An Investigation of Recursive Auto-associative Memory in Sentiment Detection

The rise of blogs, forums, social networks and review websites in recent years has provided very accessible and convenient platforms for people to express thoughts, views or attitudes about topics of interest. In order to collect and analyse opinionated content on the Internet, various sentiment detection techniques have been developed based on an integration of part-of-speech tagging, negation handling, lexicons and classifiers. A popular unsupervised approach, SO-LSA (Semantic Orientation from Latent Semantic Analysis), uses a term-document matrix to detect the semantic orientation of words according to their similarities to a predefined set of seed terms. This paper proposes a novel and subsymbolic approach in sentiment detection, with a level of accuracy comparable to the baseline, SO-LSA, using a special type of Artificial Neural Networks (ANN), an auto-encoder called Recursive Auto-Associative Memory (RAAM).