Multi-granularity Embedding for Question Classification using Recurrent Neural Network

• 1、State Key Laboratory of Networking and Switching Technology, Beijing University of Posts and Telecommunications, Beijing 100876

Abstract：The open-field-oriented retrieval-based question answering system involves a variety of tasks, including question analysis, information retrieval, answer extraction, answer combination, answer ranking, etc., of which the performance of the key part of the question analysis directly affects the accuracy of the answer extraction. Based on the existing syntactic structure analysis results, this paper proposes a MGE-RNN (Multi-granularity Embedding for Question Classification for Recurrent Neural Network) . This algorithm automatically learns distributed representations of the backbone and interrogative with its their ancillary components, and also fits the non-linear relationship between integrated multi-granularity features and question classification. The loss functions include training loss of multi-granularity word embeddings and classification error in Softmax classifier, and finally MGE-RNN performance is better than the state-of-the-art question classifier on a open dataset.

Keywords： Question answering system Question classification Multi-granularity embedding Syntactic structure