Paraphrase Detection in Monolingual Specialized/Lay Comparable Corpora

Paraphrases are a key feature in many natural language processing applications, and their extraction and generation are important tasks to tackle. Given two comparable corpora in the same language and the same domain, but displaying two different discourse types (lay and specialized), specific paraphrases can be spotted which provide a dimension along which these discourse types can be contrasted. Detecting such paraphrases in comparable corpora is the goal of the present work. Generally, paraphrases are identified by means of lexical and/or structural patterns. In this chapter, we present two methods to extract paraphrases across lay and specialized French monolingual comparable corpora. The first method uses lexical patterns designed according to intuition and linguistic studies, while the second is empirical, based on n-gram matching. The two methods appear to be complementary: the n-gram method confirms the initial lexical patterns and identifies other patterns. Besides, differences in the direction of application of paraphrase patterns highlight differences between specialized and lay discourse.