Semi-supervised vs. Cross-domain Graphs for Sentiment Analysis

Proceedings of Recent Advances in Natural Language Processing, pages 571–578, Hissar, Bulgaria, 7-13 September 2013. Semi-supervised vs. Cross-domain Graphs for Sentiment Analysis Natalia Ponomareva University of Wolverhampton, UK nata.ponomareva@wlv.ac.uk Mike Thelwall University of Wolverhampton, UK m.thelwall@wlv.ac.uk Abstract The lack of labeled data always poses challenges for tasks where machine learn- ing is involved. Semi-supervised and cross-domain approaches represent the most common ways to overcome this dif- ficulty. Graph-based algorithms have been widely studied during the last decade and have proved to be very effective at solv- ing the data limitation problem. This pa- per explores one of the most popular state- of-the-art graph-based algorithms - label propagation, together with its modifica- tions previously applied to sentiment clas- sification. We study the impact of mod- ified graph structure and parameter vari- ations and compare the performance of graph-based algorithms in cross-domain and semi-supervised settings. The re- sults provide a strategy for selecting the most favourable algorithm and learning paradigm on the basis of the available la- beled and unlabeled data. 1 Introduction Sentiment classification is an active area of re- search concerned with the automatic identification of sentiment strength or valence in texts. Being a special case of topic classification, it can bene- fit from all well-known classification algorithms. However, as sentiment classification relies on sen- timent markers rather than frequent topic words, it potentially needs more data for satisfactory perfor- mance. When a limited amount of labeled data is available, cross-domain learning (CDL) or semi- supervised learning (SSL) approaches are com- monly used. CDL techniques endeavour to ex- ploit existing annotated data from a different do- main (i.e. different topic and/or genre) but their success largely depends on how similar the source and target domains are. In contrast, SSL relies on a small amount of labeled data from the same do- main which requires additional annotations. Graph-based (GB) learning has been intensively studied in the last ten years (Zhu et al., 2003; Joachims, 2003; Talukdar and Crammer, 2009; Subramanya and Bilmes, 2011) and applied to many NLP tasks. In particular, in the field of sen- timent analysis GB models have been employed for sentiment classification (Pang and Lee, 2004; Goldberg and Zhu, 2006; Wu et al., 2009), auto- matic building of sentiment lexicons (Hassan and Radev, 2010; Xu et al., 2010), cross-lingual sen- timent analysis (Scheible et al., 2010) and social media analysis (Speriosu et al., 2011). The popu- larity of GB algorithms is not accidental: they not only represent a competitive alternative to other SSL techniques (co-training, transductive SVM, etc.) but also feature a number of remarkable properties, including scalability (Bilmes and Sub- ramanya, 2011) and easy extension to multi-class classification (Zhu et al., 2003). GB algorithms exploit the ability of the data to be represented as a weighted graph where instances are vertices and edges reflect similarities between instances. Higher edge weights correspond to more similar instances and vice versa. GB approaches assume smoothness of the label function on the graph so that strongly connected nodes belong to …

Content truncated to 500 words