|
 |
A Self-Training Framework for Automatic Exploratory Discourse Detection
Dr Zhongyu Wei M.Eng., B.Eng. This event took place on 26th June 2012 at 12:30pm (11:30 GMT)
Knowledge Media Institute, Berrill Building, The Open University, Milton Keynes, United Kingdom, MK7 6AA
The webcast was open to 100 users
| |||||||||||||||
 |
 |
With the development of online learning platforms such as SocialLearn, learning resources are uploaded to the internet at a dramatically increasing rate, which makes it difficult for individuals to identify information in need. Ferguson and Buckingham Shum (2011) aim at seeking qualitative understanding of context and meaning of the information and identify “exploratory dialogues” to facilitate users to decide if a resource is useful based on sociocultural discourse analysis (Mercer, 2004). In this project, we extend the previously proposed self-training framework (He and Zhou, 2011) to detect exploratory dialogues from online learning materials automatically. We cast the problem of detection of exploratory dialogues as a binary classification task which classifies a given piece of text as exploratory or non-exploratory. We first train an initial maximum entropy (MaxEnt) classifier based on a small set of manually annotated dataset. The trained classifier is then applied on the large amount of unseen data. Texts classified with high confidence (refer to pseudo-labeled instances) are added into the training data pool for iteratively updating the classifier. Apart from incorporating pseudo-labeled instances directly into the MaxEnt training process, we also explore the use of pseudo-labeled features to constrain the MaxEnt training. Our extensive experiments on the transcribed text from online conferences and the learning paths data downloaded from the SocialLearn platform show that with the self-training framework, the performance of MaxEnt improves significantly. The improvement is more prominent when facing with a smaller number of annotated training instances. The proposed approach will be integrated into the SocialLearn platform for highlighting exploratory discourses in learning paths.
Click below to play the event (50 minutes)