Antecedent Selection Techniques for High-Recall Coreference Resolution

Yannick VersleyEMNLP 2007

TL;DR

This paper tackles the challenging problem of "coreferent bridging" in natural language processing, where mentions refer to entities not explicitly introduced with the same wording. The authors propose and compare various techniques, demonstrating that combining different information sources, particularly those leveraging semantic similarity and web-based patterns, significantly improves recall for this task without substantial precision loss.

What Problem Does This Paper Address?

Coreference resolution, the task of identifying mentions in text that refer to the same real-world entity, is crucial for many NLP applications like information extraction and question answering. While resolving simple anaphora (e.g., "John saw him") is relatively straightforward, this paper focuses on a harder problem: resolving definite descriptions (e.g., "the car") where the antecedent isn't directly stated with the same lexical head. This is known as "coreferent bridging." For instance, in "Peter was found dead in his flat... the deceased," "the deceased" refers to Peter, but they don't share a common head word. This type of coreference is significantly more difficult because it relies on inferring relationships rather than surface-level similarity. The paper aims to improve the recall of coreference resolution systems by effectively handling these bridging cases.

Methods

The paper investigates and compares several approaches for antecedent selection in coreferent bridging. These methods can be broadly categorized:

  • Lexical Semantic Resources: Utilizing resources like WordNet to find semantic relationships such as synonymy, hypernymy, and near-synonymy. Different strategies involve using node distance, paths through WordNet, or specific relation types.
  • Corpus-Based Association Measures: Employing statistical methods on large text corpora to identify words that frequently co-occur or have related semantic contexts. This includes measures like Lin's similarity and Padó and Lapata's dependency-based semantic space models.
  • Web-Based Patterns: Mining the World Wide Web for shallow linguistic patterns (e.g., "China and other countries") that indicate specific relationships like "is-a."
  • Semantic Classes: Using automatically assigned semantic classes (e.g., person, organization) to constrain candidate antecedents.
  • Combination Approaches: The core contribution lies in exploring how to effectively combine these diverse information sources. This includes strategies like first resolving same-head cases, then using semantic similarity and association measures, and also exploring distance-bounded filtering and combining evidence from multiple sources.

A novel aspect is the explicit separation of antecedent selection from discourse-new detection, allowing for a clearer evaluation of bridging resolution techniques and providing an upper bound for the full coreference task.

Datasets

  • TüBa-D/Z corpus: The authors used the first 125 articles of this coreferentially annotated German newspaper text corpus (totaling 2239 sentences with 633 discourse-old definite descriptions) for their experiments.
  • GermaNet: This is a German lexical semantic network, used for hypernymy and node distance calculations.
  • die tageszeitung: Texts from this German newspaper, comprising about 11 million sentences, were used to induce similarity and association measures. A chunked version of this corpus was used for extracting anaphor-antecedent candidates.

Key Insights and Ideas

  • Bridging is Hard, but Solvable with Combinations: While no single information source perfectly resolves coreferent bridging, combining diverse sources (lexical, corpus-based, web-based) significantly boosts recall.
  • Distributional Similarity Outperforms Association Measures (for Ranking): Interestingly, distributional methods (like Lin's and Padó/Lapata's) often outperformed Garera and Yarowsky's association measure when used for ranking potential antecedents, possibly due to noise or sparse data issues with the latter.
  • Semantic Filtering is Crucial: Simply relying on similarity or association measures can be too noisy. Adding semantic class filtering and distance constraints dramatically improves precision for coreferent bridging cases.
  • Web Patterns Offer Complementary Evidence: While GermaNet is strong for hypernymy, web-based patterns are valuable for capturing instance relations and other context-dependent links that lexical resources might miss.
  • Separating Antecedent Selection is Informative: Evaluating antecedent selection independently of discourse-new detection provides a clearer picture of the challenges and potential of bridging resolution techniques, offering an upper bound for full systems.

Key Results

  • Same-head resolution achieved a precision of 86.5% and recalled 49.8% of all cases.
  • Coreferent bridging resolution was significantly harder. A simple baseline of resolving to the nearest compatible antecedent yielded only 12% precision for bridging cases.
  • Combining GermaNet hypernymy lookup with web-based patterns achieved a recall of 60.5% for bridging cases, but with a low precision of 39%.
  • The best combination approach (combining GermaNet, web patterns, and distance-bounded semantic similarity) achieved a recall of 69.7% for coreferent bridging cases, with a precision that did not suffer significantly compared to less comprehensive approaches.
  • The paper highlights that while Garera and Yarowsky's TheY approach had lower initial precision, it benefited significantly from semantic filtering, reaching 47% precision.

Limitations and Future Work

  • Discourse-New Detection: The paper acknowledges that near-perfect discourse-new detection is not yet achievable, which limits the overall performance of a full coreference resolution system.
  • Training Data Availability: Distributional similarity methods could benefit from more training data, but this is constrained by the availability of similar text genres.
  • Processing Costs: Distributional similarity measures, especially those requiring parsed input, can be computationally expensive.
  • Future Work: The authors suggest that integrating more information sources, particularly those covering different types of relations (e.g., hypernymy vs. instance relations), will lead to further improvements. They also emphasize the importance of good precision, especially when integrating antecedent selection with discourse-new identification. The potential for larger improvements by better identifying discourse-old mentions is also noted.

Summary updated

BibTeX
@inproceedings{versley2007antecedent,
  author = {Yannick Versley},
  title = {{Antecedent Selection Techniques for High-Recall Coreference Resolution}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2007},
}
Figure