Génération de phrase : entrée, algorithmes et applications
Claire Gardent
Abstract : Sentence Generation maps abstract linguistic representations into sentences. A necessary part of any natural language generation system, sentence generation has also recently received increasing attention in applications such as transfer based machine translation (cf. the LOGON project) and natural language interfaces to knowledge bases (e.g., to verbalise, to author and/or to query ontologies). One outstanding issue in Sentence Generation is what it starts from. What is the abstract linguistic representation it generates from? In my talk, I will explore sentence generation from two main input formats (flat semantic formulae and dependency structures) and discuss their impact on efficiency, algorithms and applications. I will start by describing an algorithm that generates from flat semantic formulae, explain why it is computationally intractable and presenting ways of optimising it to make it usable in practice. I will then show how this algorithm can be used to generate paraphrases; to support error mining and to generate teaching material for language learners from an ontology. In the second part of the talk, I will focus on generation from dependency structures. Based on the input data recently made available by the Generation Challenges Surface Realisation Shared Task, I will show how the algorithm previously used to generate from flat semantic formulae can be adapted to generate from dependency structures. I will moreover discuss various issues raised by the GenChal data such as, missing lexical entries and mismatches between dependency and grammar structures.