Global Learning of Noun Phrase Anaphoricity in Coreference Resolution via Label Propagation

Guodong Zhou, Fang KongEMNLP 2009

Introduction

Coreference resolution—identifying which noun phrases (NPs) in a text refer to the same real-world entity—is a fundamental task in Natural Language Processing. A critical challenge in this process is anaphoricity determination: deciding whether an NP refers back to a previous entity (anaphoric) or not (non-anaphoric). Non-anaphoric NPs, such as the pleonastic "it" in "It is raining" or unique definite descriptions like "The sun," often lead to errors when systems incorrectly attempt to find antecedents for them.

Traditional approaches rely on local classifiers or heuristic rules, which often fail to capture the broader context. This paper proposes a global learning framework using a graph-based Label Propagation (LP) algorithm to improve anaphoricity detection and, consequently, coreference resolution performance.

Methodology: Global Learning via Label Propagation

The authors frame anaphoricity determination as a semi-supervised learning problem on a graph.

Graph-Based Label Propagation

In this framework, every NP (both labeled training instances and unlabeled test instances) is represented as a vertex in a connected graph. Labels (anaphoric vs. non-anaphoric) propagate from labeled nodes to their neighbors based on edge weights, which represent the similarity between NPs. This global approach allows the system to leverage the "natural clustering" of data across an entire corpus, leading to more consistent decisions than local SVM classifiers.

Similarity Kernels and Structural Context

To compute edge weights, the authors explore two types of kernels:

  1. Feature-based RBF Kernel: Utilizes 17 lexical, syntactic, and semantic features (e.g., pronoun status, string matching).
  2. Convolution Tree Kernel: Captures the syntactic "neighborhood" of an NP by analyzing its surrounding parse tree structure.
The full parse tree provides the complete syntactic context for an NP.

The paper highlights that while flat features work well for pronouns, structural information is vital for definite NPs. The authors evaluate several tree representations, including the Common Tree (CT) and Minimum Tree (MT), and specifically propose a Dynamically Extended Tree (DET) that enriches the representation with potential antecedent information.

The Minimum Tree (MT) is one variation of a tree kernel structure used to represent the syntactic relationship between mentions.

Experimental Results

The system was evaluated on the ACE 2003 corpus across three domains: NWIRE (Newswire), NPAPER (Newspaper), and BNEWS (Broadcast News).

  • Improved Accuracy: The LP-based method significantly outperformed SVM baselines. In the NWIRE domain, it achieved 74.8% accuracy for anaphoric mentions ($Acc^+$) and 84.4% for non-anaphoric mentions ($Acc^-$).
  • Coreference Resolution Boost: Integrating the global anaphoricity filter improved the overall coreference system's F1-measure. For NWIRE, the F1-measure rose from 59.4% to 61.8% when using the convolution tree kernel.
  • Efficient Filtering: The model effectively filtered out approximately 82% of non-anaphoric NPs, drastically reducing the search space and potential for error in subsequent resolution steps.

Conclusion

By shifting from local classification to a global, graph-based learning approach, the authors demonstrate a significant improvement in anaphoricity determination. The use of convolution tree kernels, particularly the Dynamically Extended Tree, proves essential for handling complex definite NPs. Future work involves exploring the integration of different kernel types and developing joint models that further unify anaphoricity and coreference resolution.

Summary updated

BibTeX
@inproceedings{zhou2009global,
  author = {Guodong Zhou and Fang Kong},
  title = {{Global Learning of Noun Phrase Anaphoricity in Coreference Resolution via Label Propagation}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2009},
}
Figure