Opinion Mining in Hungarian based on textual and graphical clues

Opinion Mining aims at recognizing and categorizing or extracting opinions found in unstructured text resources and is one of the most dynamically evolving subdiscipline of Computational Linguistics showing some resemblance to document classification and information extraction tasks. In this paper we propose a novel approach in Opinion Mining which combines Machine Learning models based on traditional textual and graphical clues as well. By examining subjective messages in a given forum topic dealing with a specific voting question, our system makes a prediction about the opinion of unknown people, which can be utilized to predict the forthcoming result of a referendum. The novelty of the work is that beside the regular textual clues (i.e. uni-bigrams), decisions are enhanced by using knowledge derived from a so-called response graph, which represents the interactions between the forum members. Our experimental results showed that with the help of such a graph we were able to achieve better results and significantly outperform the baseline accuracy. The promising results have reinforced our expectations that such an application can be easily adapted to any future Opinion Mining task in the election domain. Key-Words: sentiment analysis, opinion mining, text mining, information extraction, data mining, graph-based techniques