Opinion Mining in Newspapers for a Media Response Analysis

A part of the broad research domains Knowledge Discovery and Information Retrieval deals only with Data Mining in texts: Text Mining. In general, Text Mining tries to obtain knowledge by identifying patterns in textual data. One of its most important areas is Opinion Mining, which is the main topic of this thesis. Opinion Mining is a far-reaching research area, because it is potentially interesting for many different fields of application as well as its results are very valuable: Opinions are analysed in reviews of products, services, etc. to create very detailed reports about the subject of the reviews or to identify fake or spam reviews. Furthermore, contributions for Opinion Mining in Social Media try to discover opinions in these networks such as Twitter, Facebook, and Youtube. We concentrate on Opinion Mining in news articles, because automatically extracted opinions from news have a high economic value, especially for media monitoring services, but at the same time, this domain has been rather neglected by approaches for Opinion Mining. Thus, we complement this research area by tasks of a Media Response Analysis, which includes the extraction of statements, the classification of the tonality, and the determination of viewpoints. To establish these tasks within the Opinion Mining community, we published an own dataset of a Media Response Analysis (MRA). A major challenge is the extraction of statements for an MRA. In this step, the text parts of a news article have to be identified, which are most relevant for analysis objects and contain opinions, even if the tonality of the opinion is neutral. The classification of the tonality for a given text or text part represents the most difficult task for almost every Opinion Mining approach. Many contributions involve only this step and apply a broad spectrum of techniques to tackle this problem: The creation of sentiment dictionaries, the analysis of contextual information, machine learning, heuristic rules, profoundly linguistic analyses, and many more. During this thesis we investigate many characteristics for the determination of tonality in our domain in contrast to recent research and propose a very well working approach for the tonality classification of statements in newspaper articles, which is adjusted to the requirements of a practical solution and achieves better results for our task than current state-ofthe-art techniques. Extracted and rated statements are difficult to assess for MRA, if they do not contain any information about the viewpoint. To complete a fully automated solution of Opinion Mining for a Media Response Analysis, we explain and evaluate our ontologybased approach for the determination of viewpoints.