Ubiquitous Knowledge Processing (UKP) Lab

𝗧𝗵𝗿𝗼𝘄𝗯𝗮𝗰𝗸 𝘁𝗼 #EMNLP2024 𝗶𝗻 𝘀𝘂𝗻𝗻𝘆 𝗠𝗶𝗮𝗺𝗶🌞 The UKP Lab had an amazing time at this year’s #EMNLP in Florida🌴! Our team presented a total of 13 papers, including 11 in the Main track and 2 in the Findings track, showcasing our latest research to a vibrant international audience. Beyond the exciting presentations and poster sessions, the conference was a fantastic opportunity to connect with researchers from around the world. The social event at the Phillip and Patricia Frost Museum of Science was a highlight, gathering more than 3,000 attendees for a memorable evening. Catching up with colleagues from Technische Universität Darmstadt, hessian.AI, MBZUAI (Mohamed bin Zayed University of Artificial Intelligence), and many other institutions, while exploring one of the most stunning aquariums, made the long queue to enter more than worthwhile. Workshops like FEVER and Blackbox NLP sparked engaging discussions. Congratulations to our colleagues Max Glockner & Andreas Waldis for receiving an Outstanding Reviewer Award – well-deserved recognition for his dedication to the field!🎉 And many thanks to Jonathan Tonglet  for his impressions and photos. We’re already looking forward to #EMNLP2025 in Suzhou!

Welcome to Dr. Simone Balloccu who has joined us and Technische Universität Darmstadt to lead a junior independent research group on ‘#NLProc in Expert Domains’ associated with the related ongoing Research Center proposal #ExAIC of the TU Darmstadt and Ludwig-Maximilians-Universität München. Simone obtained his PhD at the University of Aberdeen and has worked until recently at the Institute of Formal and Applied Linguistics, Charles University. His research focuses on AI for behaviour change, safety and controllability evaluation, and more generally on AI applied to expert domains. Within ExpNLP, he will work on the cooperation between AI agents and domain experts for domains like sw engineering, data science and machine learning. Interested in how NLP can help domain experts? Get in touch with Simone Balloccu to discover research opportunities at #ExpNLP!

💡 𝗛𝗼𝘄 𝗰𝗮𝗻 𝘄𝗲 𝘁𝗮𝗰𝗸𝗹𝗲 𝘁𝗵𝗲 𝗴𝗿𝗼𝘄𝗶𝗻𝗴 𝗰𝗵𝗮𝗹𝗹𝗲𝗻𝗴𝗲 𝗼𝗳 𝘃𝗶𝘀𝘂𝗮𝗹 𝗺𝗶𝘀𝗶𝗻𝗳𝗼𝗿𝗺𝗮𝘁𝗶𝗼𝗻? Jonathan Tonglet from the UKP Lab at Technische Universität Darmstadt and from LOEWE emergenCITY's Information (INF) Program Area shares his insights on detecting and combating issues like deepfakes and out-of-context images.

I am proud to announce that summetix has been awarded the Technische Universität Darmstadt Spin-off label, which is awarded by the university and its startup hub HIGHEST. The award is given to a selective group of startups that have come out of TU Darmstadt, and meets related criteria. Thank you to Rainer Hetzer/ Horst ;) for the great pitch together   As a startup born out of a research project at TU Darmstadt, we will wear this recognition with pride! We would like to express our heartfelt thanks to the Technical University of Darmstadt and the team from HIGHEST – the Innovation and Start-up Center and Ubiquitous Knowledge Processing (UKP) Lab – for your support. Sabine Remmert Simone Eva Maria Lühl Iryna Gurevych Harald Holzer Gudrun Lantelme Bartosz Kajdas Tristan Miller Andreas Rücklé Nils Reimers Christian Stab Chris Stahlhut Nandan Thakur Tilman Beck With: Maximilian Fries - thank you for keeping our price cool! MAGNOTHERM Dr. Rafael Fietzek Montgomery Wagner Carmen Rommel Johannes Daxenberger, Benjamin Schiller, Simon Sundberg, Bilge Nur Yılmaz, Tim Altendorf,Andreas Lukic, Max Gudzent , Tim Kaiser, Leonie Kollmar, Jojo Danyun Zhou Foto: HIGHEST/ Benjamin Schenk 📷 Fotocredit: HIGHEST/Benjamin Schenk #Innovation #StartUp #innovation #llm #argumentmining #donthaveyourheadinthewordcloud

𝗨𝗞𝗣 𝗟𝗮𝗯 𝗰𝗼𝘂𝗿𝘀𝗲𝘀 𝘀𝗵𝗶𝗻𝗲 𝗶𝗻 𝘁𝗼𝗽 𝗿𝗮𝗻𝗸𝗶𝗻𝗴𝘀 🌟 The student council for Computer Science, TU Darmstadt has just published the lecture rankings for the past summer semester, which are based on aggregated evaluation results. We are happy to report that all of our courses – DL4NLP, Ethics4NLP and InfMan – rank among the top 7 of 23 computer science courses and DL4NLP is even among the top-rated courses. It is nice to see that our work resonates with students 😊

