Nowadays, users do not only use the user-generated content from various Web-based platforms for entertainment, but also to inform themselves about the latest news topics. Because these topics are disseminated over different platforms, methods are needed which extract those topics and present them in a useful way. Approaches already deal with the extraction of these trending topics. However, usually a summary is created or the content is obtained from a specific platform. With regard to the described issues, this thesis aims to provide answers to the following questions: What are the topics people have been talking about on different Web-based platforms over a specific time period, e.g. the last month? What topics can therefore be extracted as trending topics? What is the temporal correlation between the detected trending topics? How can the system integrate user feedback for a better adaption to time and context? To answer these questions, a prototype consisting of three main components was implemented. The first component, the Topic Analysis, determines the trending topics. To ensure our approach is platform- and language-independent, different Named Entity Recognition tools are combined. In order to get good results, these tools had to be evaluated before the implementation. The second component, Correlation Analysis, calculates the correlations of the previously identified topics. The third component, the Influence Assessment, enables the user to adapt the relevant and non-relevant topics. For this purpose a neural network model, called Word2Vec, is used to determine the explicit and implicit topics. The evaluation showed, that the topics determined by the Topic Analysis component and the identified topics by the Influence Assessment component give a good overview of the user-generated content.
Masterprüfung mit Defensio, Pichler Stefanie
15.11.2018 11:15 - 12:45
Organiser:
SPL 5
Location:
Besprechungsraum 5.35
Währinger Straße 29
1090 Wien
1090 Wien