An Incremental Model for the Coreference Resolution Task of BioNLP 2011

Don Tuggener, Manfred Klenner, Gerold Schneider, Simon Clematide, Fabio RinaldiACL 2011 · OpenReview

Abstract

Coreference resolution in biomedical texts is a critical task for identifying mentions of biological entities like proteins and genes. Traditional mention-pair models often suffer from computational inefficiency and logical inconsistency. This paper presents an incremental coreference resolution system developed for the BioNLP 2011 Shared Task. By processing mentions sequentially and utilizing "virtual prototypes" of coreference sets, the system enforces global constraints like transitivity without complex machine learning. Relying on deep syntactic parsing and an empirical salience measure, the model achieves high precision and competitive performance relative to other participants.

Introduction

Identifying when different terms in a text refer to the same biological entity is a fundamental challenge in biomedical text mining. Traditional coreference resolution systems typically employ "mention-pair" models, which evaluate every possible pair of mentions. This approach faces three primary limitations: quadratic computational growth, underspecification of individual mentions (e.g., pronouns), and logical inconsistencies where transitivity (if A=B and B=C, then A=C) is not naturally enforced.

To address these issues, the researchers developed an incremental model that constructs coreference chains dynamically as it processes the text. This approach reduces the search space and ensures that decisions made for one mention are consistent with the entire coreference set.

Methods

The system utilizes a linguistically-driven, incremental pipeline rather than traditional machine learning. The process begins with the OntoGene preprocessing pipeline, which performs deep syntactic parsing to identify sentence boundaries, parts of speech, and dependency structures.

Incremental Processing and Virtual Prototypes

As the system encounters a new noun phrase (NP), it compares it against existing "coreference sets." To solve the underspecification problem, the system aggregates morphological and semantic information from all mentions in a set into a "virtual prototype." The new mention is then compared to this prototype rather than individual members.

Algorithm for incremental coreference resolution.

The algorithm follows a specific logic: for each mention, it checks compatibility with existing coreference sets and licensed buffer elements. If no candidate is found, a new set is created. If multiple candidates exist, an empirical salience measure—based on dependency labels (Subject > Object > Pobject)—is used to select the most likely antecedent.

Results

The model was evaluated on the BioNLP 2011 Shared Task on Protein/Gene interaction, focusing on anaphoric references to pre-identified terms.

Performance comparison against BioNLP 2011 participants.

As shown in the results table, the system achieved a Precision (P) of 55.45, a Recall (R) of 21.48, and an F1-score of 30.96. In the context of the competition, this model ranked second among the listed participants, trailing only "Team A" and outperforming several other systems. The high precision indicates that the linguistic filters and salience measures effectively minimized false positive links.

Discussion and Conclusion

The incremental architecture naturally enforces transitivity, improving consistency and efficiency. By using syntax as a proxy for importance, the system demonstrates that grammatical roles are reliable indicators of coreference in scientific writing.

However, the system's recall is notably lower than its precision, suggesting that the strict linguistic rules may miss valid coreferences that do not fit the predefined criteria. Future work will focus on handling "bridging anaphora" and refining the implementation of Binding Theory to improve coverage while maintaining high precision.

Summary updated

BibTeX
@inproceedings{tuggener2011incremental,
  author = {Don Tuggener and Manfred Klenner and Gerold Schneider and Simon Clematide and Fabio Rinaldi},
  title = {{An Incremental Model for the Coreference Resolution Task of BioNLP 2011}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2011},
  url = {https://openreview.net/forum?id=a12bx9jD3S},
}
Figure