From Matching to Generation: A Survey on Generative Information Retrieval

• DOI: 10.48550/arxiv.2404.14851
• Publication Date: 2024-04-23
• Authors: Xiaoxi Li; Jiajie Jin; Yujia Zhou; Yuyao Zhang; Peitian Zhang; Yutao Zhu; Zhicheng Dou
• Keywords: FOS: Computer and information sciences; Computer Science - Computation and Language; Artificial Intelligence (cs.AI); Computer Science - Artificial Intelligence; Computation and Language (cs.CL); Information Retrieval (cs.IR); Computer Science - Information Retrieval

Information Retrieval (IR) systems are crucial tools for users to access information, widely applied in scenarios like search engines, question answering, and recommendation systems. Traditional IR methods, based on similarity matching return ranked lists of documents, have been reliable means information acquisition, dominating the field years. With advancement pre-trained language models, generative retrieval (GenIR) has emerged as a novel paradigm, gaining increasing attention recent Currently, research GenIR can be categorized into two aspects: document (GR) response generation. GR leverages model's parameters memorizing enabling by directly generating relevant identifiers without explicit indexing. Reliable generation, other hand, employs models generate seek, breaking limitations traditional terms granularity relevance matching, offering more flexibility, efficiency, creativity, thus better meeting practical needs. This paper aims systematically review latest progress GenIR. We will summarize advancements regarding model training, identifier, incremental learning, downstream tasks adaptation, multi-modal recommendation, well generation aspects internal knowledge memorization, external augmentation, with citations personal assistant. also evaluation, challenges future prospects offer comprehensive reference researchers field, encouraging further development this area.