Year: 2024
Type: dissertation
Abstract: Doctor ofin BM25 together.Through experiments, we have shown that SPLADE with AdaLM is, on average, more accurate in zero-shot retrieval than existing methods in the bio-medical and scientific domains, where the vocabulary and word frequency differ significantly from the source domain.In addition, we showed that AdaLM is more accurate than pseudo-queries, which is an effective unsupervised domain adaptation method for retrieval.This verified the effectiveness of AdaLM as an unsupervised domain adaptation method.Finally, since AdaLM can be applied to all IR models that use pre-trained language models, we applied it to several IR models, including dense retrieval, and demonstrated its effectiveness.We also showed that applying AdaLM to multiple retrieval models and ensembling them further improved the retrieval accuracy.
Authors: 飯田, 大貴, Iida, Hiroki
Institution: Tokyo Institute of Technology
Language: Japanese
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7146228176
