The NLP Seminar is organised by the Linguistic Engineering Group at the Institute of Computer Science, Polish Academy of Sciences (ICS PAS). It takes 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.

7 September 2026

Varvara Magomedova (University of Nova Gorica)

Veje – a future treebank of Slovenian dialectal texts  

Slovenian dialect research has produced a lot of literature, but not as many data sources. The lack of available syntactically annotated corpora limits research for those, who do not have access to native speakers. This talk presents the first phase of Veje – to become a Slovenian dialect treebank. The present version of the pipeline is designed to reduce the manual work. The pipeline has two modules – normalization and annotation. In the first stage, dialectal surface forms are normalized towards standard Slovenian through a sequence of corrected lexical replacements, deterministic phonological rules, Sloleks-based morphological validation, and a constrained GaMS residual corrector. The process preserves token order and alignment wherever possible and records substitutions that cannot be morphologically verified for later review. In the second stage, normalized sentences are parsed with the non-standard Slovenian CLASSLA-Stanza model, while SloBERTa identifies whatever could not be normalized and sends it to GaMS for a second check of the annotation. The resulting extended CoNLL-U representation retains the original dialect form alongside its normalized equivalent, Universal Dependencies annotation, dialect metadata, and audit information. The current output is intentionally silver-standard: native speakers and dialectologists will correct selected data into a gold subset, which will subsequently support evaluation and model fine-tuning. I will also show the web-platform developed based on user needs study held with corpora users.

21 September 2026

Mirosław Koziarski (independent researcher)

The Digital Edition of the “Varsovian Dictionary”  

The Dictionary of the Polish Language, edited by Jan Karłowicz, Adam Kryński, and Władysław Niedźwiedzki and commonly known as the Varsovian Dictionary, is one of the most important and extensive works of Polish lexicography. Its eight volumes, published between 1900 and 1927 and comprising almost 7,800 pages, document Polish from a wide range of periods, regions, registers, and domains. Although scans of the dictionary are available in digital libraries, its content has largely remained locked within images of printed pages.

In this presentation, I will introduce the project of a digital edition of the Varsovian Dictionary, which develops the methodology and tools designed as part of my doctoral dissertation. I will discuss the successive stages of processing: source selection and analysis, image processing and segmentation, OCR—now also supported by AI-based tools—text correction and normalisation, hierarchical parsing of dictionary entries, validation, indexing, and the generation of derived data. The core component is a parser that transforms a linear transcription into a structured XML representation. It identifies several dozen types of segments, including headwords, variants, grammatical information, senses and subsenses, definitions, qualifiers, usage examples, etymologies, cross-references, and phraseological units, as well as the relations between them. Given the complexity and inconsistency of the source material, the process is iterative and combines automatic methods with manual verification.

The online edition extends an earlier research prototype into a gradually expanding platform. Users can compare a formatted entry with its source transcription, raw XML, and the corresponding facsimile. Structuring the content also makes it possible to create new paths of access to the data: a corpus of examples, collections of derivative pairs, a consolidated index of abbreviations, quantitative statistics, and a matrix comparing the dictionary’s headword inventory with those of other Polish dictionaries.

Finally, I will present the current state of the project, its quality-control mechanisms, and the limitations of automatic processing of historical lexicographic material. I will also outline planned developments.

19 October 2026

Justyna Gromada, Natalia Krawczyk (Orange Research)

Evaluation of Conversational Agents and Interpretable Satisfaction Modeling in Sales Dialogues   

The first part of the talk will introduce the CURIE research program, within Orange Research in Poland. We will focus on the evaluation of conversational agents, particularly in sales dialogues. The second part will present an approach to modeling user satisfaction in sales dialogues. Using 5200 simulated phone-sales conversations between the bot and the user, we developed an annotation framework for offers, cross-selling attempts, and customer decisions. We then derived metrics capturing aspects such as sales outcomes, negotiation patterns, and final cart contents. A model based on these metrics achieved results close to those of a model analyzing the dialogue text, while also helping to identify what influences customer satisfaction.

