Generating Multilingual Text

This paper explores a multilingual generation system in the context of machine translation. France Telecom's third-generation rule-based machine translation engine use semantic graphs as an interlingual representation. We show the detailed working process from an interlingual semantic representation to generating multilingual text as a target language: generating proto-trees (quasi-syntactic trees) from a semantic representation, generating unordered syntactic trees from proto-trees, word ordering, and ranking in generated text. We propose a robustness test which we refer to as a transparency test to verify not only the generativity of our system but also the expressivity of the interlingual semantic representation.