Extracting Appraisal Expressions from Short Texts

Short texts such as tweets and E-commerce reviews can reflect people’s opinions on interested events or products, which are much beneficial to many applications. However, one opinion word may have different sentiment polarities when modifying different targets. Therefore, in this paper we propose to extract “

appraisal expressions

” that are represented by tuples of (

opinion word

,

target

), indicating an opinion word and the target modified by the word. By extracting appraisal expressions, we can further construct target-sensitive sentiment dictionaries and improve the effectiveness of sentiment analysis on short texts. Consequently, we propose a

filtering-refinement

framework to extract appraisal expressions from short texts. In the

filtering

step, we extract appraisal-expression candidates, and in the

refinement

step, we use SVM to extract appraisal expressions and present a

dependency-grammar-based

approach to automatically label training data. Comparative experiments between our proposal and three baseline methods suggest the superiority and effectiveness of our proposal.