Product search engines (PSEs) play an essential role in retail websites as they make it easier for users to retrieve relevant products within large catalogs. Despite the continuous progress that has led to increasingly accurate search engines, a limited focus has been given to their performance on queries with negations. Indeed, while we would expect to retrieve different products for the queries “iPhone 13 cover with ring” and “iPhone 13 cover without ring”, this does not happen in popular PSEs with the latter query containing results with the unwanted ring component. The limitation of modern PSEs in understanding negations motivates the need for further investigation.
In this work, we start by defining the negation intent in users queries. Then, we design a transformer-based model, named Negation Detector for Queries (ND4Q), that reaches optimal performance in negation detection (+95% on accuracy metrics). Finally, having built the first negation detector for product search queries, we propose a negation-aware filtering strategy, named Filtering Irrelevant Products (FIP). The promising experimental results in improve the PSE relevance performance using FIP (+9.41% on 𝑛𝐷𝐶𝐺@16 for queries where the negation starts with "without") pave the way to additional research effort towards negation-aware PSEs.
Improving the relevance of product search for queries with negations
2023
Research areas
