Identifying Emotion Labels from Psychiatric Social Texts Using Independent Component Analysis

Accessing the web has been an efficient and effective means to acquire self-help knowledge when suf- fering from depressive problems. Many mental health websites have developed community-based ser- vices such as web forums and blogs for Internet users to share their depressive problems with other us- ers and health professionals. Other users or health professionals can then make recommendations in re- sponse to these problems. Such communications produce a large number of documents called psychiat- ric social texts containing rich emotion labels representing different depressive problems. Automatically identify such emotion labels can make online psychiatric services more effective. This study proposes a framework combining latent semantic analysis (LSA) and independent component analysis (ICA) to ex- tract concept-level features for emotion label identification. LSA is used to discover latent concepts that do not frequently occur in psychiatric social texts, and ICA is used to extract independent components by minimizing the term dependence among the concepts. By combining LSA and ICA, more useful la- tent concepts can be discovered for different emotion labels, and the dependence between them can also be minimized. The discriminant power of classifiers can thus be improved by training them on the inde- pendent components with minimized term overlap. Experimental results show that the use of concept- level features yielded better performance than the use of word-level features. Additionally, combining LSA and ICA improved the performance of using each LSA and ICA alone.