Incorporating Structural Information for Better Coreference Resolution

Fang Kong, Jian FuIJCAI 2019 · DOI

TL;DR

This paper introduces a novel neural coreference resolution system that effectively leverages syntactic parse tree information. By using parse tree nodes to filter mention candidates, encoding context through tree traversal sequences, and incorporating additional structural features, the model significantly improves performance over existing neural approaches, particularly for Chinese language.

What Problem Does This Paper Address?

Coreference resolution, the task of identifying mentions in text that refer to the same real-world entity, is a fundamental challenge in Natural Language Processing (NLP). While recent end-to-end neural models have achieved state-of-the-art results by treating all text spans as potential mentions, they often overlook valuable structural information present in syntactic parse trees. This omission can lead to increased computational complexity due to noisy mention candidates and potentially suboptimal representations of mentions. The paper aims to bridge this gap by demonstrating how incorporating syntactic structure can enhance both the efficiency and accuracy of neural coreference resolution systems.

Methods

The authors propose a three-pronged approach to integrate structural information into a neural coreference resolution framework:

  1. Node-based Mention Candidate Filtering: Instead of considering all text spans within a length limit, the system now uses nodes from syntactic parse trees as a constraint. Only text spans that directly correspond to a node in the parse tree are considered as potential mention candidates. This effectively prunes the search space, reducing computational complexity and noise.
  2. Traversal Node Sequence for Contextual Encoding: The traditional approach uses word sequences to encode context. This paper introduces a novel method that traverses the parse tree (using a post-order traversal) to generate a node sequence. This sequence implicitly captures hierarchical information. The contextual embedding layer then processes this sequence, allowing it to learn from the structural relationships between nodes, not just linear word dependencies. Furthermore, child node representations are used to refine parent node representations.
  3. Additional Structural Features: Beyond the parse tree structure itself, the model incorporates several explicit structural features to enrich mention representations. These include:
  4. Degrees: The number of child nodes.
  5. Siblings: The number of left and right siblings.
  6. Category: The syntactic category of the node (e.g., NP, VP).
  7. Path: The path from the current node to the root of the parse tree.

These features are encoded and combined with the span representation to create a more robust mention representation.

Datasets

The experiments were conducted on the CoNLL-2012 Shared Task dataset, which is a standard benchmark for coreference resolution. The paper reports results for both the English and Chinese portions of this dataset.

Key Insights and Ideas

  • Syntactic Structure is Crucial: The paper strongly suggests that ignoring syntactic structure in neural coreference resolution is a missed opportunity. Even simple structural constraints can significantly improve performance.
  • Chinese Language Benefits More: While the proposed model improves performance for both English and Chinese, the gains are particularly notable for Chinese. This might be due to the inherent structural richness of Chinese parse trees or differences in how span pruning performs across languages.
  • Hierarchical Context is Key: Encoding context not just linearly but hierarchically through parse tree traversal leads to better representations, highlighting the importance of capturing relationships beyond immediate neighbors.
  • Noise Reduction: Using parse tree nodes as filters effectively reduces the number of noisy, irrelevant mention candidates, leading to a more efficient system.

Key Results

The proposed model, incorporating structural information, significantly outperforms the baseline (a direct replication of Lee et al., 2017) on the CoNLL-2012 Shared Task.

  • For English, the model achieved an average F1 score of 68.57, an improvement of 2.68% over the baseline (66.01). This improvement was particularly noticeable in precision.
  • For Chinese, the model achieved an average F1 score of 63.85, an improvement of 3.03% over the baseline (60.82).
  • The paper also shows that integrating structural information incrementally (S1: node constraint, S2: node-based context, S3: additional features) leads to consistent improvements, with the combination of all three schemes yielding the best results. The node-based contextual encoder (S2) was found to be the most impactful single component.

Limitations and Future Work

  • ELMo Integration: The authors acknowledge that there is still a significant performance gap with models that incorporate ELMo embeddings, suggesting that future work should explore integrating such advanced word representations.
  • Antecedent Identification: While the paper focuses on improving mention representation and filtering, future work could also explore more sophisticated antecedent identification strategies.
  • Computational Resources: The authors mention that their baseline implementation was faster than the original due to batch processing of span pruning, but they also note that GPU resource limitations affected their fine-tuning process for the baseline.
  • Language Specificity: While the model shows promise for Chinese, further investigation into language-specific structural nuances could yield even greater improvements.

Summary updated

BibTeX
@inproceedings{kong2019incorporating,
  author = {Fang Kong and Jian Fu},
  title = {{Incorporating Structural Information for Better Coreference Resolution}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2019},
  doi = {10.24963/ijcai.2019/700},
}
Figure