Aspect-based Sentiment Analysis using Semi-supervised Learning in Bipartite Heterogeneous Networks

Abstract

Aspect-based Sentiment Analysis (ABSA) allows to analyze the sentiment from each product aspect, e.g., the camera quality, operating system and the storage capacity of a smartphone. Two main tasks to perform ABSA are: (i) identifying which terms/words are aspects and (ii) performing sentiment analysis for each aspect. Several approaches to treat these tasks are found in the literature, such as those based on the sentiment lexicon, syntactical relations and topic models. The main disadvantages of these methods are the time required and the need of specialists to build the lexicon or to define the rules for different languages and domains. Alternatively, supervised machine learning techniques are employed to perform both aspect identification and analyze the sentiments of the extracted aspects. Although these techniques are language and domain independent and avoid the use of a predefined lexicon or the manual building of rules, the use of supervised learning for ABSA requires labeling a significant amount of aspects and their polarities, which limits the use of such approach in real applications. In this paper we propose an approach to perform ABSA through semi-supervised learning techniques, i.e., a significant less amount of labeled data are required to perform learning. This makes the use of our proposal to perform ABSA easier in real applications. In our proposal we model the data into networks to perform the semi-supervised learning. Specifically, we propose the use of bipartite networks to represent the data since network-based approaches have been successfully used to perform semi-supervised learning and the bipartite networks are parameter-free and fast to be generated. So, this type of network is advisable to be used in practical situations. The results obtained in a rigorous experimental evaluation demonstrate that the proposed approach for ABSA obtains better results than existing approaches based on machine learning for ABSA.