Aligning Formal Meaning Representations with Surface Strings for Wide-coverage Text Generation

International audienceStatistical natural language generation from abstract meaning representations presupposes large corpora consisting of text–meaning pairs. Even though such corpora exist nowadays, or could be constructed using robust semantic parsing, the simple alignment between text and meaning representation is too coarse for developing robust (statistical) NLG systems. By reformatting semantic representations as graphs, fine-grained alignment can be obtained. Given a precise alignment at the word level, the complete surface form of a meaning representations can be deduced using a simple declarative rule