←
Webis Group413 posts
Webis Group avatar

Webis Group

@webis_de

Research group working the fields of Information Retrieval, Natural Language Processing, Data Mining, Machine Learning, and Artificial Intelligence.

◷ Joined September 2019▧ 133 Photos and videos
Webis Group avatar
Webis Group@webis_de·27 Oct 2025

The data spans 7 text domains: 🌐 Web: Wikis, GitHub, social media 💬 Political: Parliamentary proc., speeches ⚖️ Legal: Court decisions, law 📰 News: Newspaper archives 🏦 Economics: Public tenders 📚 Cultural: Heritage collections 🔬 Scientific: Papers, books, journals

Post attachment
Webis Group avatar
Webis Group@webis_de·27 Oct 2025

This means: ✅ Every document has verifiable usage rights (min. CC-BY-SA 4.0 + allows commercial use) ✅ Full institutional provenance for reduced compliance risks ✅ Systematic PII removal + quality filtering, ready for training ✅ Rich metadata for downstream customization

Webis Group avatar
Webis Group@webis_de·27 Oct 2025

The current problem: training data sourced from Web crawls, which give you scale but unclear licensing. This blocks models from commercial deployment and research. We took a different path: systematically collecting text from 41 institutional sources with explicit open licenses.

Maik Fröbe avatar
⟳ Webis Group retweeted
Maik Fröbe@maik_froebe·27 Jun 2025

Do not forget to participate in the #TREC2025 Tip-of-the-Tongue (ToT) Track :) The corpus and baselines (with run files) are now available and easily accessible via the ir_datasets API and the HuggingFace Datasets API. More details are available at: trec-tot.github.io/guideline…

Post attachment
Webis Group avatar
Webis Group@webis_de·22 Jun 2025

Results on BEIR demonstrate that our method matches teacher distillation effectiveness, while using only 13.5% of the data and achieving 3-15x training speedup. This makes effective bi-encoder training more accessible, especially for low-resource settings.

Post attachment
Webis Group avatar
Webis Group@webis_de·22 Jun 2025

The key idea: we can use the similarity predicted by the encoder itself between positive and negative documents to scale a traditional margin loss. This performs implicit hard negative mining and is hyperparameter-free.

Post attachment
Webis Group avatar
Webis Group@webis_de·22 Jun 2025

Our paper on self-distillation for training bi-encoders got accepted at #ICTIR2025! By exploiting pretrained encoder capabilities, our approach eliminates expensive teacher models and batch sampling while maintaining the same effectiveness.

Post attachment