Low-resource AMR-to-Text Generation: A Study on Brazilian Portuguese

Autor(es)

Sobrevilla Cabezudo, Marco Antonio ; Salgueiro, Thiago Alexandre

• Tipo de documento: Artigo Científico
• Fonte: PROCESAMIENTO DEL LENGUAJE NATURAL; v. N/A, n. 68, p. 13-pg., 2022-03-01.

Resumo

This work presents a study of how varied strategies for tackling low-resource AMR-to-text generation for three approaches are helpful in Brazilian Portuguese. Specifically, we explore the helpfulness of additional translated corpus, different granularity levels in input representation, and three preprocessing steps. Results show that translation is useful. However, it must be used in each approach differently. In addition, finer-grained representations as characters and subwords improve the performance and reduce the bias on the development set, and preprocessing steps are helpful in different contexts, being delexicalisation and preordering the most important ones. (AU)

• Processo FAPESP: 19/07665-4 - Centro de Inteligência Artificial
• Beneficiário: Fabio Gagliardi Cozman
• Modalidade de apoio: Auxílio à Pesquisa - Programa eScience e Data Science - Centros de Pesquisa em Engenharia
• Processo FAPESP: 13/07375-0 - CeMEAI - Centro de Ciências Matemáticas Aplicadas à Indústria
• Beneficiário: Francisco Louzada Neto
• Modalidade de apoio: Auxílio à Pesquisa - Centros de Pesquisa, Inovação e Difusão - CEPIDs