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 October 2024

Janusz S. Bień (University of Warsaw, profesor emeritus)

Identifying glyphs in some 16th century fonts: a case study  

Some glyphs from 16th century fonts, described in the monumental work “Polonia Typographica Saeculi Sedecimi”, can be more or less easily identified with the Unicode standard characters. Some glyphs don't have Unicode codepoints, but can be printed with an appropriate OpenType/TrueType fonts using typographic features. For some of them their encoding remains an open question. Some examples will be discussed.

14 October 2024

Alexander Rosen (Charles University in Prague)

Lexical and syntactic variability of languages and text genres. A corpus-based study  

This study examines metrics of syntactic complexity (SC) and lexical diversity (LD) as tools for analyzing linguistic variation within and across languages. Using quantifiable measures based on cross-linguistically consistent (morpho)syntactic annotation (Universal Dependencies), the research utilizes parallel texts from a large multilingual corpus (InterCorp). Six SC and two LD metrics – covering the length and embedding levels of nominal and clausal constituents, mean dependency distance (MDD), and sentence length – are applied as metadata for sentences and texts.

The presentation will address how these metrics can be visualized and incorporated into corpus queries, how they reflect structural differences across languages and text types, and whether SC and LD vary more across languages or text types. It will also consider the impact of language-specific annotation nuances and correlations among the measures. The analysis includes comparative examples from Polish, Czech, and other languages.

Preliminary findings indicate higher SC in non-fiction compared to fiction across languages, with nominal and clausal metrics being dominant factors. The results suggest distinct patterns for MDD and sentence length, highlighting the impact of structural differences (e.g., analytic vs. synthetic morphology, dominant word-order patterns) and the influence of source text type and style.

28 October 2024

Rafał Jaworski (Adam Mickiewicz University in Poznań)

Framework for aligning and storing of multilingual word embeddings for the needs of translation probability computation  

The presentation will cover my research in the field of natural language processing for computer-aided translation. In particular, I will present the Inter-language Vector Space algorithm set for aligning sentences at the word and phrase level using multilingual word embeddings.

The first function of the set is used to generate vector representations of words. They are generated using an auto-encoder neural network based on text data – a text corpus. In this way vector dictionaries for individual languages are created. The vector representations of words in these dictionaries constitute vector spaces that differ between languages.

To solve this problem and obtain vector representations of words that are comparable between languages, the second function of the Inter-language Vector Space set is used. It is used to align vector spaces between languages using transformation matrices calculated using the singular value decomposition method. This matrix is calculated based on homonyms, i.e. words written identically in the language of space X and Y. Additionally, a bilingual dictionary is used to improve the results. The transformation matrix calculated in this way allows for adjusting space X in such a way that it overlaps space Y to the maximum possible extent.

The last function of the set is responsible for creating a multilingual vector space. The vector space for the English language is first added to this space in its entirety and without modification. Then, for each other vector space, the transformation matrix of this space to the English space is first calculated. The vectors of the new space are multiplied by this matrix and thus become comparable to the vectors representing English words.

The Inter-language Vector Space algorithm set is used in translation support systems, for example in the author's algorithm for automatic transfer of untranslated tags from the source sentence to the target one.

4 November 2024

Jakub Kozakoszczak (Deutsche Telekom)

ZIML: A Markup Language for Regex-Friendly Linguistic Annotation  

Attempts at building regex patterns that match information annotated in the text with embedded markup lead to prohibitively unmanageable patterns. Regex and markup combine even worse when the pattern must use distances as a matching condition because tags disrupt the text format. On the other hand, fully externalized markup preserves text format but leaves regex patterns without reference points.

I introduce the Zero Insertion Markup Language (ZIML), where every combination of characters and labels in the annotated text is represented by a unique "allocharacter". Regex patterns also translate to appropriate patterns with allocharacters, preserving text span matches in standard regex engines. As the main result, ZIML extends regex semantics to include label referencing by matching allocharacters that represent them.

I will give a proof of correctness for ZIML translation and demonstrate its implementation, including a user-facing pattern language that integrates labels into regex syntax. I hope to discuss potential applications of ZIML in linguistics and natural language processing. A basic understanding of model theory and regex functionality is recommended.

