TL;DR
This paper presents a novel approach to anaphor resolution in unrestricted texts by combining partial parsing with Slot Unification Grammar (SUG). By leveraging the output of a part-of-speech tagger and automatically generating syntactic information, the system efficiently resolves pronominal, surface-count, and one-anaphora with improved accuracy.
What Problem Does This Paper Address?
Anaphor resolution, the task of identifying what a pronoun or other referring expression refers to, is a fundamental challenge in Natural Language Processing (NLP). Traditional approaches often rely on deep, full syntactic parsing, which can be computationally expensive and brittle, especially with noisy, real-world text. This paper addresses the need for an efficient and robust anaphor resolution system that can operate on less syntactically complete input, making it more applicable to a wider range of NLP tasks and frameworks that may not have access to sophisticated parsers.
Methods
The core of the proposed method lies in the integration of a Part-of-Speech (POS) tagger with a novel parsing formalism called Slot Unification Grammar (SUG).
- POS Tagging: The system begins by taking the output of a POS tagger, which provides lexical, morphological, and syntactic information for each word.
- Partial Parsing with SUG: Instead of full parsing, the paper employs SUG to perform partial parsing. SUG is a logical formalism based on unification, extending Definite Clause Grammars (DCGs) by allowing optional constituents to be represented more concisely. This partial parsing focuses on identifying key constituents like noun phrases (np), prepositional phrases (pp), pronouns (p), conjunctions (conj), and verbs (verb).
- Slot Structure (SS): The output of the SUG parser is a "slot structure" (SS) that captures syntactic, morphologic, and semantic information. Crucially, the SS marks optional constituents that were not present in the text, providing valuable information for anaphor resolution.
- Anaphor Resolution Algorithm: The paper proposes an algorithm that operates on the SS. It identifies anaphors (pronouns, surface-count anaphora, and one-anaphora) and searches for potential antecedents within the SS and previous sentences. The algorithm uses a set of constraints (e.g., morphosyntactic agreement) and preferences (e.g., recency, syntactic parallelism) to select the most likely antecedent.
What's new here is the automatic generation of syntactic information through partial parsing with SUG, which avoids the need for a full parser and allows for a more modular NLP system.
Datasets
The system was evaluated on a portion of the CRATER corpus, specifically the International Telecommunications Union CCITT handbook (also known as "The Blue Book"). This corpus contains approximately 5 million words in English, French, and Spanish versions. The Spanish version of the Xerox tagger was used to automatically tag this corpus.
Key Insights and Ideas
- Partial Parsing as a Solution: The paper strongly advocates for partial parsing as a practical solution to the challenges of anaphor resolution in unrestricted texts, highlighting its efficiency and robustness compared to full parsing.
- Modularity of SUG: Slot Unification Grammar's design allows for a modular NLP system where the grammar, lexicon, and anaphor resolution modules can be developed and modified independently.
- Leveraging SS for Anaphora: The slot structure generated by SUG is particularly well-suited for anaphor resolution, as it explicitly encodes information about constituent presence and relationships, which can be directly used to infer antecedents.
- Constraints and Preferences: The effectiveness of the anaphor resolution algorithm hinges on a carefully designed set of constraints and preferences tailored to different types of anaphora, demonstrating the nuanced nature of resolving these linguistic phenomena.
Key Results
- The system detected 100% of the anaphors described in the paper.
- The partial parsing achieved an 81% coverage of words with a simple grammar.
- For pronominal references, the system achieved an 83% accuracy in detecting the antecedent's position, a slight improvement over previous work (75%).
- For one-anaphora and surface-count anaphora, the results were less significant due to a small number of instances (5 anaphors with 80% accuracy).
- Failures were primarily attributed to the lack of semantic information and issues with attachments between parsed constituents.
Limitations and Future Work
- Lack of Semantic Information: The authors acknowledge that the current system relies primarily on lexical, morphologic, and syntactic information. The lack of semantic information is identified as a significant limitation, particularly for resolving certain anaphors.
- Attachments: Problems with attachments between different parsed constituents also contributed to some resolution failures.
- Future Work: The authors plan to incorporate semantic information by using a dictionary that can be consulted automatically. They also aim to improve the grammar to include more constituents (e.g., adverbs, punctuation marks) and explore how to handle quantification problems.
Summary updated