A two-step hybrid unsupervised model with attention mechanism for aspect extraction

Abstract

Social networking sites have a wealth of user-generated unstructured text for fine-grained sentiment analysis regarding the changing dynamics in the marketplace. In aspect-level sentiment analysis, aspect term extraction (ATE) task identifies the targets of user opinions in the sentence. In the last few years, deep learning approaches significantly improved the performance of aspect extraction. However, the performance of recent models relies on the accuracy of dependency parser and part-of-speech (POS) tagger, which degrades the performance of the system if the sentence doesn't follow the language constraints and the text contains a variety of multi-word aspect-terms. Furthermore, lack of domain and contextual information is again an issue to extract domain-specific, most relevant aspect terms. The existing approaches are not capable of capturing long term dependencies for noun phrases, which in turn fails to extract some valid aspect terms. Therefore, this paper proposes a two-step mixed unsupervised model by combining linguistic patterns with deep learning techniques to improve the ATE task. The first step uses rules-based methods to extract the single word and multi-word aspects, which further prune domain-specific relevant aspects using fine-tuned word embedding. In the second step, the extracted aspects in the first step are used as label data to train the attention-based deep learning model for aspect-term extraction. The experimental evaluation on the SemEval-16 dataset validates our approach as compared to the most recent and baseline techniques.