
My talk on LLM Evaluation presented at Data on Campus 2 (in Polish) youtu.be/i2wFqIsy-UE?si=ySJ_…

Assistant Professor at @poznanAI. Leader of Conversational Systems Team at the Center for Artificial Intelligence @UAM_Poznan. #AI #NLProc

My talk on LLM Evaluation presented at Data on Campus 2 (in Polish) youtu.be/i2wFqIsy-UE?si=ySJ_…

Our joint study on augmenting spoken language corpora was presented at @FedCSIS
A joint study by @poznanAI researchers and Samsung Electronics Polska engineers was presented at @FedCSIS 2024. The paper investigates the impact of augmenting spoken language corpora with domain-specific synthetic samples. arxiv.org/abs/2406.07090

Today was tle last day of workshops at @aclmeeting conference. Yesterday at @wassa_ws, @PSkorzewski and @piotrjablo represented our department with paper „POLygraph: Polish Fake News Dataset” check full paper at aclanthology.org/2024.wassa-…

We have created POLygraph, a dataset of Polish fake news. Details below:
POLygraph - our team @PSkorzewski @marekkubis @FilipGralinski @piotrjablo and others, prepared a unique resource for fake news detection in Polish which will be presented at @wassa_ws workshop during @aclmeeting 2024 🚀🚀 Paper available at arxiv.org/abs/2407.01393 📄

POLygraph - our team @PSkorzewski @marekkubis @FilipGralinski @piotrjablo and others, prepared a unique resource for fake news detection in Polish which will be presented at @wassa_ws workshop during @aclmeeting 2024 🚀🚀 Paper available at arxiv.org/abs/2407.01393 📄

Our work (@a_karlinska, @cezaryrosinski_, @marekkubis @patt_hub, @JanWieczorekPWr ) on using bibliodata LODification to create metadata-enriched literary corpora in line with FAIR principles is now available at aclanthology.org/2024.lrec-m…

The preprint of our paper "Two Approaches to Diachronic Normalization of Polish Texts" accepted to LaTeCH-CLfL 2024 is now available at arxiv.org/abs/2402.01300 #NLProc #DH

In exactly 20 minutes @marekkubis @PSkorzewski @tzietkiewicz and Marcin Sowański will speak about Back Transcription as a Method for Evaluating Robustness of NLU Models to Speech Recognition Errors. Join us online or in person. We start at 11.00 am CET wmi.amu.edu.pl/zycie-naukowe…

We are participating in the aUPaEU workshop in Turin, Italy, on the presentation of the concept of the Agora. We are a part of a team developing tools for collecting and searching of information for effective cooperation for scientists and HEIs in Europe. @WideningEU @poznanAI

Our work (co-authors: @PSkorzewski, Marcin Sowański, @tzietkiewicz) on using back transcription for evaluating robustness of #NLU models to speech recognition errors was featured on Samsung Research Blog research.samsung.com/blog/Ba… #AI #NLProc @poznanAI @samsungresearch

A joint study of @UAM_Poznan researchers and Samsung Electronics Poland engineers on evaluating robustness of #NLU models to speech recognition errors was presented at #EMNLP2023 by @marekkubis and @tzietkiewicz arxiv.org/abs/2310.16609

The robustness criteria that we formulate are then used to construct a model for detecting speech recognition errors that impact the NLU model in the most significant way.

Contrary to conventional adversarial attacks, which aim at determining the samples that deteriorate the model performance under study, our method also takes into consideration samples that change the NLU outcome in other ways.

The augmented dataset is used to evaluate natural language understanding models and the outcomes of the evaluation serve as a basis for defining the criteria of NLU model robustness.

The method that we propose relies on the use of back transcription, a procedure that combines a text-to-speech model with an automatic speech recognition system to prepare a dataset contaminated with speech recognition errors.

The preprint of our paper "Back Transcription as a Method for Evaluating Robustness of Natural Language Understanding Models to Speech Recognition Errors" accepted to #EMNLP2023 is now available at arxiv.org/abs/2310.16609 #NLProc #VoiceAI #AI