[Back]


Diploma and Master Theses (authored and supervised):

K. Stary:
"Answerer Ranking Framework for Community-Driven Question and Answer Platforms";
Supervisor: S. Dustdar, B. Satzger; Institut für Informationssysteme, AB Verteilte Systeme, 2012; final examination: 2012-06-19.



English abstract:
Web-based Question and Answer communities have become very popular within the last decade. Today, there are huge communities, where tens of thousands of questions are posted every single day. This work aims at improving such communities via automatically forwarding new questions to appropriate answerers. In addition, askers should not be required to categorize their questions. To achieve these goals a framework is presented which contains several components in order to
estimate how suitable users are for answering a specific question. The three core components are Expertise/Knowledge, Authority and Availability, which are used to predict the accuracy, trustworthiness
and response time of a userīs potential answer regarding a particular new question. These components are crucial in the process of determining if a certain user is likely to give a satisfying answer to a specific question. The framework implements different approaches of these components, which can be combined for the user ranking calculation. Concrete realizations of the Expertise component are based on the Vector Space Model (VSM) and on the Query Likelihood
Language Model. VSM is improved via Term Frequency, Inverse Document Frequency (TF-IDF), where IDF is based on the user collection. To the best of our knowledge, this is a novel interpretation, because throughout the literature, IDF is based on the question collection.
Realizations of the Authority component are, for example, InDegree, PageRank, and ZScore. User activity is realized via a novel Activity Filter, which, in all conscience, has not been used by related works. It removes inactive users prior to the actual ranking determination of potential answerers. User profiles, which are the foundation for all these components, are built from previously best answered questions. The implementation of the Ranking Engine, which produces the ranking of potential answers in relation to their likelihood of answering specific new questions, is optimized for modern multi-core processors by leveraging concurrent programming. Based on a dataset, provided by the Yahoo! Research Alliance Webscope program, different ranking variants are evaluated and the most accurate one is determined.

Created from the Publication Database of the Vienna University of Technology.