Multi-View AdaBoost for Multilingual Subjectivity Analysis

Proceedings of COLING 2012: Technical Papers, pages 2851–2866, COLING 2012, Mumbai, December 2012. Multi-View AdaBoost for Multilingual Subjectivity Analysis Min X iao Yuhong Guo Department of Computer and Information Sciences Temple University, Philadelphia, PA, USA minxiao@temple.edu, yuhong@temple.edu ABSTRACT Subjectivity analysis has received increasing attention in natural language processing field. Most of the subjectivity analysis works however are conducted on single languages. In this paper, we propose to perform multilingual subjectivity analysis by combining multi-view learn- ing and AdaBoost techniques. We aim to show that by boosting multi-view classifiers we can develop more effective multilingual subjectivity analysis tools for new languages as well as increase the classification performance for English data. We empirically evaluate our two multi-view AdaBoost approaches on the multilingual MPQA dataset. The experimental results show the multi-view AdaBoost approaches significantly outperform existing monolingual and multilingual methods. KEYWORDS: Multi-view learning, AdaBoost, Multilingual subjectivity analysis. 2851 1 Introduction Subjectivity analysis has received increasing interest in natural language processing (NLP) area (Banea et al., 2010; Alm, 2011; Abdul-Mageed and Diab, 2011; Abdul-Mageed et al., 2011). Subjectivity refers to the expression of emotions, sentiments, opinions, beliefs, speculations, evaluations, as well as other private states (Banfield, 1982; Wiebe, 1994). Subjectivity clas- sification aims to distinguish whether a given text expresses subjective or objective meaning (Wiebe and Cardie, 2005; Abdul-Mageed et al., 2011; Banea et al., 2008, 2010). Subjectiv- ity analysis has been intensively studied, particularly motivated by the prevalent need for opinion-related applications, including mining opinions from product reviews (Pang et al., 2002; Hu and Liu, 2004) or political news (Abbott et al., 2011), and recognizing stances in online debates (Somasundaran and Wiebe, 2009, 2010). Moreover, many NLP tasks employ subjectivity analysis as an additional layering to filter data. Research that benefited from this phase ranges from conversation summarization (Seki et al., 2005; Carenini et al., 2008) and information extraction (Riloff et al., 2005) to text semantic analysis (Wiebe and Mihalcea, 2006) and question answering (Li et al., 2008; Yu and Hatzivassiloglou, 2003). Although subjectivity analysis has been widely studied in NLP area, much work has only fo- cused on English data. Recently, some researchers propose to carry out subjectivity analysis in a multilingual framework based on machine translation, where resources or tools of sub- jectivity analysis developed in one language are used to support developing resources or tools in another language (Mihalcea and Banea, 2007; Banea et al., 2008, 2010). The approaches in (Mihalcea and Banea, 2007; Banea et al., 2008) however only exploit the translated target- language-view of the data to develop a subjectivity analysis tool, which is a waste of resources in a multilingual setting since possible parallel views of the data are ignored. Banea et al. (2010) propose to overcome this shortcoming by conducting subjectivity analysis based on concatenated multilingual input feature vectors. This simple feature combination method nev- ertheless is still very preliminary in exploring the capacity of multi-view learning for subjectiv- ity analysis on multilingual data. In this paper, we propose to use multi-view AdaBoost approaches for …

Content truncated to 500 words