21 November 2024

Christian Chiarcos (University of Augsburg)

Aspects of Knowledge Representation for Discourse Relation Annotation  

Semantic technologies comprise a broad set of standards and technologies including aspects of knowledge representation, information management and computational inference. In this lecture, I will describe the application of knowledge representation standards to the realm of computational discourse, and especially, the annotation of discourse relations. In particular, this includes the formal modelling of discourse relations of different theoretical frameworks by means of modular, interlinked ontologies, the machine-readable edition of discourse marker inventories with OntoLex and techniques for the induction of discourse marker inventories.

2 December 2024

Participants of PolEval 2024

Presentation of the Shared Task results  

Welcome to PolEval 2024 (Łukasz Kobyliński, Maciej Ogrodniczuk, Filip Graliński, Ryszard Staruch, Karol Saputa)

PolEval 2024 Task 1: Reading Comprehension (Ryszard Tuora / Aleksandra Zwierzchowska)

Optimizing LLMs for Polish Reading Comprehension: A Comparative Study of Ensemble and Unified Approaches (Krzysztof Wróbel)

PolEval 2024 Task 2: Emotion and Sentiment Recognition (Jan Kocoń, Bartłomiej Koptyra)

Emotion and Sentiment Recognition in Polish Texts Using Large Language Models: A Winning Approach to PolEval 2024 (Krzysztof Wróbel)

Ensemble as a Variance Reduction Method for Emotion and Sentiment Recognition (Tomasz Warzecha)

Emotion and Sentiment Recognition Using Ensemble Models (Jakub Kosterna)

Zero-shot Approach Using Bielik LLM for Emotion Recognition in Polish (Paweł Cyrta)

PolEval 2024 Task 3: Polish Automatic Speech Recognition Challenge (Michał Junczyk, Iwona Christop, Piotr Pęzik)

Augmenting Polish Automatic Speech Recognition System with Synthetic Data (Łukasz Bondaruk, Jakub Kubiak, Mateusz Czyżnikiewicz)

Exploration of training Zipformer and E-Branchformer models with Polish language BIGOS dataset (Paweł Cyrta)

19 December 2024

Piotr Przybyła (Pompeu Fabra University / Institute of Computer Science, Polish Academy of Sciences)

Adaptive Attacks on Misinformation Detection Using Reinforcement Learning  

The presentation will cover XARELLO: a generator of adversarial examples for testing the robustness of text classifiers based on reinforcement learning. This solution is adaptive: it learns from previous successes and failures in order to better adjust to the vulnerabilities of the attacked model. It reflects the behaviour of a persistent and experienced attacker, which are common in the misinformation-spreading environment. We will cover the evaluation of the approach using several victim classifiers and credibility-assessment tasks, showing it generates better-quality examples with less queries, and is especially effective against the modern LLMs.

17 February 2025

Alicja Martinek (NASK National Research Institute, AGH University of Kraków), Ewelina Bartuzi-Trokielewicz (NASK National Research Institute, Warsaw University of Technology)

Detecting deepfakes and false ads through analysis of text and social engineering techniques  

Existing deepfake detection algorithm frequently fail to successfully identify fabricated materials. These algorithms primarily focus on technical analysis of video and audio, often neglecting the meaning of content itself. In this paper, we introduce a novel approach that emphasizes the analysis of text-based transcripts, particularly those from AI-generated deepfake advertisements, placing the text content at the center of attention. Our method combines linguistic features, evaluation of grammatical mistakes, and the identification of social engineering techniques commonly used in fraudulent content. By examining stylistic inconsistencies and manipulative language patterns, we enhance the accuracy of distinguishing between real and deepfake materials. To ensure interpretability, we employed classical machine learning models, allowing us to provide explainable insights into decision-making processes. Additionally, zero-shot evaluations were conducted using three large language model based solutions to assess their performance in detecting deepfake content. The experimental results show that these factors yield a 90\% accuracy in distinguishing between deepfake-based fraudulent advertisements and real ones. This demonstrates the effectiveness of incorporating content-based analysis into deepfake detection, offering a complementary layer to existing audio-visual techniques.

