A latent aspect mining framework from textual reviews

Abstract

The problem of latent aspect mining from textual customer reviews is important in knowledge discovery and natural language processing. Given a collection of review texts, it is required to automatically determine aspect ratings for each textual review and identify important aspects for all textual reviews. Existing works on aspect-based sentiment analysis has used deep learning techniques with architectures designed based on neural networks. However, they take a lot of time to learn models and need computer configuration to be big enough. This paper proposes a framework (which contains three sub-tasks: (1) aspect term extraction, (2) aspect category detection, (3) aspect ratings and important aspects determination) using simpler and more efficient techniques based on constrained-KMeans algorithm and word2vec. In experiment, we use a data set of 174615 reviews of 1768 hotels with five common aspects including cleanliness, location, service, room and value. Experimental results show that aspects represented by averaging word vectors are more effective than represented by a bag of word. In general, our model outperforms some other benchmark algorithms.