22 October 2026

Nina Smirnova (GESIS – Leibniz Institute for the Social Sciences)

Open Political Corpora: Structuring, Searching, and Analyzing Parliamentary Transcripts  

This seminar presents our work on processing and providing access to parliamentary transcripts as part of a specialized information service for political science. Our goal is to create a unified, searchable collection of parliamentary speeches from different countries and languages, with a particular focus on making these resources accessible to researchers without extensive programming expertise.

At the center of this work is PoliCorp, a web-based platform for searching and exploring processed political texts. In contrast to initiatives such as Parla-CLARIN and related parliamentary corpora, PoliCorp emphasizes intuitive, browser-based access to the data. Its advanced search functionality allows researchers to combine or exclude criteria using logical operators and to search across multiple metadata and text fields. This makes it possible to construct targeted subcorpora and investigate complex patterns within large collections of parliamentary debates. Selected data can subsequently be downloaded in JSON format for further computational analysis.

The current collection includes transcripts from the German Bundestag and German state parliaments. We are extending the corpus to include parliamentary debates from France and Poland, moving toward a multilingual and cross-national research infrastructure.

The seminar will introduce the main stages involved in building such a resource, from preprocessing and structuring parliamentary transcripts to storing, indexing, and making them searchable on the web. We will discuss both the technical and methodological challenges of harmonizing heterogeneous parliamentary data across institutions, countries, and languages.

Finally, we will demonstrate how these corpora can support substantive research in political science and computational social science.

9 November 2026

Piotr Pęzik, Filip Żarnecki, Wojciech Janowski, Jakub Kwiatkowski, Paweł Wilk, Łukasz Stolarski (University of Łódź)

LLMs in corpus search. From orchestration to exploration  

Talk summary will be made available shortly.

19 November 2026

Sabina Tomkins (University of Michigan)

The title of the talk will be made available soon  

Talk summary will be made available shortly.

23 November 2026

Persuasion in Parliament and Beyond: Computational Analysis of Political Discourse

Mini-conference supported by the Polish Academy od Sciences (MOST PAN)  

ParlaMint – Comparable and Interoperable Parliamentary Corpora (Tomaž Erjavec, Jožef Stefan Institute)

The talk presents the ParlaMint corpora of parliamentary debates of 29 European countries and autonomous regions, covering at least the period from 2015 to 2022, and containing over 1 billion words. The corpora are uniformly encoded, contain rich metadata about their 24 thousand speakers, and are linguistically annotated, as well as machine translated to English. We present the compilation of the corpora, including the encoding infrastructure, use of GitHub, the production of individual corpora, the common pipeline for producing their distribution, and use of CLARIN services for dissemination. We also present associated efforts, such corpora of additional countries, and the ParlaSpeech corpora. Finally, the use of the corpora and further work are discussed.

Beyond the Polish Parliamentary Corpus (Maciej Ogrodniczuk, Institute of Computer Science, Polish Academy of Sciences)

The Polish Parliamentary Corpus, which is also represented in ParlaMint, served as a model for a series of Polish corpora containing similar types of material, which are ideally suited to research into persuasion techniques. The presentation will introduce these datasets: the Round Table Corpus, documenting the opposition’s negotiations with the communist authorities of the Polish People’s Republic in 1989, and the Polish Local Government Debates Corpus – a collection of transcripts from the proceedings of provincial assemblies, city councils, county councils and local councils from 2018 to 2026.

Multilingual Corpora of Persuasion Techniques in News and Parliamentary Debates (Jakub Piskorski, Joint Research Centre of the European Commission)

This talk provides an overview of recently developed multilingual corpora with text span-level annotation of persuasion techniques in online news, parliamentary debates, and other types of texts. The talk will elaborate on the corpora characteristics, taxonomies, creation process, and annotation complexity aspects. The talk will also reflect on the experience gained from shared tasks revolving around detection and classification of persuasion techniques which exploit these corpora, and covers state-of-the-art performance and cross-corpora comparison.

