An Unsupervised Multiple Word-Embedding Method with Attention Model for Cross Domain Aspect Term Extraction

Abstract

Aspect-based sentiment analysis is mainly a domain-dependent problem. In recent years, the existing approaches have the problem of domain adaptation for the poor performance when the domain on which the model is applied is other than the one on which the aspect extraction model is trained. However, a large sample corpus for all domains is required to train an accurate model to extract the most relevant domain-specific aspects. Furthermore, the manual annotation of large labels for the supervised model for the different domain is expensive and time-consuming. In this paper, first, we used an unsupervised method using multiple word embeddings to extract domain-specific aspects belonging to different aspects and then use these aspects as label data to train an attention-based cross-domain model for better prediction. The proposed method is evaluated on SemEval-14 and SemEval-16 datasets and competitive results are shown for baseline and most recent approaches.