A Hybrid Embedding Approach to Noisy Answer Passage Retrieval

Answer passage retrieval is an increasingly important information retrieval task as queries become more precise and mobile and audio interfaces more prevalent. In this task, the goal is to retrieve a contiguous series of sentences (a passage) that concisely addresses the information need expressed in the query. Recent work with deep learning has shown the efficacy of distributed text representations for retrieving sentences or tokens for question answering. However, determining the relevancy of answer passages remains a significant challenge, specifically when there exists a lexical and semantic gap between the text representation used for training and the collection’s vocabulary. In this paper, we demonstrate the flexibility of a character based approach on the task of answer passage retrieval, agnostic to the source of embeddings and with improved performance in P@1 and MRR metrics over a word based approach as the collections degrade in quality.