Answer Selection Using Interactive Attention Mechanism

Answer selection is a crucial subtask of a question answering system that focuses on ranking candidate answer sentences from an answer pool based on how relevant and useful it is to answer the question. Conventional deep learning methods like RNNs and CNNs suffer from obtaining local and global context information from the sentence representation. Also, utilizing question context information to generate better answer sentence representation will also contribute to better learning. Thus, our proposed approach uses an interactive attention mechanism using both co-attention for learning question context and self-attention for learning global context. We also adopt the attentive pooling network for compressing features where each element in a question-answer sentence pair can influence the representation of the other. We evaluate our proposals on the TREC-QA dataset and compare using the metrics of MRR and MAP. Our proposed model shows better performance on these evaluation metrics compared to existing baseline models.