FEATURE PHRASE DIRECTED SENTIMENT EXTRACTION FROM TEXT

Abstract

Sentiments about a given subject, be it a commercial product or any other topic of interest can contain a wealth of information not only of the generic opinion toward the subject but also individual opinions of its several features or sub-topics. Automated, context unaware extraction of sentiment polarities toward the individual features that make up the subject has seen extensive research and generally involves two phases. The first phase involves the extraction of feature phrases from the document. This involves the separation of the topic-specific part from the rest of the document. The method chosen for implementation, due to its absolute non-requisite of contextual knowledge, is probabilistic mixture-model based and constructs multinomial distributions of occurrence frequencies for candidate feature phrases, in both the document (topic specific language model), and the whole corpus (generic language model). Then, a maximum likelihood estimation routine is run on the candidate phrases and their respective occurrence frequencies in the corpus and document to filter out the feature phrases from the others. Thus, features are extracted without any knowledge of topic or any other form of context. In the second phase, individual words of a feature phrase are provided to a pre-defined word-sentiment database and the net sentiment polarity is decided from the sum of the results, which provides for a decently accurate estimate owing to the additive nature of the English language (polarity changing words like ‗not‘ are not mined in the features. Due to this, we can assume that the features, along with the sentiment polarities are closely related (semantically speaking) to the document even if they don‘t fall within the crux of what is being conveyed. ). This project will attempt to implement and test this algorithm on different sets of test data and compare the performance results to try and infer the effectiveness of context unaware methods of text mining, in general, for different domains of text data. Keywords: feature phrase, context awareness, sentiment polarity, corpus, language model, occurrence frequency, maximum likelihood estimation, word-sentiment database.