The NLP Seminar is organised by the Linguistic Engineering Group at the Institute of Computer Science, Polish Academy of Sciences (ICS PAS). It will restart in October and will take place on (some) Mondays, usually at 10:15 am, often online – please use the link next to the presentation title. All recorded talks are available on YouTube.

15 September 2025

Louis Esteve (Universite Paris-Saclay)

Diversity and dataset size – a quantitative perspective  

The field of Natural Language Processing (NLP) studies the abilities of computer systems to process and generate natural language, and has received increasing attention from the general population since the democratisation of generative and conversational models. However, behind the scenes, state-of-the-art NLP models are trained on ever-larger datasets, reaching trillions of tokens. It may be argued that the creation and use of such immense datasets is motivated by the idea that 'the larger the dataset, the more diverse it is', and that in turn 'if the training set is more diverse, it shall yield better models'. However, these statements thus far remain intuitions and need to be properly tested. To this end, this presentation will tackle methods and caveats of formal diversity quantification including limitations of the literature, a preliminary discussion on the link between diversity and dataset size, as well as their impact on downstream applications.

6 October 2025

Stan Matwin (Dalhousie University / Institute of Computer Science, Polish Academy of Sciences)

Deep, multi-faceted learning of diagnosing mental disorders from clinical interview records   

The key characteristics of mental illnesses are reflected in audio recordings of clinical interviews with patients and their families. We have developed a deep learning method that automatically extracts the relevant features necessary for the diagnosis of mental illnesses (ADHD, depression, bipolar disorder and schizophrenia) from such interviews. We use a variety of pre-trained models to extract representations from both the audio segments of these interviews and their text versions. We use several modern representation techniques (embeddings). We apply a Big Data approach by exploring existing audio and text corpora annotated with emotional labels. We address the problem of annotated data scarcity by using parametric model fine-tuning (Parameter Efficient Fine-Tuning). All these representations are then combined into a single multimodal form. To diagnose the above mental disorders, we use contrastive learning and model synthesis using a committee of experts (Mixture of Experts). The results show that through multimodal analysis of clinical interviews, mental disorders can be diagnosed with satisfactory accuracy (project conducted in collaboration with H. Naderi and R. Uher).

20 October 2025

Arkadiusz Modzelewski (University of Padua / Polish-Japanese Academy of Information Technology)

The Why and How of Disinformation: Datasets, Methods and Language Models Evaluation  

What language tools do disinformation agents employ? Can incorporating persuasion and intent knowledge enhance the ability of large language models to detect disinformation? And how effective are LLMs at identifying disinformation in Polish and English? In this talk, I will present findings from my PhD research on disinformation, persuasion, and the intent behind misleading information. I will introduce one of the largest Polish disinformation datasets, alongside a novel English dataset, both designed to capture manipulative techniques and intent of disinformation agents. Drawing on these and other resources, I will discuss how well current LLMs perform in detecting disinformation, persuasion, and intent, and highlight promising directions for improving their effectiveness in disinformation detection.

3 November 2025

Gražina Korvel (Vilnius University)

Developing Speech Corpora for Low-Resource Languages   

Developing diverse, well-annotated speech corpora is essential for training modern machine learning models. This presentation discusses the principles and methodologies involved in creating large-scale speech corpora, with a focus on the Lithuanian language as a case study. It presents the Great Lithuanian Speech Corpus (LIEPA-3) project, outlining strategies for collecting, annotating, and ensuring the quality of data, as well as ensuring balanced representation across dialects, genders, and age groups. The talk also addresses challenges related to ethical data collection and corpus standardization.

24 November 2025

Jan Eliasz, Mikołaj Langner, Jan Kocoń (Wrocław University of Science and Technology)

Language, Culture, and Ideology: Personalizing Offensiveness Detection in Political Tweets with Reasoning LLMs  

We investigate two complementary strategies for improving the reliability of Large Language Models in classification settings. First, we show that decomposing multi-label classification into a set of independent binary decisions offers clear practical advantages over structured output formulations: it substantially reduces parsing errors, works seamlessly with decoder-only architectures, and delivers faster inference when combined with prefix caching, without requiring any model retraining.

Divide, Cache, Conquer. Dichotomic Prompting for Efficient Multi-Label LLM-Based Classfication  

Second, we demonstrate that reasoning-enabled LLMs are markedly better at tasks requiring contextual sensitivity, such as offensive-language annotation. When prompted to adopt a specific role, reasoning models maintain that role more consistently and make more accurate, fine-grained judgments than their non-reasoning counterparts. Viewed together, these findings highlight a unifying principle: LLMs become both more efficient and more context-aware when their decision process is made more structured, whether through task decomposition or through explicit reasoning.

1 December 2025

Filip Kucia, Anna Wróblewska, Bartosz Grabek, Szymon Trochimiak (Warsaw University of Technology)

How to Make Museums More Interactive? Case Study of the “Artistic Chatbot”  

This presentation examines the challenges of deploying large language model (LLM)-powered chatbots in public cultural spaces, based on our experience with Artistic Chatbot – a voice-based conversational agent used during a month-long art exhibition at the Warsaw Academy of Fine Arts. We focus on two intertwined issues: how to make a system answer questions about a multilingual artistic collection, and how to evaluate the quality of those answers. On the technical side, we discuss strategies for building a retrieval-augmented knowledge base from heterogeneous, multilingual exhibition materials and the trade-offs between native-language models and pivot-language approaches based on translation. From the perspective of interaction design, we outline a fully voice-based setup in a gallery space, in which visitors walk up to a ceiling-mounted microphone and address the system through spoken trigger expressions, without screens or keyboards. The core of the talk is a post-hoc evaluation. We analyse interaction logs and conduct a human annotation study to compare different modelling and retrieval configurations along dimensions such as factual precision, coherence and relevance to the exhibition domain. Using this case study, we ask how to define and measure a “good” answer in conversational AI for cultural heritage, and how choices about language, translation and voice interaction should influence future deployments in museums and galleries.

Please see also the talks given in 2000–2015 and 2015–2025.