“Yes we can?”: Subjectivity Annotation and Tagging for the Health Domain

Proceedings of Recent Advances in Natural Language Processing, pages 666–671, Hissar, Bulgaria, 12-14 September 2011. ”Yes we can?”: Subjectivity Annotation and Tagging for the Health Domain Muhammad Abdul-Mageed Department of Linguistics and School of Library & Info. Science, Indiana University, Bloomington mabdulma@indiana.edu Mohammed Korayem School of Informatics & Computing Indiana University, Bloomington mkorayem@indiana.edu Ahmed YoussefAgha Applied Health Science Department Indiana University, Bloomington ahmyouss@indiana.edu Abstract The area of Subjectivity and sentiment analy- sis (SSA) has been witnessing a flurry of novel research. However, only few attempts have been made to build SSA systems for the health domain. In the current study, we report ef- forts to partially bridge this gap. We present a new labeled corpus of professional articles collected from major Websites focused on the Obama health reform plan (OHRP). We in- troduce a new annotation scheme that incor- porates subjectivity as well as topics directly related to the OHRP and describe a highly- successful SSA system that exploits the anno- tation. In the process, we introduce a number of novel features and a wide-coverage polarity lexicon for the health domain. 1 Introduction In recent years, searches and processing of data be- yond the limiting level of surface words are becom- ing more important than it used to be (Diab et al., 2009). One of the areas that has been witnessing a swelling interest is that of Subjectivity and senti- ment analysis (SSA). Subjectivity in natural language refers to aspects of language used to express opin- ions, feelings, evaluations, and speculations (Ban- field, 1982; Wiebe, 1994) and it, thus, incorporates sentiment. Subjectivity classification refers to the task of classifying texts into either Objective (e.g., The Obama Health Committe submitted a report last week.) or Subjective. Subjective text is further clas- sified with sentiment or polarity. For sentiment clas- sification, the task refers to identifying whether a subjective text is positive (e.g., Obama’s reform plan will solve all our health problems!), negative (e.g., The proposed ideas will lead to definite failure!), neutral (e.g., The president may make changes to some of the ideas proposed.), and, sometimes, mixed (e.g., The plan is bad, but I like Obama.). In spite of the great interest in SSA, only few stud- ies have been conducted on the health domain. The quick dissemination of information characteristic of our world today, makes opinions expressed in these media more important than they traditionally used to be, and hence building SSA systems on top of these media is a valuable endeavor. In the current paper, we present a paragraph-level novel annotation scheme for professional articles from the health do- main that incorporates customized topic annotation. More specifically, we focus on articles treating the Obama Healthcare Reform Plan (OHRP). The rest of the paper is organized as follows: In Section 2, we motivate work on the news genre. In Section 3, we introduce data set, summarize subjec- tivity and topic annotations, and provide examples of categories in our data. In Section 4 we describe our approach. In Section 5, we …

Content truncated to 500 words