TL;DR
This paper presents a rule-based system for Chinese coreference resolution, adapting a successful English system by carefully ordering manually designed filters. The approach achieved competitive results on the CoNLL-2012 Shared Task, demonstrating the effectiveness of a structured, deterministic method even without extensive machine learning.
What Problem Does This Paper Address?
Coreference resolution, the task of identifying all expressions in a text that refer to the same real-world entity, is a fundamental challenge in Natural Language Processing. While machine learning approaches have become dominant, they often require large amounts of labeled data, which can be scarce for specific languages or domains. This paper tackles the problem of coreference resolution for Chinese, a language with unique linguistic features that can complicate existing methods. The authors aim to build a robust system by leveraging a deterministic, rule-based approach, inspired by prior work on English, but adapted to the nuances of Chinese.
Methods
The core of the proposed method is a "deterministic coreference model" that employs a series of "filtering sieves" applied in a specific order. This ordered filtering is designed to progressively refine potential coreferent mentions, starting with general rules and moving to more specific ones. The system is an adaptation of the Stanford English coreference system. Key modifications for Chinese include:
- Head Finding Rules: Adapting existing Chinese head-finding rules, which are crucial for several sieves that rely on identifying the main word in a phrase.
- Pronoun Handling: Expanding the set of Chinese pronouns to include common spoken forms and honorifics, which are not always present in standard pronoun lists.
- Plurality Detection: Introducing new rules to identify plural nouns in Chinese, as Chinese POS tags do not inherently distinguish between singular and plural forms. This is achieved by looking for plural markers like "们" (men) and conjunctions like "和" (hé, and).
- Semantic-Based Sieves: The authors removed semantic-based sieves due to resource constraints, focusing on a simplified, rule-based pipeline.
The sieves are ordered from highest to lowest precision, meaning that once a mention is identified as coreferent by a sieve, it is not reconsidered by subsequent sieves.
Datasets
The paper primarily evaluates its system on the CoNLL-2012 Shared Task dataset for both Chinese and English. The task involves a closed track, meaning systems are evaluated on test data using only the provided training data. Specific details about the dataset's composition (e.g., news articles, broadcast conversations) are discussed in the context of experimental results.
Key Insights and Ideas
- Effectiveness of Ordered Filtering: The paper demonstrates that a carefully designed, ordered sequence of deterministic filters can achieve competitive performance, even without relying on complex machine learning models. This suggests that linguistic knowledge, when structured effectively, can be a powerful tool for coreference resolution.
- Language-Specific Adaptations are Crucial: The success of adapting an English system to Chinese hinges on addressing language-specific features, such as pronoun usage and pluralization markers. The authors highlight the importance of these manual adjustments.
- Pronouns as Key Indicators: The paper emphasizes the utility of considering special POS tags and common pronouns as significant indicators for improving coreference resolution performance in Chinese.
- Trade-off in Mention Detection: The authors note that their mention detection component prioritizes high recall over precision to ensure fewer mentions are missed, accepting that some spurious mentions might be filtered out later.
Key Results
The system achieved the following scores on the CoNLL-2012 Shared Task (closed track):
- Chinese: 51.83 F1
- English: 59.24 F1
These results are competitive within the context of the shared task. The paper also presents detailed performance breakdowns on development sets across different data types (broadcast conversations, news, etc.) and mention types (proper nominal, pronominal, other nominal), showing variations in performance across these categories. For instance, the system generally performs better on broadcast news and conversational speech than on web data.
Limitations and Future Work
The authors acknowledge several limitations and suggest future directions:
- Resource Constraints: The removal of semantic-based sieves due to resource limitations is a noted constraint. Future work could explore incorporating these or similar semantic features.
- Manual Rule Design: The reliance on manually designed rules, while effective, is labor-intensive and may not scale as easily as learning-based approaches.
- Future Improvements: The authors suggest extending the system by involving more feasible filtering tiers or developing automatic rule generation methods. They also point to the potential for further improving Chinese mention detection, which showed lower performance compared to English.
Summary updated