Semantic Interpretation of Tweets: A Contextual Knowledge-Based Approach for Tweet Analysis

Tweets are cryptic and often laced with insinuation. Hence, interpretation of tweets cannot be done in isolation. Human beings can interpret the tweets because they possess the requisite Contextual Knowledge. This knowledge enables them to understand the context of tweets and interpret the text. Emulating interpretation ability in machines requires the machine to acquire this contextual knowledge. Tweets pertaining to political and societal issues contain domain-specific terms. Interpretation of such tweets solely on the basis of sentiment orientation of words produces incorrect sentiment tags. Polarity of terms is based on the topic of reference. Thus, an understanding of the pertinent domain terms and their associated sentiment is essential to guide the sentiment mining process. A resource of relevant domain-specific contextual terms and associated sentiments can help to achieve an enhanced sentiment mining performance. With the objective of equipping the machine with the contextual knowledge to facilitate semantic interpretation, we tap the Web resources, process them and structure them as Contextual Knowledge Structures (CKS). We then leverage the CKS to enable a semantic interpretation of tweets. We construct a CKS-based training set to train the Naive Bayes classifier and classify the tweets. We further transform the CKS into sentiment training set (STS) and use it for detecting sentiment polarity tags for tweets. CKS provide the necessary background knowledge pertaining to issues, events, and the related domain-specific terms, thus facilitating semantic sentiment mining. All our experiments are conducted in the context of political/public policy, trending topic, and event-related tweets with an objective of obtaining a pulse of the political climate in India. Our CKS-based classifier exhibits an accuracy of 94.23% in mapping the tweets to the political topic. The distance-based CKS-Sentiment mining algorithm exhibits a consistent performance with an accuracy of 70.90%. The relevance of this contribution is: (a) a novel method which leverages the Web content to derive an optimum training set for tweet analysis, (b) a high degree of Accuracy, Precision, and Recall in tweet classification and sentiment mining with a small CKS-based training set, (c) a topic-adaptive model which can adapt to any domain or topic and exhibit improved tweet analysis performance.