Classification of Opinions with Non-affective Adverbs and Adjectives

International Conference RANLP 2009 - Borovets, Bulgaria, pages 421–427 Classification of Opinions with Non-affective Adverbs and Adjectives Marina Sokolova EHIL, Children’s Hospital of Eastern Ontario Ottawa, Ontario, Canada msokolova@ehealthinformation.ca Guy Lapalme RALI, Universit´e de Montr´eal Montr´eal, Qu´ebec, Canada lapalme@iro.umontreal.ca Abstract We propose domain-independent language pat- terns that purposefully omit the affective words for the classification of opinions. The informa- tion extracted with those patterns is then used to analyze opinions expressed in the texts. Em- pirical evidence shows that opinions can be dis- covered without the use of affective words. We ran experiments on four sets of reviews of con- sumer goods: books, DVD, electronics, kitchen, and house ware. Our results support the practi- cal use of our approach and its competitiveness in comparison with other data-driven methods. This method can also be applied to analyze texts which do not explicitly disclose affects such as medical and legal documents. 1 Introduction Opinion and sentiment analysis has recently received much attention from researchers in the Natural Lan- guage Processing (NLP) and Machine Learning (ML) communities. Most of the research addresses the pri- mary clues of sentiments in a binary setting, e.g. pos- itive/negative word orientation (good vs evil), sub- jective/objective statements (The movie was awesome vs We saw the movie yesterday), texts belonging to positive/negative opinion categories (I recommend this camera . . . vs This book is awful . . . ). In this work, we concentrate on learning opinion from complete texts, classifying the texts as positive or negative. In opinion learning, NLP and ML research mainly concentrates on the emotional polarity of texts which is expressed through the use of affective words (This is an excellent view shows positive polarity, excellent is an affective word). In this work, we, however, pro- pose that learning opinions should allow for the use of the word categories other than affective. We present a method which uses non-affective adjectives and ad- verbs (future, full, perhaps), supplemented by degree pronouns and mental and modal verbs, to determine whether a text bears a positive or a negative opinion label. The method engineers features by using the intrinsic characteristics of a language and avoids extensive and elaborate computational mechanism. Methods used to classify complete texts according to opinions and sen- timents usually employ automated feature selection, e.g., [17, 15]. Although such methods can be applied to different domains, they sometimes involve complex optimization problems, e.g., NP-hard approximation problems [2]. We concentrate on expressions of stance (maybe, necessary), degree (extremely, any), time (ago, now), frequency (rare, again), size (short, high), quantity (many, few), and extent (big). We show that these indicators reliably represent texts in opinion learn- ing. We organize the corresponding word cate- gories – stance/degree/time/frequency adverbs, fre- quency/size/quantity adjectives, degree pronouns – into a semantic hierarchy. Its lowest level works with words; the middle level generalizes word categories into groups; the highest level applies to the text as a whole. The hierarchy avoids the use of emotionally- charged words. We use the hierarchy to …

Content truncated to 500 words