A Deep Learning Model for Extracting User Attributes from Conversational Texts

Abstract

Extracting user attributes is an important task in a Personal Artificial Intelligence (P.A.I) system to acquire information and knowledge through conversations between the system and humans. In this paper, we proposed a deep learning model for extracting user attributes in the form of SAO triples (subject, attribute, object) from conversational texts in Japanese. We apply a joint CNN-RNN model which combines strength of both Convolution and RNN architectures. In the embedding layer, we propose to combine word, part-of-speech, named-entity, and position embeddings. Experimental results show that the proposed deep learning model outperforms a baseline feature-based model by a large margin.