Paraphrase Extraction from Validated Question Answering Corpora in Spanish

Basing on the debate around the definition of paraphrase, this work aims to empirically clarify what is considered a paraphrase by humans. The experiment accomplished has its starting point in one of the several campaigns that every year generate large amounts of validated textual data, which can be reused for different purposes. This paper describes in detail a simple method –based on pattern–matching and deletion and insertion operations–, able to extract a remarkable amount of paraphrases from Question Answering assessed corpora. An assessment of the corpus obtained was accomplished by experts, and an analysis of this process is shown. This work has been developed for Spanish.