Deep stochastic sentence generation : resources and strategies
Year: 2014
Type: dissertation
Abstract: The present Ph.D. thesis addresses the problem of deep data-driven Natural Language Generation (NLG), and in particular the role of proper corpus annotation schemata for stochastic sentence realization. The lack of multilevel corpus annotation has prevented so far the development of proper statistical NLG systems starting from abstract structures. We first detail a methodology for annotating corpora at different levels of linguistic abstraction (namely, semantic, deep-syntactic, surface-syntactic, topological, and morphological levels), and report on the actual annotation of such corpora, manually for Spanish and automatically for English. Then, using the resulting annotated data for our experiments, we train and evaluate deep stochastic NLG tools which go beyond the current state of the art, in particular thanks to the absence of rules in non-isomorphic transductions. Finally, we show that such data can also serve well other purposes such as statistical surface and deep dependency parsing.
Author: Simon Mille
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W2497558203