Aspect-based sentiment analysis of conference review forms

The aim of this thesis is to create a system for extracting opinions and sentiment from conference paper reviews and group these opinions by the different criteria a paper is judged on. The theoretical part of the thesis describes the existing methods of sentiment analysis and natural language processing, thus providing necessary context. The reviewing process of conferences focused on semantic technology and the structure of the reviews are explored. A set of criteria is identified, based on the fields of different conference review forms and used as a foundation for the extraction of terms that are used to express these criteria. A sentiment lexicon is created specifically for the domain of conference paper reviews. In the practical section a dictionary-based sentiment lexicon analysis method is implemented and applied to a set of reviews from 3 different conferences. The results are then evaluated by comparing the numerical scores estimated by the algorithm with the numerical scores from the reviews. The outcome is then explored further, by inspecting the accuracy of criterion identification and sentiment analysis on a sentence level. The precision of criterion identification is evaluated at 57.38 % and the recall at 53.44 %, while the sentiment polarity is correct in over 75 % of cases. The rationale behind this outcome is explained and a set of recommendations is given for future improvements.