Identifying Event-Sentiment Association using Lexical Equivalence and Co-reference Approaches

In this paper, we have identified event and sentiment expressions at word level from the sentences of TempEval-2010 corpus and evaluated their association in terms of lexical equivalence and co-reference. A hybrid approach that consists of Conditional Random Field (CRF) based machine learning framework in conjunction with several rule based strategies has been adopted for event identification within the TimeML framework. The strategies are based on semantic role labeling, WordNet relations and some handcrafted rules. The sentiment expressions are identified simply based on the cues that are available in the sentiment lexicons such as Subjectivity Wordlist, SentiWordNet and WordNet Affect. The identification of lexical equivalence between event and sentiment expressions based on the part-of-speech (POS) categories is straightforward. The emotional verbs from VerbNet have also been employed to improve the coverage of lexical equivalence. On the other hand, the association of sentiment and event has been analyzed using the notion of co-reference. The parsed dependency relations along with basic rhetoric knowledge help to identify the co-reference between event and sentiment expressions. Manual evaluation on the 171 sentences of TempEval-2010 dataset yields the precision, recall and F-Score values of 61.25%, 70.29% and 65.23% respectively.