Aspect term extraction via BLSTM-CRF with attention on word dependencies

Abstract

This work proposes a neural network architecture using deep learning structures, and minimal feature engineering, to solve the problem of aspect term extraction in opinionated documents, like reviews of products or restaurants. In Opinion Mining and Sentiment Analysis we have three levels of Sentiment Classification: document level, sentence level and aspect level. Aspect Based Sentiment Analysis aims to find the sentiment at the most detailed level, and one of its tasks is the Aspect Term Extraction (ATE): the task of identifying aspects (attributes or characteristics) that have been explicitly evaluated in a sentence. For example, in the sentence “The picture quality of this camera is amazing” the aspect term is “picture quality”. The proposed architecture consists of a BLSTM-CRF classifier with a BLSTM encoder that uses an attention mechanism to permit the incorporation of grammatical relations between words as an additional feature. We also used the Part-of-speech tag (POS tags) as another relevant feature. The proposed architecture obtained state-of-the-art results, with the advantage of using no linguistic rules or lexicons, only minimal feature engineering.