Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CWRCzech: 100M Query-Document Czech Click Dataset and Its Application to Web Relevance Ranking

Josef Vonášek, Milan Straka, Rostislav Krč, Lenka Lasoňová, Ekaterina Egorova, Jana Straková, Jakub Náplava

TL;DR

CWRCzech presents a large-scale Czech click dataset (100M Q–D pairs) with rich user-behavior signals (27.6M clicks, 10.8M dwell times) and a 50k-query annotated test set to advance web relevance ranking. The authors compare training with automatically harvested user signals against human-annotated data, showing that, at scale, behavior data can match or surpass annotated data for both cross-encoder and bi-encoder models. They introduce a principled label design (ClickDwellRank) and augment training with soft negatives and contrastive training, achieving up to +4.5 NDCG@10 points over baselines and a roughly linear gain with data size. The work demonstrates the practical value of non-English, large-scale click datasets for IR research and provides guidance on data aggregation, labeling, and training strategies under non-commercial licensing. Overall, CWRCzech significantly lowers the barrier to effective Czech IR model training and offers a model-agnostic resource for future studies.

Abstract

We present CWRCzech, Click Web Ranking dataset for Czech, a 100M query-document Czech click dataset for relevance ranking with user behavior data collected from search engine logs of Seznam$.$cz. To the best of our knowledge, CWRCzech is the largest click dataset with raw text published so far. It provides document positions in the search results as well as information about user behavior: 27.6M clicked documents and 10.8M dwell times. In addition, we also publish a manually annotated Czech test for the relevance task, containing nearly 50k query-document pairs, each annotated by at least 2 annotators. Finally, we analyze how the user behavior data improve relevance ranking and show that models trained on data automatically harnessed at sufficient scale can surpass the performance of models trained on human annotated data. CWRCzech is published under an academic non-commercial license and is available to the research community at https://github.com/seznam/CWRCzech.

CWRCzech: 100M Query-Document Czech Click Dataset and Its Application to Web Relevance Ranking

TL;DR

CWRCzech presents a large-scale Czech click dataset (100M Q–D pairs) with rich user-behavior signals (27.6M clicks, 10.8M dwell times) and a 50k-query annotated test set to advance web relevance ranking. The authors compare training with automatically harvested user signals against human-annotated data, showing that, at scale, behavior data can match or surpass annotated data for both cross-encoder and bi-encoder models. They introduce a principled label design (ClickDwellRank) and augment training with soft negatives and contrastive training, achieving up to +4.5 NDCG@10 points over baselines and a roughly linear gain with data size. The work demonstrates the practical value of non-English, large-scale click datasets for IR research and provides guidance on data aggregation, labeling, and training strategies under non-commercial licensing. Overall, CWRCzech significantly lowers the barrier to effective Czech IR model training and offers a model-agnostic resource for future studies.

Abstract

We present CWRCzech, Click Web Ranking dataset for Czech, a 100M query-document Czech click dataset for relevance ranking with user behavior data collected from search engine logs of Seznamcz. To the best of our knowledge, CWRCzech is the largest click dataset with raw text published so far. It provides document positions in the search results as well as information about user behavior: 27.6M clicked documents and 10.8M dwell times. In addition, we also publish a manually annotated Czech test for the relevance task, containing nearly 50k query-document pairs, each annotated by at least 2 annotators. Finally, we analyze how the user behavior data improve relevance ranking and show that models trained on data automatically harnessed at sufficient scale can surpass the performance of models trained on human annotated data. CWRCzech is published under an academic non-commercial license and is available to the research community at https://github.com/seznam/CWRCzech.
Paper Structure (25 sections, 9 equations, 5 figures, 9 tables)

This paper contains 25 sections, 9 equations, 5 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Datasets
  4. CWRCzech Click Dataset
  5. CWRCzech Annotated Test Set
  6. DaReCzech Dataset
  7. Methodology
  8. Architectures
  9. Cross-encoder
  10. Bi-encoder
  11. Pretrained Models
  12. Training Details
  13. Label Design
  14. Aggregation
  15. Correlations
  16. ...and 10 more sections

Figures (5)

  • Figure 1: Comparison of query length in words (separated by a blank space) between CWRCzech and ORCAS Craswell2020Orcas.
  • Figure 2: Statistics of clicks, dwell times, and ranks within the CWRCzech dataset.
  • Figure 3: A visualization of the proposed architecture of a bi-encoder relevance model.
  • Figure 4: An illustration of a cross-encoder relevance model.
  • Figure 5: Relationship between user behavior dataset size and Small-E-Czech cross-encoder performance on the CWRCzech test set. The x-axis represents number of query-document pairs in the dataset used for training.