Leveraging Persuasion and Intent for Analysis and Reasoning-based Detection of Disinformation with Large Language Models (Arkadiusz Modzelewski, Uniwersytet Padewski / NASK)

This talk introduces novel human-annotated disinformation datasets that enable the analysis of manipulation techniques and malicious intent in Polish and English. It presents reasoning approaches that explicitly incorporate persuasion and intent to improve LLM-based disinformation detection across domains, genres, and languages. Finally, it introduces a multilingual benchmark of human- and AI-generated persuasive content, showing that AI-generated persuasive text exhibits linguistic differences and poses new challenges for detection. Overall, the talk presents new datasets, reasoning frameworks, and empirical insights toward more transparent and generalizable disinformation detection in the era of generative AI.

Talk title will be made available soon (Philipp Mayr-Schlegel, GESIS – Leibniz Institute for the Social Sciences)

Talk summary will be made available soon.

ParlSent-EU: Three Decades of Interpersonal Sentiment in European Parliamentary Debates Since the 1990s (Kamil Orzechowski, Warsaw University of Technology)

Parliamentary corpora offer rich speech texts and speaker metadata but rarely capture interpersonal sentiment between legislators. Research on legislative relationships relies mainly on roll-call votes and co-sponsorship, reflecting behavioural alignment rather than expressed attitudes such as approval, criticism, or hostility. We introduce ParlSent-EU, a cross-national dataset of directed interpersonal sentiment relations from plenary speeches in the Polish Sejm, German Bundestag, and French Assemblée nationale, spanning over three decades from the 1990s. It was developed by Warsaw University of Technology (OMINO project) and GESIS – Leibniz Institute for the Social Sciences (POLLUX project). Each relation records a speaker, a targeted legislator, and the expressed sentiment (positive, neutral, or negative), with temporal, parliamentary, and demographic metadata, enabling the reconstruction of directed, signed, and temporal networks. Relations are identified by a pipeline combining lexical filtering, LLM-based sentiment extraction, and rule-based entity disambiguation, validated against manually annotated gold standards with performance metrics and inter-annotator agreement. ParlSent-EU enables comparative, longitudinal studies of elite affective polarization, cross-party hostility, intra-party cohesion, and gendered political communication in European parliamentary democracies.

From Detecting Persuasion to Predicting Its Success: Persuasion Strategies as Predictive Signals (Tiziano Labruna, Fondazione Bruno Kessler)

Computational work on persuasion has largely focused on detecting rhetorical techniques in text. This talk would ask a complementary question: do those techniques tell us anything about whether an argument actually changes someone's mind? Multi-Strategy Persuasion Scoring is a zero-shot framework in which a large language model reasons about each of six persuasion strategies independently and produces a per-strategy score, which can then be aggregated directly or used as input to a lightweight classifier. The strategy has been tested across three datasets, proving that strategy-guided reasoning consistently outperforms both direct pairwise comparison and generic chain-of-thought baselines.

30 November 2026

AI without Borders

Mini-conference supported by the Polish Academy od Sciences (MOST PAN)  

Universalist modelling and processing of idiomaticity in the PARSEME and UniDive framework – recent developments and the state of Polish (Agata Savary, Paris-Saclay University)

Multiword expressions (MWEs), like "biały kruk" (lit. a white crow, 'a rare person or thing') or "krew kogoś zalewa" (lit. blood floods someone, 'someone get furious'), a "być komuś na rękę (lit. to be on hand to someone, 'to be convenient'), or "wchodzić w grę" (lit. to enter the play, 'to come into play'), are combinations of words which exhibit idiosyncratic behavior on the lexical, morphological, syntactic, and semantic levels. Their most outstanding feature is their semantic non-compositionality, i.e. the fact that their meaning cannot be straightforwardly deduced form the meanings of their components.

