Context-Sensitive Convolution Tree Kernel for Pronoun Resolution

Guodong Zhou, Fang Kong, Qiaoming ZhuIJCNLP 2008 · OpenReview

Introduction

Pronoun resolution—the task of linking an anaphor (e.g., "she") to its correct antecedent (e.g., "Mary")—is a fundamental challenge in Natural Language Processing (NLP). Traditional "feature-based" methods rely on manual selection of linguistic traits like gender or grammatical role, often failing to capture complex structural relationships. While tree kernels were introduced to model these structures, they typically suffer from "context blindness" and the difficulty of determining the optimal "tree span"—the amount of structural information to include.

Methodology

The authors propose a two-fold solution to these limitations: a dynamic-expansion tree span scheme and a context-sensitive convolution tree kernel.

Dynamic-Expansion Tree Span Scheme

Rather than using a fixed window of the parse tree, the proposed algorithm dynamically constructs a relevant sub-tree. It begins with the shortest path between the pronoun and a candidate antecedent and then expands to include:

  • Competitors: Other potential candidates that match the pronoun's profile, facilitating learning through comparison.
  • Predicates: Verbs or actions associated with the candidates, providing essential semantic context.
  • Pruning: Trimming redundant nodes to minimize noise.
Figure 1: The step-by-step process of the Dynamic-Expansion Tree Span Scheme, from initial minimal expansion to the final pruned tree.

Context-Sensitive Convolution Tree Kernel

Standard kernels compare sub-trees in isolation. The authors' context-sensitive version incorporates a "root node path" ($m$), allowing the model to consider the ancestors of a sub-tree. This ensures that a grammatical structure is interpreted differently depending on its position (e.g., within a main clause versus a subordinate clause). Empirical results suggest that looking up to three levels ($m=3$) provides the optimal balance between context and noise.

Experimental Results

The model was evaluated on the ACE 2003 corpus. The dynamic expansion scheme achieved a peak F-measure of 83.0, significantly outperforming the previous state-of-the-art "Simple-Expansion" method (81.7).

Robustness Across Sentence Distance

A key strength of this approach is its performance in difficult cases where the pronoun and its referent are separated by multiple sentences. While performance naturally degrades as distance increases, the dynamic scheme remains far more robust than its predecessors, particularly when the antecedent is two sentences away.

Figure 6: Performance comparison across varying sentence distances. The Dynamic scheme shows a significant advantage in long-distance resolution.

Conclusion

By combining a smarter, dynamic method for selecting tree spans with a kernel that respects broader structural context, the authors have advanced the state of pronoun resolution. The inclusion of "competitor" information proved particularly vital, highlighting that resolution is as much about elimination as it is about identification. Future work will focus on further refining structural context and mitigating the impact of initial parsing errors.

Summary updated

BibTeX
@inproceedings{zhou2008context,
  author = {Guodong Zhou and Fang Kong and Qiaoming Zhu},
  title = {{Context-Sensitive Convolution Tree Kernel for Pronoun Resolution}},
  booktitle = {IJCNLP},
  year = {2008},
  url = {https://openreview.net/forum?id=xR0b9rn5hl},
}
Figure