Stop Generating Simple Question as a Query!

Anonymous

Abstract: A prevailing strategy for zero-shot retrieval entails the construction of synthetic queries from documents. However, these generated queries tend to be simple and concise, hence falling short in adequately representing diverse retrieval tasks. An alternative approach harnesses the capability of large language models (LLMs) for in-context learning, enabling the retriever to adapt effectively to the target domain. Nonetheless, such endeavours to discover the unspecified intents demand massive computational resources. In this paper, we challenge the conventional approach of creating simple questions as queries. We propose \textbf{TOPiC}, which directly generates \texttt{task-oriented} queries. TOPiC achieves the highest performance on 6 non-QA datasets, as well as second on entire BeIR benchmark. Our study underscores the potential benefits of incorporating stylistic elements into the query generation procedure.

Paper Type: short

Research Area: Information Retrieval and Text Mining

Contribution Types: NLP engineering experiment, Approaches to low-resource settings, Approaches low compute settings-efficiency

Languages Studied: English

Consent To Share Submission Details: On behalf of all authors, we agree to the terms above to share our submission details.