DynamicRetriever: A Pre-trained Model-based IR System Without an Explicit Index

doi:10.1007/s11633-022-1373-9

CORC > 自动化研究所 > 中国科学院自动化研究所 > 学术期刊 > International Journal of Automation and Computing

	DynamicRetriever: A Pre-trained Model-based IR System Without an Explicit Index
	Yu-Jia Zhou 2; Jing Yao 2; Zhi-Cheng Dou 2; Ledell Wu 1; Ji-Rong Wen 2
刊名	Machine Intelligence Research
	2023
卷号	20 期号:2 页码:276-288
关键词	Information retrieval (IR) document retrieval model-based IR pre-trained language model differentiable search index
ISSN号	2731-538X
DOI	10.1007/s11633-022-1373-9
英文摘要	Web search provides a promising way for people to obtain information and has been extensively studied. With the surge of deep learning and large-scale pre-training techniques, various neural information retrieval models are proposed, and they have demonstrated the power for improving search (especially, the ranking) quality. All these existing search methods follow a common paradigm, i.e., index-retrieve-rerank, where they first build an index of all documents based on document terms (i.e., sparse inverted index) or representation vectors (i.e., dense vector index), then retrieve and rerank retrieved documents based on the similarity between the query and documents via ranking models. In this paper, we explore a new paradigm of information retrieval without an explicit index but only with a pre-trained model. Instead, all of the knowledge of the documents is encoded into model parameters, which can be regarded as a differentiable indexer and optimized in an end-to-end manner. Specifically, we propose a pre-trained model-based information retrieval (IR) system called DynamicRetriever, which directly returns document identifiers for a given query. Under such a framework, we implement two variants to explore how to train the model from scratch and how to combine the advantages of dense retrieval models. Compared with existing search methods, the model-based IR system parameterizes the traditional static index with a pre-training model, which converts the document semantic mapping into a dynamic and updatable process. Extensive experiments conducted on the public search benchmark Microsoft machine reading comprehension (MS MARCO) verify the effectiveness and potential of our proposed new paradigm for information retrieval.
内容类型	期刊论文
源URL	[http://ir.ia.ac.cn/handle/173211/55980]
专题	自动化研究所_学术期刊_International Journal of Automation and Computing
作者单位	1.Beijing Academy of Artificial Intelligence, Beijing 100084, China 2.Gaoling School of Artificial Intelligence, Renmin University of China, Beijing 100872, China
推荐引用方式 GB/T 7714	Yu-Jia Zhou,Jing Yao,Zhi-Cheng Dou,et al. DynamicRetriever: A Pre-trained Model-based IR System Without an Explicit Index[J]. Machine Intelligence Research,2023,20(2):276-288.
APA	Yu-Jia Zhou,Jing Yao,Zhi-Cheng Dou,Ledell Wu,&Ji-Rong Wen.(2023).DynamicRetriever: A Pre-trained Model-based IR System Without an Explicit Index.Machine Intelligence Research,20(2),276-288.
MLA	Yu-Jia Zhou,et al."DynamicRetriever: A Pre-trained Model-based IR System Without an Explicit Index".Machine Intelligence Research 20.2(2023):276-288.