A Neural Topic Model Based on Variational Auto-Encoder for Aspect Extraction from Opinion Texts

Aspect extraction is an important task in ABSA (Aspect Based Sentiment Analysis). To address this task, in this paper we propose a novel variant of neural topic model based on Variational Auto-encoder (VAE), which consists of an aspect encoder, an auxiliary encoder and a hierarchical decoder. The difference from previous neural topic model based approaches is that our proposed model builds latent variable in multiple vector spaces and it is able to learn latent semantic representation in better granularity. Additionally, it also provides a direct and effective solution for unsupervised aspect extraction, thus it is beneficial for low-resource processing. Experimental evaluation conducted on both a Chinese corpus and an English corpus have demonstrated that our model has better capacity of text modeling, and substantially outperforms previous state-of-the-art unsupervised approaches for aspect extraction.