Sentence Answer Selection for Open Domain Question Answering via Deep Word Matching

Abstract

Question answering is a challenging task due to the diversity of expression in human language. This work proposes a novel unsupervised approach for sentence answer selection, called Deep Word Matching (DWM), that uses both the string form and distributed representations of words, thereby capturing their latent semantic relatedness. Our method takes advantage of publicly available linguistic resources and word embeddings in order to explore various levels of word similarity and identify matching concepts between a question and the sentence that contains the answer. By evaluating on three large corpora (SQuAD, NewsQA and WikiQA), we show that the proposed method outperforms previously published baselines and is also task independent, eliminating the need for retraining and tuning on a new task. We obtain improvements between 5% and 11% on target baseline for two tasks, and best result on the third task.