Unsupervised inference of implicit biomedical events using context triggers

Event extraction from the biomedical literature is one of the most actively researched areas in biomedical text mining and natural language processing. However, most approaches have focused on events within single sentence boundaries, and have thus paid much less attention to events spanning multiple sentences. The Bacteria-Biotope event (BB-event) subtask presented in BioNLP Shared Task 2016 is one such example; a significant amount of relations between bacteria and biotope span more than one sentence, but existing systems have treated them as false negatives because labeled data is not sufficiently large enough to model a complex reasoning process using supervised learning frameworks.

We present an unsupervised method for inferring cross-sentence events by propagating intra-sentence information to adjacent sentences using context trigger expressions that strongly signal the implicit presence of entities of interest. Such expressions can be collected from a large amount of unlabeled plain text based on simple syntactic constraints, helping to overcome the limitation of relying only on a small number of training examples available. The experimental results demonstrate that our unsupervised system extracts cross-sentence events quite well and outperforms all the state-of-the-art supervised systems when combined with existing methods for intra-sentence event extraction. Moreover, our system is also found effective at detecting long-distance intra-sentence events, compared favorably with existing high-dimensional models such as deep neural networks, without any supervised learning techniques.

Our linguistically motivated inference model is shown to be effective at detecting implicit events that have not been covered by previous work, without relying on training data or curated knowledge bases. Moreover, it also helps to boost the performance of existing systems by allowing them to detect additional cross-sentence events. We believe that the proposed model offers an effective way to infer implicit information beyond sentence boundaries, especially when human-annotated data is not sufficient enough to train a robust supervised system.