Journal Article

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

Dou, Zhi-Cheng, Yao, Jing, Wu, Ledell, Wen, Ji-Rong, Zhou, Yu-Jia

2023

Overview

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.

• Publisher: Springer Nature B.V
• Subject: Documents / Information retrieval / Ranking / Search methods / Training