Investigating the Performance of Dense Retrievers for Queries with Numerical Conditions

Haruki Fujimaki, Makoto P. Kato

Abstract: This study investigates the performance of dense retrieval models for queries with numerical conditions, such as “mountains higher than 2000 m” and “laptops between 500 and 1000 dollars.” Our experimental results revealed that while dense retrieval models were able to change search results according to numerical conditions, there was a large gap between the ideal performances and those achieved by dense retrieval models. We also found that the retrieval effectiveness varied between models depending on the type of numerical condition and expression. Furthermore, experimental results suggest that knowledge acquired during pre-training might influence the search results for queries containing numerical conditions, which could lead to biased search results. This study highlights the need for further improvements in the capability of dense retrieval models to handle numerical information, which is often required in e-commerce, medical, and financial domains.