Open Domain Question Answering with Character-Level Deep Learning Models

摘要

Single-relation factoid question answering (QA) is strongly supported by rich sources of facts from knowledge bases (KB). However, there are many irrelevant information in questions and overwhelming number of facts in knowledge bases, mak-ing it difficult to capture goal entity and relation involved in a question. In order to settle these issues, firstly, a state-of-the-art sequence tagging model (BiLSTM-CRF) is adopted to detect the entity mention in a question. Then, we propose a n-gram match (NGM) algorithm with Chinese-specific rules and an attention-based siamese bidirectional long-short term memory (ASBLSTM) model to measure the lexical and se-mantic similarity between questions and candidate facts. Our whole method requires no hand-crafted template or feature engineering. In addition, character-level models are proved to be effective in solving the out of vocabulary (OOV) issue and improving the accuracy in Chinese KBQA task. Experiment results show that our system outperforms the best system with deep learning models in the KBQA share task of the Confer-ence on Natural Language Processing and Chinese Compu-ting (NLPCC) 2016 and our system achieves an AverageF1 measure of 80.97% and 37.18% on test dataset in NLPCC 2016 and 2017 respectively.