Statistical Language Generation from Semantic Structures

Semantic stochastic sentence realization is still in its fledgling stage. Most of the available stochastic realizers start from syntactic structures or shallow semantic input structures which still contain numerous syntactic features. This is unsatisfactory since sentence generation traditionally starts from abstract semantic or conceptual structures. However, a change of this state of affairs requires first a change of the annotation of available corpora: even multilevel annotated corpora of the CoNLL competitions contain syntaxinfluenced semantic structures. We address both tasks—the amendment of an existing annotation with the purpose to make it more adequate for generation and the development of a semantic stochastic realizer. We work with the English CoNLL 2009 corpus, which we map onto an abstract semantic (predicateargument) annotation and into which we introduce a novel “deep-syntactic” annotation, which serves as intermediate structure between semantics and (surface-)syntax. Our realizer consists of a chain of decoders for mappings between adjacent levels of annotation: semantic → deep-syntactic → syntactic → linearized→ morphological.