Coreference based event-argument relation extraction on biomedical text.

Area of Science:

• Biomedical Natural Language Processing
• Computational Linguistics
• Information Extraction

Background:

• Extracting event-argument (E-A) relations from biomedical texts is crucial for understanding biological processes.
• Existing methods often struggle with relations spanning sentence boundaries and leveraging discourse-level information.

Purpose of the Study:

• To develop and evaluate a new approach for biomedical E-A relation extraction using coreference information.
• To demonstrate the capability of identifying cross-sentence E-A relations through coreference transitivity.
• To compare the performance of different coreference-based models.

Main Methods:

• Proposed two coreference-based models: a pipeline Support Vector Machine (SVM) and a joint Markov Logic Network (MLN).
• Utilized the concept of salience in discourse to identify valuable E-A relations.
• Evaluated models on a biomedical event corpus, assessing their ability to extract relations, including cross-links.

Main Results:

• Both proposed coreference-based models significantly outperformed systems that did not utilize coreference information.
• The joint Markov Logic Network (MLN) model demonstrated superior performance compared to the pipeline Support Vector Machine (SVM) model when using gold coreference data.
• The approach successfully extracted a substantial number of E-A relations, including those spanning sentence boundaries.

Conclusions:

• Coreference information is a valuable resource for enhancing biomedical event-argument relation extraction.
• The joint MLN model offers a more effective strategy for E-A relation extraction than the pipeline SVM approach, particularly in complex biomedical texts.
• This work advances the field of biomedical NLP by providing a robust method for capturing nuanced event-argument relationships.