MCINext/query-to-query-sts

Dataset Summary

Query to Query STS (Query2Query) is a Persian (Farsi) dataset for the Semantic Textual Similarity (STS) task. It is part of the FaMTEB (Farsi Massive Text Embedding Benchmark). The dataset was created from real-world anonymized queries issued to the Zarrebin search engine. Semantic similarity scores between query pairs are based on the overlap of their respective search result sets.

• Language(s): Persian (Farsi)
• Task(s): Semantic Textual Similarity (STS)
• Source: Collected from anonymized Zarrebin search engine logs
• Part of FaMTEB: Yes

Supported Tasks and Leaderboards

This dataset is used to evaluate how well models can assess semantic similarity between search queries, which may include typos, abbreviations, or fragmentary text. Evaluation scores are published on the Persian MTEB Leaderboard under the STS category.

Construction

1. Anonymized query logs and corresponding search results were collected from Zarrebin.
2. For each query pair, a similarity score was computed using:
   • Overlap of their search results
   • Normalization via harmonic mean of result counts
3. Scores were grouped into three levels:
   • Unrelated
   • Partially related
   • Fully related
4. Toxic and inappropriate queries were filtered out.
5. Evaluation by human annotators found a correlation of 0.695 between similarity labels and human judgment.

Data Splits

• Train: 4,228,933 samples
• Development (Dev): 0 samples
• Test: 42,717 samples