An opinion about opinions about opinions: subjectivity and the aggregate reader

Proceedings of NAACL-HLT 2013, pages 691–696, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics An opinion about opinions about opinions: subjectivity and the aggregate reader Asad Sayeed Computational Linguistics and Phonetics / M2CI Cluster of Excellence Saarland University 66123 Saarbr¨ucken, Germany asayeed@coli.uni-saarland.de Abstract This opinion piece proposes that recent ad- vances in opinion detection are limited in the extent to which they can detect important cat- egories of opinion because they are not de- signed to capture some of the pragmatic as- pects of opinion. A component of these is the perspective of the user of an opinion-mining system as to what an opinion really is, which is in itself a matter of opinion (metasubjectivity). We propose a way to define this component of opinion and describe the challenges it poses for corpus development and sentence-level de- tection technologies. Finally, we suggest that investment in techniques to handle metasub- jectivity will likely bear costs but bring bene- fits in the longer term. 1 Introduction Opinion mining, also known as sentiment analysis (Pang and Lee, 2008), is a relatively recent area of research in natural language processing. It has grown very quickly as a research area, developing around a small number of basic approaches. How- ever, these approaches are based on particular def- initions of opinion, assumptions about opinion ex- pressions, and evaluation practices that we believe need to be expanded in order for sentiment analysis to reach new domains and applications. We are not the first to express concern over the direction of sentiment analysis as a field. This paper seeks to further expand upon the views expressed in Alm (2011) that prevailing evaluation concepts in sentiment analysis limit the kinds of models we can build, particularly through the encouragement of a focus on “high-performing” systems. The central thread that connects our view of the field is the idea that the basis of standard techniques and evaluation in information retrieval and extrac- tion that underlie existing approaches needs to be rethought for applications that are inherently subjec- tive and that the field needs to return to more theoret- ical groundwork. This will entail sacrificing some of the performance gains made in recent times, as well as potentially reducing the capacity for easily com- parable research that has been gained by the rapid adoption of corpora that are very easily produced, shared, and used. This problem is particularly relevant in the expan- sion of sentiment analysis techniques to areas such as market prediction (Bollen et al., 2010) and social science. In these areas, it is not enough to detect opinions in predefined areas of text or even to mine for the locations of opinions in large corpora, but it is necessary to be able to connect opinions across doc- uments and to reconstruct the social networks that underlie social trends. Furthermore, it must be pos- sible to do this in text that can have an arbitrary num- ber of opinions intertwined in ways that go beyond the base case …

Content truncated to 500 words