Adapting Sentiments with Context Flávio Ceci a , Rosina O Weber b , Alexandre L Gonçalves a , Roberto C S Pacheco a a Federal University of Santa Catarina, Florianopolis, SC, Brazil b Drexel University, Philadelphia, PA, USA flavio.ceci@unisul.br, a.l.goncalves@ufsc.br, rosina@drexel.edu, pacheco@egc.ufsc.br Abstract. Users of sentiment analysis applications are interested in opinions of individual aspects (e.g., excellent mileage) of a target entity rather than in their polarity (e.g., 56% are positive). This analysis is known as aspect-level sentiment analysis. In this paper, we use document-level polarization to learn patterns for contextual polarity, which refers to finding whether a sentiment bearing word changes polarity in a given context. For example, in it is cheap looking, cheap is negative; in this is good quality and cheap, cheap is positive. Our proposed case- based approach for sentiment analysis assesses contextual polarity of individual aspects in the adaptation step leading to increased classification accuracy. Key words: contextual polarity, aspect-based sentiment analysis, context, textual case-based reasoning, sentiment lexicon, ontology. 1 Introduction and Background Sentiment analysis is a valuable application of text classification because of the high volume of crowdsourced online content [1]. The typical texts targeted by sentiment analysis systems consist of opinions about an entity (e.g., individual, product, service). The value of sentiment analysis goes beyond classifying the polarity of a document; it resides in providing the aspects of the entity being reviewed and the sentiment asso- ciated with these aspects, which is known as aspect-level sentiment analysis (e.g., [2], [3]). For example, the value of the analysis of the opinion “The flash recovery time is ridiculously slow, but that I can live with, the 4 out of every 5 pictures that come out blurry I cannot” is not that its overall sentiment is negative, but it is that flash recovery time and blurry pictures are both negative aspects. Consumers are interested in the sen- timent associated to aspects of multiple opinions to make decisions about products, organizations or people [1][3]. To be useful, systems should thus aggregate information contained in multiple opinions. Furthermore, the literature [1] tells us that users prefer structured visualizations that summarize opinions over textual summaries. Fig. 1 shows the application context where a user is interested in opinions about a product x. A filter F(x) produces n opinions pi on product x (i=1,.., n) and submits each opinion pi as new case ci. Cases are pairs P, S where the problem P is an opinion pi and a solution S is a triple S(ti, aij, C) where ti are sentiment trees, aij are m (j=1,.., m) polar- ized aspects, and C is the global polarization of the opinion. The case-based sentiment