Data and Benchmarking
Research-ready datasets and benchmarks for interconnected living texts.
Learn moreNews
How do experts reason during peer review?... This question is crucial for successful expert-AI collaboration. In the new paper, we rethink peer review as a diagnostic reasoning process. We propose Natural Language Diagnostic Abductive Reasoning as a new family of text-based reasoning tasks, where experts analyze a text step by step to arrive at a verdict. Our unique dataset of over 4000k reasoning steps opens new frontiers in the study of expert-AI collaboration. Have a look at the preprint to learn more!
Are LLMs good classifiers?... To find out, we propose a framework to study LLM fine-tuning for classification with generation- and encoding-based approaches. We apply it to the edit intent classification task and create Re3-Sci2.0: a new large-scale dataset of scientific document revisions with over 94k labeled edits. Have a look at the preprint, while we prepare the camera-ready for EMNLP!
InterText at ACL-2024. Two InterText papers to appear at ACL-2024 in Bangkok! Qian Ruan will present our new dataset and approach for holistic modelling of document revision [1], and Furkan Şahinuç will talk about systematic exploration of creative multi-document NLG tasks in the age of LLMs [2]. While the authors are busy preparing their posters, take a look at the preprints and meet us at the conference!
Introducing M2QA. Language and domain are two major sources of data variation in NLP, motivating the need for joint language-domain transfer. Yet, reliable evaluation remains a challenge. To address this gap, together with colleagues, we created M2QA - a new multi-domain multi-lingual QA benchmark that allows testing for domain and/or language transfer across 4 distinct languages and domains. Find out the details in our preprint, or get the data and start experimenting!
New white paper on NLP for peer review. Peer review is at the core of modern science. Yet it is hard, time consuming and often unfair. What makes peer review challenging, how can NLP help, and where should it stand aside? A new, extensive white paper in collaboration with over 20 high-profile NLP and ML researchers lays the foundation for machine-assisted scientific quality control in the age of AI. The companion repository aggregates datasets for peer review assistance to help new researchers get started. Have a look and contribute!
InterText at EACL-2024. Long documents are often structured, making it much easier for humans to navigate large texts. Is document structure encoded in long-document transformers, and how can their structure-awareness be improved? We investigate this with a novel probing suite and structure infusion kit in our new EACL paper.
Related work from our colleagues. Peer review is one of the core objects of study in InterText. A closely related new work from our colleagues at UKP Lab and the University of Hamburg explores argumentation in peer reviews and rebuttals. Take a look at their pre-print and visit their talk at the upcoming EMNLP!
InterText at EMNLP-2023. We are excited to announce two upcoming EMNLP presentations from InterText. CiteBench is our new benchmark for citation text generation in collaboration with IBM Research. The Argument Mining workshop will host our PragTag shared task. Have a look at our preprints and meet us at the conference!
Join the PragTag-2023 Shared Task. We are hosting PragTag-2023: The First Shared Task on Pragmatic Tagging of Peer Reviews at the Argument Mining Workshop at EMNLP-2023. Learn more and register here!
Three papers accepted at ACL-2023. Our NLPeer, Inclusive Notion of Text and CARE to appear in ACL-2023! 🥳
Team
Iryna Gurevych
Principal Investigator
Ilia Kuznetsov
Postdoc
Martin Tutek
Postdoc
Jan Buchmann
PhD Student
Nils Dycke
PhD Student
Max Eichler
PhD Student
Dennis Zyska
PhD Student
Qian Ruan
PhD Student
Toru Sasaki
PhD Student