Textual entailment inference in machine translation

Translation is about generating semantically equivalent texts in a different
language. Textual entailment is a framework for modeling semantic
equivalences or implications between texts, typically within the same
language. Taking a unified view, one could think of a process where a source
sentence is first transformed to semantically equivalent or implied variants
in the source language, then gets translated to a sentence in the target
language which may be further converted to better-phrased semantic
equivalents in the target-language. Such process may be uniformly viewed as
sequentially generating semantically equivalent or implied forms, where some
of the steps are intra-lingual (entailment) and some are cross-lingual (MT).

In this talk we first review the potential roles of textual entailment
technology along such process. These include generating paraphrases or
semantic generalizations (with some loss of information) of the source or
target sentences, as well as comparing automatic and reference translations
in MT evaluation. Then, we describe in detail a concrete work in which
paraphrases and directional entailments (generalizations) were applied to
overcome the problem of translating unknown words. 



This is a joint work with Wilker Aziz, Nicola Cancedda, Mark Dymetman, Lucia
Specia and Idan Szpektor.