Semantic classification of tweets: A contextual knowledge based approach for tweet classification

In this paper we propose a novel approach and technique for tweet classification based on the Contextual Knowledge Structures (CKS). We first discover the popular, trending topics, then tap the web for the related, relevant content for the topics and harness the same to build CKS. CKS are built using text mining techniques and Computational Linguistics; they are relevant Subject-Predicate-Object triples that depict a specific topic or event. Since the tweets are sparse and most of them do not contain a hashtag, it is difficult to map them to a specific topic. We leverage the CKS to train the Naïve Bayes (NB) classifier and achieve a semantic classification of user tweets. We evaluate the performance of our CKS based classifier by comparing it with the baseline Bag-of-Words (BOW) learning model. The CKS based NB classifier exhibits a consistent performance with an accuracy of approximately 94%. This approach has a two-fold advantage: a) A small training set of knowledge structures is effectively used for machine learning, b) The model adapts to the topic. Trending topics are automatically discovered, the associated CKS are built and the classifier is trained using these dynamic CKS. This model is dynamic, topic-adaptive and efficient.