| dc.description.abstract | "With the increasing amount of reviews and opinionated content available on the internet nowadays, sentiment analysis has garnered enough industry and research interest. The goal of sentiment analysis is to find out the underlying sentiment of such opinionated topics. A more granular approach revolves around accomplishing
sentiment analysis at the aspect level, where researchers try to find out the sentiments behind individual aspects present in such opinionated texts and reviews. In order to perform aspect level sentiment analysis, the foremost thing that requires to be done is aspect extraction. Two paradigms of aspect extraction approaches that are widely used in the research community are frequency based methods and syntax based methods.
Frequency based approach has a drawback, that it is often unable to fetch less frequently talked about aspects. Double Propagation, a popular syntax based method doesn’t perform nearly as good when it comes to extraction of phrasal aspects. We have seen that Double Propagation has a tendency of accumulating noise in the process, resulting in lower precision.
In our work, we try to address the limitation of frequency based approach by introducing a bootstrapping nature to its aspect extraction, where extraction of aspects and opinion words are performed simultaneously. By doing this, we notice an improvement in the accuracy of frequency based approach for aspect extraction. We propose a compound dependency based rule for the Double Propagation algorithm which to make it perform better with the presence of phrasal aspects, which occurs frequently in technology domain. Experimental results have shown an increase in precision and recall across all the datasets used. Additionally, we try to look into the cause of accumulation of noise in Double Propagation of noise and try to address this problem by introducing a frequency based pruning technique and semantic similarity based pruning technique." | |