🌍 How can Europe navigate AI’s societal, scientific, and political challenges?  On the 8th of November 2024, Prof. Iryna Gurevych, Director of Ubiquitous Knowledge Processing (UKP) Lab, has shared her insights at the 𝗙𝗮𝗹𝗹𝗶𝗻𝗴 𝗪𝗮𝗹𝗹𝘀 𝗦𝗰𝗶𝗲𝗻𝗰𝗲 𝗦𝘂𝗺𝗺𝗶𝘁 in Berlin. 💡 She has joined the round table talk with - Georg Schütte (VolkswagenStiftung) - Martin Hullin (Bertelsmann Stiftung) - Sarah de Rijcke (Leiden University) - Michael Schönstein (German Federal Chancellery) More information on the round table at the #FallingWalls Science Summit 🔗 https://lnkd.in/dGSmZ6v2

🎙️ At ML in PL 2024 in Warsaw, Iryna Gurevych tackled a pressing challenge: novel AI techniques to fight misinformation. 🌐 Her talk highlighted how NLP and multimodal AI models can combat misleading narratives by: 1️⃣ Identifying why misinformation seems credible.  2️⃣ Explaining why it’s false.  3️⃣ Offering accurate alternative explanations. From detecting manipulated visuals to debunking misused scientific studies, this approach strengthens our defense against harmful misinformation. 💡 👉 Learn more about the conference https://lnkd.in/ewNgFb_y #UKPLab #MLinPL #NLP #AI #misinformation

On November 19, Iryna Gurevych, will speak at the ATHENE-Center conference “Cognitive Security” in the session “The Perspective of Science”. Her topic: 𝗗𝗶𝘀𝗶𝗻𝗳𝗼𝗿𝗺𝗮𝘁𝗶𝗼𝗻 𝗮𝗻𝗱 𝗙𝗮𝗰𝘁-𝗖𝗵𝗲𝗰𝗸𝗶𝗻𝗴 👇 🗓️ November 19 📍 Jügelhaus, Frankfurt am Main 👉 Find more details about the conference and the program here: https://lnkd.in/eViHax83 The conference is being held with the support of Goethe-Universität Frankfurt, Stiftung Mercator GmbH & Fraunhofer Institute for Secure Information Technology SIT #ATHENE #TUDa #UKPLab #CognitiveSecurity #Disinformation #FactChecking

If you're interested in leaderboard generation or #AI4Science themes, please feel free to stop by our poster at #EMNLP2024 in Riverfront Hall (Nov. 14 Thursday 14:00 - 15:30) #NLProc

Wanna facilitate text classification with LLMs? Curious if they make the cut as top-tier classifiers? (Fig. 1) We take on the challenging task of edit intent classification (EIC), and test generalizability across five other tasks. We show how LLM classifiers can accelerate large-scale dataset creation! ⏫ Learn more about our #EMNLP2024 paper in this post. #NLProc 📰 https://lnkd.in/ecFUSYWq Our framework for fine-tuning LLMs for classification combines generation-based and three encoding-based methods. It also allows for a systematic comparison of various training strategies, exploring three input types, five transformation functions, and different language models. Two of the approaches are: - Gen casts the task as text generation, aiming to produce the label as part of the output. - SeqC uses a linear classification layer, treating the task as sequence classification based on the last hidden state. (Fig. 2) There’s also: - SNet employs a Siamese architecture, converting outputs into a single embedding for classification. - XNet processes both sentences simultaneously through a single LLM. (Fig. 3 & 4) ✨ LLMs can be fine-tuned to operate as good classifiers, achieving SOTA. ✨ SeqC has superior performance, inference efficiency, and perfect answer inclusion.  ✨ 13B Llama2 and 8B Llama3 models show the most potential.  ✨ We validate the generalizability and significance of our findings on five additional tasks. (Fig. 5) We applied our most efficient and high-performing model to create Re3-Sci2.0: a new large-scale dataset of 1,780 scientific document revisions with over 94k labelled edits,  spanning various scholarly domains. (Fig. 6)  The main insights of Re3-Sci2.0 are: 🔵 Successful revisions involve more edits.  🔵 Clarity and Claim modifications and Fact/Evidence additions significantly boost revision success.  🔵 Human editing focus and location stay consistent within the same research domain but vary across domains. --- We offer open access to our code, paper, and dataset: 📄Paper: https://lnkd.in/ecFUSYWq   💻Code: https://lnkd.in/ep3ZFmCT  🗂️Data: https://lnkd.in/e66SZNEx --- And consider following the authors Qian Ruan, Ilia Kuznetsov and Iryna Gurevych (all Ubiquitous Knowledge Processing (UKP) Lab Lab) if you are interested in more information or an exchange of ideas. Poster presentation will take place on Wednesday 7.45 am. (virtual) See you in Miami! #EMNLP2024