Corpus-wise extraction of syntactic structures for data-to-text generators

Ricci Comella, Josep

Corpus-wise extraction of syntactic structures for data-to-text generators

License

Citation

Ricci Comella, J. Corpus-wise extraction of syntactic structures for data-to-text generators. 2023. handle: http://hdl.handle.net/10230/58317

This citation was generated automatically.

Abstract

Data-to-text NLG consists in converting structured data as found for instance in the Semantic Web (e.g. DBpedia, Wikidata, etc.) into well-formed text in the target language(s). The input to the system is typically a series of triples “Property(Subject, Object)” that encode a wide variety of knowledge types. There are currently different approaches to converting structured data into text. Lately, the most popular ones are neural machine-learning-based techniques, which lack energy efficiency and content accuracy (hallucinations, omissions). On the other hand, symbolic (rule-based) approaches that are very accurate and thus produce reliable output texts when fed with known Properties, are currently less used, in particular, due to their limitations with new (unseen) Properties when corresponding semantic representations are missed. The project aims at addressing this common problem of symbolic approaches by developing a system that converts a given triple into a plausible syntactic structure, to be further used for the creation of the required delexicalized predicate-argument structures. The main tasks of the project are retrieving a text semantically equivalent to the triple from a large corpus connected with a structured knowledge base, and transforming them into generic syntactic representations.