Extracting Food-Drug Interactions from Scientific Literature: Tackling Unspecified Relation

摘要

This paper tackles the problem of mining scientific literature to extract Food-Drug Interaction (FDI). This problem is viewed as a relation extraction task which can be solved with classification method. Since FDI need to be described in a very fine way with many relation types, we face the data sparseness and the lack of examples per type of relation. To address this issue, we propose an effective approach for grouping relations sharing similar representation into clusters and reducing the lack of examples. Since unspecified relations represent more than half the data, we propose to contrast supervised and unsupervised methods to identify the specific relation involved in these examples. The performance of our classification-based labeling approach is twice better than on initial dataset and the data imbalance is significantly reduced. Besides, how learning models combine relations can be interpreted to more effectively group relations.