Semantic Textual Similarity

Through this project, we will explore Semantic Textual Similarity a shared task in Sem Eval 2012. The problem involves comparing a pair of sentences for semantic equivalence, reporting the equivalence on a scale of 0 5 with 0 meaning no equivalence and 5 meaning semantically equivalent. This task is more challenging than determining if the two sentences are paraphrases (a simple yes/no answer), and thus is likely to have much wider applications for a host of NLP tasks like Machine Translation, Question-Answering and Information Retrieval. We build upon the winning 2012 Sem Eval entries, and in addition use a tree-kernel based Support Vector Regression that compares the top-k syntactic parses of the the two sentences as an added similarity measure. Our results seem promising, beating the 2012 winning entry overall, and even more importantly doing significantly better on the surprise test data. We also discuss future extensions that are likely to strengthen our approach.