MWEs pose severe challenges in semantically-oriented NLP tasks. For instance paraphrasing sentences containing idioms is hindered by their lexical and morpho-syntactic inflexibility. Thus, when applying usual paraphrasing techniques to "krew kogoś zalewa", we often obtain incorrect paraphrases, in which the idiomatic meaning is lost, e.g. "ktoś jest zalany krwią" (someone is flooded with blood) or "krew kogoś oblewa" (lit. blood poured onto someone). One of the ways to tackle this challenge is to identify MWEs in running text and apply dedicated treatment to them.

MWE identification has been the object of many efforts, and is one of the tasks where supervised encoder-based methods still outperform more recent generative LLMs. Paraphrasing of MWEs is an understudied task, especially in multilingual contexts. These two tasks have been the subject of longstanding effort within two European COST networks: PARSEME (2013–2017) and UniDive (2022–2026), which are notably the organizers of 5 multilingual evaluation campaigns on these topics. I will summarize major challenges and findings from these campaigns and highlight future work.

Detecting AI generated content on surface- and idea-level (Marzena Karpińska, Simon Fraser University)

Modern detectors identify only who wrote the words. However, AI-assisted writing is becoming more common. For instance, a user may use AI tools to polish their writing or produce the text from their supplied content. Hence, it is becoming important to detect who contributed the ideas not only the prose. In this talk, I will first discuss the specificities of AI writing at the prose and content levels. Then I will talk about training Idealens, an AI detector that works based on document content, rather than its prose.

Can Neural Network Models Learn Logic? (Jakub Szymanik, University of Trento)

Large language models now produce chains of reasoning that look like proofs, but it remains unclear whether they have acquired the rules of inference or merely the patterns that usually accompany them. I will describe a line of joint work with Manuel Vargas Guzmán and Maciej Malicki that tries to make this question precise enough to answer. We use syllogistic logic — a small, fully controlled fragment of natural language — as a testbed, and ask models not whether a conclusion holds but which premises are needed to derive it. This lets us separate two things usually conflated under "compositional generalization": extending a learned pattern to longer chains of reasoning, and recovering the underlying rules from complex cases to apply them to simpler ones. Models turn out to be markedly better at the first than the second, and their success depends systematically on the shape of the inference rather than its difficulty. I will end with two more encouraging results — that training on examples of the task itself substantially helps smaller models, and that even imperfect neural reasoners make effective assistants to a symbolic prover — and with what all this suggests about what "learning logic" could mean for a system that learns from data.

Faithful Explanations in the Age of Large Language Models (Mateusz Lango, Charles University in Prague)

The fluency of modern large language models may lead to a perception that natural language generation is close to solved, yet a closer look reveals a recurring pattern of failure whenever a task requires exact, discrete fidelity to some underlying structure rather than fluent pattern completion. Reasoning that is straightforward in a single turn becomes surprisingly brittle across a dialogue; generation under explicit lexical or content constraints remains difficult to guarantee; and explanations produced by LLM-as-judge systems are frequently unfaithful to the actual decision process they claim to justify. I will argue that these three failures share a common shape: each is a case where an LLM must maintain or produce something exact and discrete, but is instead relying on an implicit, continuous representation not well suited to that job. This talk focuses on the third case - faithful explanation - and asks what happens when we make the intermediate representation explicit and symbolic, rather than leaving it to the model to reconstruct implicitly. I will present two lines of work built on this idea. The first constructs faithful explanations for image classifiers via a pipeline that separates the extraction of the (symbolic, verifiable) evidence behind a decision from its realization into fluent natural language, so that plausibility and faithfulness can be optimized somewhat independently. The second uses LLMs not to generate text directly, but to author interpretable, rule-based generation systems from scratch, producing data-to-text generators that are faithful and inspectable by construction rather than by post-hoc checking. Together, these case studies suggest a general recipe for faithful explanation: let neural models handle fluency and search, and let symbolic structure carry the guarantees.

