Abstract
Anaphora resolution—the process of identifying the antecedents of pronouns and noun phrases—is a critical component of natural language understanding. Traditional systems are often language-specific and brittle. This paper presents a modular, data-driven discourse module developed at SRA that separates the resolution engine from language-specific knowledge bases. This architecture enables portability across languages like English, Spanish, and Japanese and allows for statistical optimization using tagged corpora.
Introduction
In natural language processing (NLP), anaphora resolution is the task of determining the antecedent of a reference. For example, in the sentence "The company released its report; it was profitable," a system must determine that "it" refers to "the report." Historically, these systems were monolithic and hard-coded for a specific language, making them difficult to adapt. Furthermore, real-world text is often "noisy"; upstream parsing errors frequently cause discourse modules to fail. The proposed system addresses these challenges by being language-independent, extensible to new linguistic phenomena, and robust against incomplete input.
System Architecture and Methodology
The core innovation of this research is a "plug-and-play" architecture that decouples the processing logic from linguistic data.
1. The Resolution Engine
The system is divided into a language-neutral Resolution Engine and various Discourse Knowledge Bases (KBs). The engine executes the logic of anaphora resolution without being tied to the specific rules of any one language.
2. Knowledge Sources (KS)
Resolution is decomposed into small, specialized "Knowledge Sources" (KS) rather than a single monolithic algorithm. These are categorized into three types:
- Generators: Propose potential antecedents based on structural or proximity rules.
- Filters: Eliminate candidates that violate linguistic constraints (e.g., gender, number, or semantic class).
- Orderers: Rank the remaining candidates to identify the most probable match.
3. Data Structures and File Cards
The system utilizes File Cards, a concept inspired by Irene Heim’s File Change Semantics. When a new entity is mentioned, a "discourse marker" is created. If the system resolves this marker to an existing entity, it updates a corresponding "File Card." This monotonic process ensures that information about an entity is accumulated over time rather than replaced, building a comprehensive semantic profile.
Evaluation and Key Insights
The system was evaluated using manually discourse-tagged corpora across English, Spanish, and Japanese, primarily focusing on technical and medical domains such as AIDS surveillance reports.
Key Insights
- Separation of Function and Data: By isolating the logic of a filter (e.g., "Syntactic-Gender Filter") from its language-specific data (e.g., "he" is masculine), the core engine remains portable.
- Robustness through Backup Strategies: To handle noisy input, the system employs a "Main Strategy" with precise rules and a "Backup Strategy" with looser constraints that triggers if the primary search fails.
- Statistical Tuning: The authors applied quantitative metrics—such as failure rates for generators and false positive rates for filters—to objectively select the most effective rules for specific domains.
Results and Future Work
The modular approach successfully resolved anaphora across three distinct languages and handled complex phenomena like Japanese "dou" ellipsis (omitted objects) through simple KB additions. While the system establishes a rigorous framework for precision and recall, future work aims to improve the combination of multiple orderers via weighted scoring and to incorporate large-scale lexical statistics to resolve highly ambiguous cases.
Summary updated