Composing Measures for Computing Text Similarity

We present a comprehensive study of computing similarity between texts. We start from the observation that while the concept of similarity is well grounded in psychology, text similarity is much less well-defined in the natural language processing community. We thus define the notion of text similarity and distinguish it from related tasks such as textual entailment and near-duplicate detection. We then identify multiple text dimensions, i.e. characteristics inherent to texts that can be used to judge text similarity, for which we provide empirical evidence. We discuss state-of-the-art text similarity measures previously proposed in the literature, before continuing with a thorough discussion of common evaluation metrics and datasets. Based on the analysis, we devise an architecture which combines text similarity measures in a unified classification framework. We apply our system in two evaluation settings, for which it consistently outperforms prior work and competing systems: (a) an intrinsic evaluation in the context of the Semantic Textual Similarity Task as part of the Semantic Evaluation (SemEval) exercises, and (b) an extrinsic evaluation for the detection of text reuse. As a basis for future work, we introduce DKPro Similarity, an open source software package which streamlines the development of text similarity measures and complete experimental setups.