InPars: Unsupervised Dataset Generation for Information Retrieval

Autor(es)

Bonifacio, Luiz ; Abonizio, Hugo ; Fadaee, Marzieh ; Nogueira, Rodrigo ; ACM

• Tipo de documento: Artigo Científico
• Fonte: PROCEEDINGS OF THE 45TH INTERNATIONAL ACM SIGIR CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION RETRIEVAL (SIGIR '22); v. N/A, p. 6-pg., 2022-01-01.

Resumo

The Information Retrieval (IR) community has recently witnessed a revolution due to large pretrained transformer models. Another key ingredient for this revolution was the MS MARCO dataset, whose scale and diversity has enabled zero-shot transfer learning to various tasks. However, not all IR tasks and domains can benefit from one single dataset equally. Extensive research in various NLP tasks has shown that using domain-specific training data, as opposed to a general-purpose one, improves the performance of neural models [45, 56]. In this work, we harness the few-shot capabilities of large pretrained language models as synthetic data generators for IR tasks. We show that models fine-tuned solely on our synthetic datasets outperform strong base-lines such as BM25 as well as recently proposed self-supervised dense retrieval methods. Code, models, and data are available at https://github.com/zetaalphavector/inpars. (AU)

• Processo FAPESP: 22/01640-2 - QUEST - sistema de busca e agregação de informações baseado em técnicas Zero-Shot
• Beneficiário: Rodrigo Frassetto Nogueira
• Modalidade de apoio: Auxílio à Pesquisa - Pesquisa Inovativa em Pequenas Empresas - PIPE
• Processo FAPESP: 20/09753-5 - Sistema inteligente para análise de jurisprudência usando técnicas modernas de aprendizado profundo aplicadas ao processamento de linguagem natural
• Beneficiário: Rodrigo Frassetto Nogueira
• Modalidade de apoio: Auxílio à Pesquisa - Pesquisa Inovativa em Pequenas Empresas - PIPE