A Study on Enhancing Zero-Shot Dense Retrieval Using Query and Hypothetical Document Embedding Combination 


Vol. 14,  No. 3, pp. 161-171, Mar.  2025
https://doi.org/10.3745/TKIPS.2025.14.3.161


PDF
  Abstract

Dense retrieval transforms user queries and documents into high-dimensional embedding vectors and calculates their similarity in vector space, enabling effective contextual understanding. It is widely applied in search engines, recommendation systems, and legal domains, outperforming traditional keyword-based methods in handling complex queries. However, it struggles with short, ambiguous queries and adapting to new domains in zero-shot settings. Recent advancements, such as HyDE, leverage large language models (LLMs) to generate hypothetical documents to augment queries. Yet, relying solely on hypothetical embeddings may fail to fully capture user intent. This study proposes a novel framework that combines query and hypothetical document embeddings, dynamically adjusting their contributions based on query complexity. This approach enhances semantic richness and customization for more accurate search results. Experiments on MS MARCO and BEIR datasets show up to 8% performance improvement over HyDE and demonstrate superior nDCG@10 results with dynamic weights compared to fixed-weight methods. This framework offers a scalable, efficient solution applicable to various domains and complex query environments.

  Statistics
  Cite this article

[IEEE Style]

L. Subin and H. Bae, "A Study on Enhancing Zero-Shot Dense Retrieval Using Query and Hypothetical Document Embedding Combination," The Transactions of the Korea Information Processing Society, vol. 14, no. 3, pp. 161-171, 2025. DOI: https://doi.org/10.3745/TKIPS.2025.14.3.161.

[ACM Style]

Lee Subin and Ho Bae. 2025. A Study on Enhancing Zero-Shot Dense Retrieval Using Query and Hypothetical Document Embedding Combination. The Transactions of the Korea Information Processing Society, 14, 3, (2025), 161-171. DOI: https://doi.org/10.3745/TKIPS.2025.14.3.161.

Forms

Search
(IN TITLE, AUTHOR, ABSTRACT,KEYWORDS)

Advanced Search ▼

POPULAR KEYWORDS
(TOP 10 KEYWORDS)

Recent Publications
(LAST 3 YEARS)

Old Journals

Indexing

All publications of TKIPS is indexed in DOI, EBSCO, Google Scholar, Crossref, and CrossCheck.

TKIPS is also selected as the Journal for Accreditation by NRF (National Research Foundation of Korea).

Related Journals