Computational Replication of Human Paraphrase Assessment

Two sentences are paraphrases if their meanings are equivalent but their words and syntax are different. Paraphrasing can be used to aid comprehension, stimulate prior knowledge, and assist in writing skills development. While automated paraphrase assessment is both common-place and useful, research has centered solely on artificial, edited paraphrases and has used only binary dimensions (i.e., is or is-not a paraphrase). In this study, we use 1998 natural paraphrases generated by high school students that have been assessed along 10 dimensions of paraphrase (e.g., semantic completeness). This study investigates the components of paraphrase quality emerging from these dimensions, and examines whether computational approaches (e.g. LSA, MED) can simulate those human evaluations. The results suggest that semantic and syntactic evaluations are the primary components of paraphrase quality, and that computationally light systems such as LSA (semantics) and MED (syntax) present promising approaches to simulating human evaluations of paraphrases.