# Siamese BERT-based Model for Web Search Relevance Ranking Evaluated on a New Czech Dataset

**Authors:** Matěj Kocián, Jakub Náplava, Daniel Štancl, Vladimír Kadlec  
**arXiv:** [2112.01810](https://arxiv.org/abs/2112.01810)  
**Date:** 2021-12-03  
**Summary page:** https://www.summarizepaper.com/en/paper/2112.01810v1/

## Key points

- Efficient real-time approach to document ranking problem in web search engines
- Proposal of BERT-based siamese architecture to address computational demands and improve production performance by more than 3%
- Deployment of the model providing highly relevant search results in real-time
- Release of DaReCzech dataset consisting of 1.6 million Czech user query-document pairs with manually assigned relevance levels
- Release of Small-E-Czech language model pre-trained on a large Czech corpus
- Contribution to web search relevance ranking field by introducing efficient approach, valuable datasets, and language models for further research and development

## Summary

The paper titled "Siamese BERT-based Model for Web Search Relevance Ranking Evaluated on a New Czech Dataset" introduces an efficient real-time approach to the document ranking problem in web search engines. To address the challenge of using pre-trained language transformer models like BERT that have high computational demands and are not suitable for serving highly relevant results within milliseconds, the authors propose a BERT-based siamese architecture that improves production performance by more than 3% in a commercial search engine. This model is already deployed and provides highly relevant search results in real-time. In addition to presenting their approach, the authors release two valuable resources for further research and evaluation. The first is DaReCzech, a unique dataset consisting of 1.6 million Czech user query-document pairs with manually assigned relevance levels. This dataset will be instrumental in advancing search relevance research. The second resource released by the authors is Small-E-Czech, an Electra-small language model that has been pre-trained on a large Czech corpus. This model can be used for various multilingual focused research endeavors. Overall, this paper contributes significantly to the field of web search relevance ranking by introducing an efficient real-time approach and providing valuable datasets and language models for further research and development in both search relevance and multilingual focused communities.