LEq: Large Language Models Generate Expanded Queries for Searching

Nilanjan Sinhababu, Rabina Khatun

Abstract: Large Language Models have recently been employed everywhere, from personal assistants to professional environments. With the increased capability of these models, it has never been easier for a real synthetic model to replace humans and individual models in most tasks. LLMs are one model for several tasks and this paper explores the LLMs capability to improve upon the existing information searching techniques. It is observed that smaller LLMs($\leq$ 7B) which are practical due to low inference time tend to produce hallucinated and incorrect search results for unlearned domains. To tackle this, we proposed an LLM-based query expansion scheme (LEq) that improves searching drastically. It uses the LLM inferences from the original query and assumes all the generated text tokens as expansion keywords. It then uses an RM3-based keyword weighting scheme for selecting top-m keywords which further reduces model hallucination. We observe a performance improvement of 22.4593% in nDCG@10 and 12.517% improvement in MAP scores across TREC-Deep Learning 2020 and 2019 datasets.

External IDs:dblp:conf/icccnt/SinhababuK24