Unsupervised inference of implicit biomedical events using context triggers.

Area of Science:

• Biomedical text mining
• Natural Language Processing

Background:

• Biomedical event extraction often focuses on single sentences.
• Cross-sentence events, like Bacteria-Biotope relations, are challenging due to limited labeled data.
• Existing systems struggle with implicit events spanning multiple sentences.

Purpose of the Study:

• To develop an unsupervised method for inferring cross-sentence biomedical events.
• To overcome the limitations of supervised learning in event extraction due to data scarcity.
• To improve the detection of implicit and long-distance events.

Main Methods:

• Propagating intra-sentence information to adjacent sentences using context trigger expressions.
• Collecting trigger expressions from unlabeled text using syntactic constraints.
• An unsupervised, linguistically motivated inference model.

Main Results:

• The unsupervised system effectively extracts cross-sentence events.
• Performance surpasses state-of-the-art supervised systems when combined with intra-sentence methods.
• The system also detects long-distance intra-sentence events, rivaling deep neural networks without supervision.

Conclusions:

• The model successfully detects implicit, cross-sentence events without relying on training data or knowledge bases.
• It enhances existing systems by enabling detection of additional cross-sentence events.
• Offers an effective solution for inferring implicit information beyond sentence boundaries, especially with insufficient annotated data.