Sentence-Level Polarity Classification
Speaker: Michael Wiegand
Abstract:
In this talk, I will present some of my results on feature design for supervised polarity classification at sentence level. Sentence-level classification is more suitable for various NLP applications, such as question answering or text summarization, than document-level classification since these tasks require a more fine-grained classification. Unlike document-level classification, the addition of features more expressive than plain bag of words is effective as, due to data sparseness, bag of words alone provide too little information for a reasonable classification at sentence level. In this context, I will present linguistic features, such as part-of-speech information, clause types, depth of word constituents and WordNet hypernyms, and features containing explicit polarity information obtained from a manually constructed polarity lexicon.