Virus Causes Flu: Identifying Causality in the Biomedical Domain Using an Ensemble Approach with Target-Specific Semantic Embeddings

Identification of Cause-Effect (CE) relation is crucial for creating a scientific knowledge-base and facilitate question-answering in the biomedical domain. An example sentence having CE relation in the biomedical domain (precisely Leukemia) is: viability of THP-1 cells was inhibited by COR. Here, COR is the cause argument, viability of THP-1 cells is the effect argument and inhibited is the trigger word creating a causal scenario. Notably CE relation has a temporal order between cause and effect arguments. In this paper, we harness this property and hypothesize that the temporal order of CE relation can be captured well by the Long Short Term Memory (LSTM) network with independently obtained semantic embeddings of words trained on the targeted disease data. These focused semantic embeddings of words overcome the labeled data requirement of the LSTM network. We extensively validate our hypothesis using three types of word embeddings, viz., GloVe, PubMed, and target-specific where the target (focus) is Leukemia. We obtain a statistically significant improvement in the performance with LSTM using GloVe and target-specific embeddings over other baseline models. Furthermore, we show that an ensemble of LSTM models gives a significant improvement (\(\sim \)3%) over the individual models as per the t-test. Our CE relation classification system’s results generate a knowledge-base of 277478 CE relation mentions using a rule-based approach.