In this article, an original view on how to improve phrase translation estimates is proposed. This proposal is grounded on two main ideas: first, that appropriate examples of a given phrase should participate more in building its translation distribution; second, that paraphrases can be used to better estimate this distribution. Initial experiments provide evidence of the potential of our approach and its implementation for effectively improving translation performance. 1

Rephrasing text spans is a common task when revising a text. However, traditional dictionaries often cannot provide direct assistance to writers in performing this task. In this article, we describe an approach to obtain a monolingual phrase lexicon using techniques used in Statistical Machine Translation. A part to be rephrased is first translated into a pivot language, and then translated back into the original language. Models for assessing fluency, meaning preservation and lexical divergence are used to rank possible rephrasings, and their relative weight can be tuned by the user so as to better address her needs. An evaluation shows that these models can be used successfully to select rephrasings that are likely to be useful to a writer. 1

The ability to generate or to recognize paraphrases is key to the vast majority of NLP applications. As correctly exploiting context during translation has been shown to be successful, using context information for paraphrasing could also lead to improved performance. In this article, we adopt the pivot approach based on parallel multilingual corpora proposed by (Bannard and Callison-Burch, 2005), which finds short paraphrases by finding appropriate pivot phrases in one or several auxiliary languages and back-translating these pivot phrases into the original language. We show how context can be exploited both when attempting to find pivot phrases, and when looking for the most appropriate paraphrase in the original subsentential “envelope”. This framework allows the use of paraphrasing units ranging from words to large sub-sentential fragments for which context information from the sentence can be successfully exploited. We report experiments on a text revision task, and show that in these experiments our contextual sub-sentential paraphrasing system outperforms a strong baseline system. 1