Multilingual Subjectivity: Are More Languages Better?

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 28–36, Beijing, August 2010 Multilingual Subjectivity: Are More Languages Better? Carmen Banea, Rada Mihalcea Department of Computer Science University of North Texas carmenbanea@my.unt.edu rada@cs.unt.edu Janyce Wiebe Department of Computer Science University of Pittsburgh wiebe@cs.pitt.edu Abstract While subjectivity related research in other languages has increased, most of the work focuses on single languages. This paper explores the integration of features originating from multiple languages into a machine learning approach to subjectiv- ity analysis, and aims to show that this enriched feature set provides for more ef- fective modeling for the source as well as the target languages. We show not only that we are able to achieve over 75% macro accuracy in all of the six lan- guages we experiment with, but also that by using features drawn from multiple languages we can construct high-precision meta-classifiers with a precision of over 83%. 1 Introduction Following the terminology proposed by (Wiebe et al., 2005), subjectivity and sentiment analysis focuses on the automatic identification of private states, such as opinions, emotions, sentiments, evaluations, beliefs, and speculations in natural language. While subjectivity classification labels text as either subjective or objective, sentiment or polarity classification adds an additional level of granularity, by further classifying subjective text as either positive, negative or neutral. To date, a large number of text processing ap- plications have used techniques for automatic sen- timent and subjectivity analysis, including auto- matic expressive text-to-speech synthesis (Alm et al., 1990), tracking sentiment timelines in on-line forums and news (Balog et al., 2006; Lloyd et al., 2005), and mining opinions from product reviews (Hu and Liu, 2004). In many natural language processing tasks, subjectivity and sentiment clas- sification has been used as a first phase filtering to generate more viable data. Research that benefited from this additional layering ranges from ques- tion answering (Yu and Hatzivassiloglou, 2003), to conversation summarization (Carenini et al., 2008), and text semantic analysis (Wiebe and Mi- halcea, 2006; Esuli and Sebastiani, 2006a). Although subjectivity tends to be preserved across languages – see the manual study in (Mi- halcea et al., 2007), (Banea et al., 2008) hypoth- esize that subjectivity is expressed differently in various languages due to lexicalization, formal versus informal markers, etc. Based on this obser- vation, our research seeks to answer the following questions. First, can we reliably predict sentence- level subjectivity in languages other than English, by leveraging on a manually annotated English dataset? Second, can we improve the English sub- jectivity classification by expanding the feature space through the use of multilingual data? Sim- ilarly, can we also improve the classifiers in the other target languages? Finally, third, can we ben- efit from the multilingual subjectivity space and build a high-precision subjectivity classifier that could be used to generate subjectivity datasets in the target languages? The paper is organized as follows. We intro- duce the datasets and the general framework in Section 2. Sections 3, 4, and 5 address in turn each of …

Content truncated to 500 words