Do Neighbours Help? An Exploration of Graph-based Algorithms for Cross-domain Sentiment Classification

Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 655–665, Jeju Island, Korea, 12–14 July 2012. c⃝2012 Association for Computational Linguistics Do Neighbours Help? An Exploration of Graph-based Algorithms for Cross-domain Sentiment Classification Natalia Ponomareva Statistical Cybermetrics Research group, University of Wolverhampton, UK nata.ponomareva@wlv.ac.uk Mike Thelwall Statistical Cybermetrics Research group, University of Wolverhampton, UK m.thelwall@wlv.ac.uk Abstract This paper presents a comparative study of graph-based approaches for cross-domain sentiment classification. In particular, the paper analyses two existing methods: an optimisation problem and a ranking algorithm. We compare these graph-based methods with each other and with the other state-of- the-art approaches and conclude that graph domain representations offer a competitive solution to the domain adaptation problem. Analysis of the best parameters for graph- based algorithms reveals that there are no optimal values valid for all domain pairs and that these values are dependent on the characteristics of corresponding domains. 1 Introduction The sentiment classification (SC) is an active area of research concerned automatic identification of sentiment strength or valence of texts. SC of product reviews is commercially important and widely researched but it typically needs to be optimised separately for each type of product (i.e. domain). When domain-specific data are absent or insufficient the researchers usually seek solution in semi-supervised, unsupervised or cross-domain approaches. In this paper, we focus on cross-domain methods in order to take advantage of the huge amount of annotated sentiment data available on the Internet. Our aim is to find out to what extent it is possible to learn sentiment phenomena from these data and transfer them to new domains rather than induce them from scratch for each new domain. Previous research has shown that models trained on one data usually give much worse results on another, especially when both data sets belong to completely different domains. This is largely because the sentiment words and their valences depend a lot on the domain where they are expressed. The first problem concerns the words that can convey opposite sentiments with respect to the context or domain. For example, a word “ridiculous” in book reviews may express a negative meaning when talking about a book content, however for reviews on electronics this word can bear a positive meaning when talking about prices. Another and more common problem is related to sentiment words that are specific for each domain. For instance, words like “boring”, “inspiring”, “engaging” are very common in book reviews but it is almost impossible to find them in reviews on electronics. At the same time, the electronics domain can contain words like “defective”, “refund”, “return”, “customer service”, which are very unusual for book reviews. Several cross-domain approaches have been suggested recently to solve the problem of accuracy loss in cross-domain sentiment classification, namely Structural Correspondence Learning (SCL) (Blitzer et al., 2007), the graph-based approach (Wu et al., 2009) and Spectral Feature Alignment (SFA) (Pan et al., 2010). In this paper, we explore graph-based algorithms which refer to a …

Content truncated to 500 words