Language Independent Sentence-Level Subjectivity Analysis with Feature Selection

Copyright 2012 by Aditya Mogadala and Vasudeva Varma 26th Pacific Asia Conference on Language,Information and Computation pages 171–180 Language Independent Sentence-Level Subjectivity Analysis with Feature Selection Aditya Mogadala Search and Information Extraction Lab IIIT-H Hyderabad India aditya.m@research.iiit.ac.in Vasudeva Varma Search and Information Extraction Lab IIIT-H Hyderabad India vv@iiit.ac.in Abstract Identifying and extracting subjective informa- tion from News, Blogs and other user gen- erated content has lot of applications. Most of the earlier work concentrated on English data. But, recently subjectivity related re- search at sentence-level in other languages has increased. In this paper, we achieve sentence-level subjectivity classification us- ing language independent feature weighing and selection methods which are consistent across languages. Experiments performed on 5 different languages including English and South Asian language Hindi show that En- tropy based category coverage difference cri- terion (ECCD) feature selection method with language independent feature weighing meth- ods outperforms other approaches for subjec- tive classification. 1 Introduction Subjective text expresses opinions, emotions, senti- ment and beliefs, while objective text generally re- port facts. So the task of distinguishing subjective from objective text is useful for many natural lan- guage processing applications like mining opinions from product reviews (M. Hu and B. Liu, 2004), summarizing different opinions(K. Ganesan.et.al, 2010), question answering (A. Balahur.et.al, 2009) etc. But research work performed earlier on subjectiv- ity analysis has been applied only on English and mostly at document-level and word-level. Some methods (Wiebe and Riloff, 2005) which concen- trated at sentence-level to learn subjective and ob- jective expressions are boot-strapping algorithms which lacks scalability. But, recently focus shifted to multilingual space (R. Mihalcea.et.al, 2007). Banea (C. Banea.et.al, 2008) worked on sentence- level subjectivity analysis using machine translation approaches by leveraging resources and tools avail- able for English . Another approach (C. Banea.et.al, 2010) used multilingual space and meta classifiers to build high precision classifiers for subjectivity clas- sification. However, aforementioned work (C. Banea.et.al, 2008) concentrated more on language specific at- tributes due to variation in expression of subjec- tivity in different languages. This create a prob- lem of portability of methods to different languages. Other approach (C. Banea.et.al, 2010) which tried achieving language independence created large fea- ture vectors for subjectivity classification. Different languages parallel sentences are taken into consider- ation to build high-precision classifier for each lan- guage. This approach not only increases the com- plexity and time for classification but also com- pletely dependent on parallel corpus to get good ac- curacies. A weakly supervised method (C. Lin.et.al, 2011) for sentence-level subjectivity detection us- ing subjLDA tried to reduce training data is avail- able only for English. There are some experi- ments conducted for Japanese (H. Kanayama.et.al, 2006), Chinese (T. Zagibalov.et.al, 2008), Romanian (C. Banea.et.al, 2008; R. Mihalcea.et.al, 2007) lan- guages data. But these approaches are performed at document level and not language independent. In this paper, we try to address three major prob- lems highlighted from earlier approaches. First, can 171 language portability problem be eliminated by se- lecting language independent features. Second, can …

Content truncated to 500 words