Feature Combination for Sentence Similarity

摘要

The possible combinations of features traditionally used for sentence similarity amount to a very large feature space. Considering all possible combinations and training a support vector machine on the resulting meta-features in a two step process significantly improves performance. The proposed method is trained and tested on the SemEval-2012 Semantic Textual Similarity (STS) Shared Task data, outperforming the task's highest ranking system.