Comparing semantic associations in sentences and paragraphs for opinion detection in blogs

Opinion Detection is one of the most interesting and challenging work in the field of Information Retrieval. Lot of research work already exists in this area with some distinctive work. A review of the reveals that researchers have been working on different levels of granularity like documents, passages, sentences and words for the task of opinion detection. In this work we revise our previous approach that combines document level heuristics with a semantic similarity based method. We evaluate this semantic similarity approach on a huge data collection using three different setups involving both sentences and passages and then compare the performance of our approach with these different setups. For evaluation purposes, we are using TREC Blog 2006 collection (148 GB) with 50 topics of TREC Blog 2006 over baseline obtained through Terrier Information System Platform. Results show that our approach improves the baseline opinion MAP by 28.89%, 30.13% and 32.26% using setup one, two and three respectively.