Towards Temporal Intelligence in Language Technologies: Modeling, Retrieval, and Reasoning (Adam Jatowt, University of Innsbruck)

Time is central to how we understand the world. It shapes how we interpret events, construct narratives, and reason about change, causality, and evidence. In this talk, I will first examine how well large language models handle temporal information and temporal reasoning, an essential but still less explored dimension of language intelligence. I will then discuss recent advances in incorporating temporal awareness into language technologies, including time-aware representation learning, retrieval from evolving document collections, and the development of new resources for training and evaluation. I will also introduce temporal reasoning-oriented question answering and retrieval benchmarks that highlight the challenges of retrieving time-sensitive information, answering complex temporal questions, and reasoning over changing evidence. The talk will conclude with open challenges for building language technologies that can better understand, retrieve, and reason over information across time.

In Search for Universal Language Similarity Metric with Application in Multilingual AI (Michał Ptaszyński, Kitami Institute of Technology)

How similar must two languages be for an AI model trained on one to work on the other, and can we know before training? Multilingual AI models are usually trained on a few high-resource languages and then applied to many others. This lecture presents qWALS, an open-source metric of structural distance between more than 2,600 languages, built on nearly 200 features from the World Atlas of Language Structures (WALS), including word order, morphology, phonology, negation and grammatical gender. Unlike vector-based measures, every qWALS feature can be inspected and re-weighted for a given task.

I show that cross-lingual transfer works better the more similar the source and target languages are. Once the features relevant to a task are selected, qWALS predicts transfer performance in syntactic parsing almost perfectly (up to r = −0.99). For Polish, a model trained on other Slavic languages clearly outperforms one trained on English. To additionally valuidate qWALS, three professional linguists judged 769 language-pair similarity questions, and their answers agreed with qWALS at least as closely as with each other.

Finally, I discuss what this means for multilingual models and LLMs, namely, choosing source languages for transfer, estimating performance before training, and deciding which languages to include in training data. This matters most for low-resource and endangered languages, where data is scarce and the choice of source language matters most.

AI Argues Differently: Distinctive Persuasive Patterns and Communicative Strategies of LLMs (Agnieszka Faleńska, University of Stuttgart)

Large language models are becoming remarkably good at producing language that sounds convincing, respectful, and socially appropriate. But does this mean they are becoming better communicators?

In this talk, I explore this question through a series of studies examining how language models differ from human writers, construct persuasive arguments, and adapt to their conversational partners. We will see that AI-generated language exhibits distinctive linguistic patterns, while often appearing more agreeable and persuasive than human writing. Beyond the quality of their arguments, models also reproduce subtle communicative strategies, such as expressing trust, acknowledging different perspectives, and establishing common ground. Yet the very linguistic choices that make communication seem natural and engaging can also carry unintended social meanings, signalling familiarity, hierarchy, or social expectations that may be inappropriate in a given context.

Together, these findings highlight a fundamental challenge for human–AI interaction: producing human-like language is not the same as navigating its social meaning. Understanding AI communication requires looking beyond fluency and correctness to the social functions of language and how people interpret them.

Safety Vulnerabilities in Spoken Language Models and Web Agents (Karolina Stańczak, ETH Zurich)

LLM-based AI systems are rapidly expanding beyond single-modality text generation. Spoken language models combine speech and text to enable natural voice-based interaction, while web agents now act autonomously on real websites, completing complex tasks on a user's behalf. Both introduce new attack opportunities that existing safety evaluations were not designed to capture. This presentation will explore these two topics. First, I will present JAMA (Joint Audio-text Multimodal Attack), which shows that jailbreaking spoken language models by perturbing a single modality is insufficient. Jointly optimizing adversarial perturbations across both text and audio increases jailbreak success by up to 20x over unimodal attacks, while a sequential approximation keeps this attack computationally efficient. Second, I will introduce SafeArena, the first benchmark designed to evaluate the deliberate misuse of web agents across categories such as misinformation, illegal activity, and cybercrime. Our Agent Risk Assessment framework reveals that leading agents, including GPT-4o and Claude-3.5 Sonnet, complete an alarming number of harmful tasks. I will conclude this talk by discussing why safety evaluation must keep pace with how LLM-based systems are deployed, accounting for both their expanding modalities and growing agentic capabilities.

Please see also the talks given in 2000–2015 and 2015–2026.