24 March 2025

Maciej Rapacz, Aleksander Smywiński-Pohl (AGH University of Krakow)

Interlinear Translation of Ancient Greek Texts: How Morphological Tags Enhance Machine Translation Quality   

Interlinear translation prioritizes preserving the original syntactic structure by placing target language words directly below their source text counterparts, maintaining the original word order rather than natural fluency. Although interlinear translations often deviate from the linguistic norms of the target language, they serve as a valuable tool for those wishing to deeply understand texts in their original form, especially in the case of sacred and ancient texts.

In our research, we conducted the first attempt to apply machine translation to generate interlinear translations from Ancient Greek to Polish and English. We compared the performance of specialized models (GreTa, PhilTa) pretrained on Ancient Greek texts with a general-purpose multilingual model (mT5). We examined 144 different model configurations, manipulating the base model, morphological tag encoding method, tag set, and text normalization approach, using the Greek New Testament texts as our corpus.

During the presentation, we will describe our research methodology and discuss the results. The best results were achieved by models in which we implemented new dedicated embedding layers for encoding morphological information, which yielded results up to 35-38% better (BLEU) compared to the baseline scenario. Additional detailed study showed that PhilTa performs better than mT5, particularly in scenarios with limited data availability. PhilTa achieved the highest results in translation to English (60.40 BLEU), while mT5-large performed best with Polish (59.33 BLEU).

14 April 2025

Ryszard Staruch, Filip Graliński (Adam Mickiewicz University in Poznań)

Leveraging Large Language Models for the Grammatical Error Correction Task  

Large Language Models (LLMs) currently represent the state-of-the-art in many natural language processing tasks. However, their effectiveness in correcting language errors in texts written in Polish remains unclear. To address this gap, a dedicated dataset for Polish text correction has been developed. During the talk, this dataset will be presented along with the evaluation results of selected LLM-based solutions. In the second part of the seminar, new techniques for adapting LLMs to the task of minimal-edit text correction will be discussed, focusing on texts written by language learners — using English as a case study.

28 April 2025

Manfred Stede (Universität Potsdam)

Discourse structure in the Potsdam Commentary Corpus: Human annotation, human disagreement, and automatic parsing  

The talk gives a brief introduction to Rhetorical Structure Theory (RST, Mann/Thompson 1988) and then explains the design decisions for the Potsdam Commentary Corpus (PCC), which brings together RST, coreference, and other annotation layers on 175 German news editorials. After illustrating cross-layer queries on the corpus in the ANNIS linguistic database, we turn to the intricacies of manual RST annotation. I will give an overview of the annotation guidelines and their motivations, and present results from an (ongoing) study on annotator disagreements, from which we derive ideas for redesigning the annotation scheme (and potentially the underlying theory), with a comparison to the recent proposal of "eRST" by Zeldes et al. (2025). In the last part of the talk, I outline our results on automatic parsing using the system by Ji and Eisenstein (2014).

26 May 2025

Deniz Zeyrek (Middle East Technical University)

Building monolingual and multilingual discourse banks and implications for discourse structure  

In this talk, I will overview the Turkish Discourse Bank (TDB), and the TED-MDB (TED Multilingual Discourse Bank), both annotated at the discourse level by native speakers. The TDB is a resource of over 3800 implicitly or explicitly conveyed discourse relations built over a multi-genre corpus of 40.000 words. The TED-MDB is a multilingual corpus of six English TED talks with translations into five languages (Turkish, Polish, European Portuguese, Russian, and German, recently extended to a sixth language, Lithuanian) with about 600 relation annotations per language. While both corpora follow the rules and principles of the Penn Discourse Treebank (PDTB), they also consider the language-specific characteristics of individual languages. I will summarize the characteristics of both corpora and the work of our research team where these corpora are exploited, discussing implications on discourse structure.

2 June 2025

Maciej Ogrodniczuk, Aleksandra Tomaszewska, Bartosz Żuk, Alina Wróblewska (Institute of Computer Science, Polish Academy of Sciences)

The title of the talk (on the Polish Large Language Model) will be given shortly  

The summary of the talk will be given shortly.

23 June 2025

Aleksandra Tomaszewska, Bartosz Żuk, Dariusz Czerski, Maciej Ogrodniczuk (Institute of Computer Science, Polish Academy of Sciences)

The title of the talk (on the NeoN tool for detecting lexical innovations) will be given shortly  

The summary of the talk will be given shortly.

Please see also the talks given in 2000–2015 and 2015–2024.