). It involves figuring out which noun phrase (antecedent) a pronoun refers to. This is tricky because multiple strategies are needed: morphosyntactic agreement (gender, number), syntactic rules, semantics, and pragmatics. While syntactic constraints, especially those derived from Chomsky's Binding Theory, are powerful, directly implementing them is computationally expensive. Existing approaches often simplify these constraints or process them in isolation, leading to potential errors due to interdependencies between antecedent decisions. This paper tackles the problem of creating an efficient and accurate anaphor resolution system that effectively incorporates complex syntactic constraints without succumbing to computational infeasibility.
Methods
The core of the paper is an anaphor resolution algorithm designed for the KONTEXT text analysis system. It combines several strategies:
- Morphosyntactic Agreement: Basic checks for person, number, and gender.
- Syntactic Constraints (Binding Theory): The algorithm implements Chomsky's Binding Theory principles (A, B, and C) which govern how anaphors, pronouns, and referring expressions relate to their antecedents within specific "binding categories" (local domains).
- Interdependency-Sensitive Antecedent Selection: This is the key innovation. Instead of processing antecedent decisions independently, the algorithm dynamically re-verifies binding constraints. If choosing an antecedent for one anaphor would violate a binding principle for another anaphor whose antecedent has already been determined, that choice is rejected. This is crucial for handling complex sentence structures and discourse phenomena.
- Preference Criteria: To rank potential antecedents, the algorithm uses heuristics like recency, subject preference, and a novel "case role inertia" criterion. This criterion suggests that pronouns tend to take antecedents that occupy similar semantic or syntactic roles.
- Surface Structure Representation: The system uses a simplified dependency grammar-based representation that captures essential structural details for binding principle verification, such as local domains and subject-object asymmetry.
Datasets
Not discussed in the paper. The paper focuses on theoretical analysis and algorithm design, using examples from linguistic literature and architect biographies for evaluation.
Key Insights and Ideas
- Syntactic Constraints are Crucial: The paper strongly argues that syntactic restrictions, when properly implemented, play a central role in accurate anaphor resolution, outperforming purely lexical or shallow semantic approaches.
- Interdependencies are Real and Costly: Ignoring the dependencies between antecedent decisions for different anaphors can lead to incorrect resolutions. The proposed algorithm's strength lies in its ability to manage these interdependencies efficiently.
- Case Role Inertia is a Powerful Heuristic: This preference criterion, which favors antecedents with similar roles, proved very effective in practice, helping to resolve ambiguities and promote complementary distributions of antecedents.
- Pragmatics is Sometimes Necessary: The paper acknowledges that purely syntactic approaches have limits. In certain cases, determining the correct antecedent requires pragmatic inferencing, especially when defining the boundaries and participants of local domains.
Key Results
The algorithm demonstrated high performance:
- Accuracy: Approximately 90% accuracy for type B pronouns (including possessives) and all occurrences of reflexives in a test corpus of architect biographies.
- Efficiency: The algorithm avoids the exponential time complexity of naive generate-and-test approaches, achieving a worst-case complexity of O(n³), where n is the number of NP nodes.
- Practical Performance: The approach proved to be a satisfactory substitute for generate-and-test strategies, balancing accuracy and computational feasibility.
Limitations and Future Work
- Pragmatic Inferencing: The authors acknowledge that for some cases, especially with complex noun phrases and adjectival participles, semantic and pragmatic inferencing are necessary to define local domains and their participants, which the current syntactic approach doesn't fully capture.
- Interdependency between Parsing and Anaphor Resolution: The current system performs parsing before anaphor resolution. The authors suggest that a more integrated approach, handling parsing and anaphor resolution simultaneously, could improve performance, as parsing decisions (like relative clause attachment) can impact anaphor resolution.
- Further Research: Future work could explore the interdependency between parse tree construction and anaphor resolution more deeply, and investigate how to better integrate semantic and pragmatic knowledge.
Summary updated