Automatically Facilitating Discussion in Online Forums

Abstract:

Online discussion forums provide users a platform to learn from the collective wisdom of the community. Forum users ask questions, share anecdotal observations with others in the community, in the hope of getting relevant information from them. The growing number of users, turning to these forums for fulfilling their information need and the continuous influx of new topics to discuss, pose significant challenges in making the discussion forums run in an efficient manner. In this thesis, we investigate how Natural Language Processing, and Information Retrieval techniques such as Recommendation Engines can be designed to help the users navigate through the online discussion forums efficiently.

Firstly, we propose building a recommendation system to improve the visibility of threads in online discussion forums. We develop a probabilistic graphical model to consider the interests explicitly mentioned by the user to recommend
her posts she is likely to be interested in. We also show that our framework can provide explanation behind the
recommendations. Unlike traditional recommendation system settings, discussion forum also suffers from new posts
being generated all the time. We propose a deep neural network based framework that can represent a post based
on the words used in it and utilize them to identify the potentially interested users for it. We propose to address
this problem as an Extreme Multi-Class Multi-Labelling problem and show that this formulation works well in
practice.

The open nature of the online forums attracts a large number of users to participate in the discussion. Although
this is desirable, often the large number of posts in response to a discussion topic quickly becomes unmanageable as many repetitive or even irrelevant posts are frequently posted. To this end, we propose a neural network based framework that can identify helpful posts in a discussion thread automatically. Experimenting with large real-world datasets we show that our model performs significantly better compared to existing state-of-the-art solutions.