Suggestions
Vasily Konovalov
PhD in Computer Science | Kaggle Competition Expert
Vasily Konovalov is a highly accomplished professional in the field of Natural Language Processing (NLP) and Machine Learning. He currently holds the position of Staff Research Scientist at AIRI (Artificial Intelligence Research Institute).1
Key aspects of Vasily Konovalov's profile include:
Education and Expertise
- Holds a PhD in Computer Science from Moscow Institute of Physics and Technology (State University)
- Obtained a Master's Degree in Computer Science from Bar-Ilan University
- Specializes in Natural Language Processing, Machine Learning, and Dialogue Systems3
Professional Experience
- Staff Research Scientist at AIRI
- Former NLP Team Lead at DeepPavlov2
- Previously worked as an NLP Research Engineer at Bar-Ilan University4
Research and Contributions
- Active researcher with numerous publications in the field of NLP and AI
- Google Scholar profile shows 113 citations for his work3
- Recent projects focus on detecting LLM-generated text, LLM hallucinations, and fine-tuning LLMs for specific tasks
Online Presence
- LinkedIn username: vaskonov4
- Twitter handle: @vaskon5
- Actively shares insights and research findings on professional platforms2
Vasily Konovalov is recognized as a passionate NLP expert with a strong background in open-source development. His work at AIRI contributes to the institute's mission of creating universal artificial intelligence systems that solve real-world problems.1
Highlights
With a slight delay, I'd like to present our joint solution with Anastasia Voznyuk for #SemEval-2024 Task 8, which focused on detecting the boundary between human-written and LLM-generated text. We proposed a straightforward supervised approach, where our DeBERTa-based model, fine-tuned on limited training data, classifies the boundaries. However, one may wonder: if it's a supervised problem, why not simply augment the training data in other words cut data around the boundaries? By adopting this strategy, we achieved a MAE of 13.38 whereas organizers baseline is 23.16.
More details can be found in our paper https://lnkd.in/eVzQecsh
I'd like to briefly promote our solution for #TextGraphs-17, which we developed with Julia B. and Evgeniy Beliakin, and with which we achieved 🚀 2nd place in the competition (incl post-competition). As it turns out, simply incorporating as much knowledge as possible into the LLM context and employing RAG isn't the most efficient method. Instead, we propose selecting the most useful context through techniques from Uncertainty Estimation. In our experiments, a simple margin probability approach, comparing the best two candidates for each context source, demonstrated the best performance.
More details in our paper https://lnkd.in/eQUSGwx4
And don't forget to check out other solutions to TextGraph-17 task at #ACL2024
