In my research, I focus on making sure that large language models (LLMs) generate accurate text.

The projects I have worked on include:

• domain-independent and low-resource data-to-text generation [1], [2], [3],
• evaluating semantic accuracy of generated text [4], [5], [8], [9], [10],
• investigating the role of data labels [6],
• a software toolkit for data-to-text generation [7].

During my internship at Mila–Quebec AI institute, I was also working on applying LLMs for autonomous web navigation [11].

If I only had enough time, I would dig into many more things! For example, I am interested in how is the information inside the language models represented, how this all relates to human cognition, and how can we use AI tools to make our lives better.

Selected publications

Large Language Models as Span Annotators

Zdeněk Kasner, Vilém Zouhar, Patrícia Schmidtová, Ivan Kartáč, Kristýna Onderková, Ondřej Plátek, Dimitra Gkatzia, Saad Mahamood, Ondřej Dušek, Simone Balloccu

2025, arXiv

Finding errors in the text by highlighting the part that relates to the error sounds like a great idea... until you have to start paying human annotators for doing it. So the question is: can we use LLMs to do that for us? We try it on three tasks: evaluation of data-to-text generation, finding errors in machine translation and propaganda techniques annotation. We find that for simpler, well-defined tasks, LLMs are catching up with skilled crowdworkers. The results are more mixed for complex tasks, but we still see LLMs for span annotation as the path forward!

arXiv Website Code

factgenie: A Framework for Span-based Evaluation of Generated Texts

Zdeněk Kasner, Ondřej Plátek, Patrícia Schmidtová, Simone Balloccu, Ondřej Dušek

INLG 2024 – Demo Track

Have you ever found youself staring at a meaningless wall of results from corpus-level metrics? Not anymore! Our factgenie framework promotes a more fine-grained and telling way of evaluating generated texts. Any error is highlighted on the level of individual text spans. How can you do that? Either the "old-fashioned way" using human annotators, or using LLM-as-evaluators – it's up to you to choose! All conveniently packaged in a lightweight, self-hosted, Flask-based web app.

ACL Anthology arXiv Code

Data-to-Text Generation with Neural Language Models

Zdeněk Kasner

PhD Thesis

My PhD Thesis I submitted at Charles University in June 2024 (on the day of my 30th birthday!). If want to know more about the interplay of neural language models and data-to-text generation, this tome will guide you through. It also contains a variety of papers of mine in a condensed form, with conclusions updated so that they reflect the current state of the art.

PDF

Beyond Reference-Based Metrics: Analyzing Behaviors of Open LLMs on Data-to-Text Generation

Zdeněk Kasner, Ondřej Dušek

ACL 2024

Bechmarks for data-to-text generation are, well... not great. Most of them are small-scale, subject to data contamination, and slowly becoming saturated. But there are plenty of novel structured data online! In this paper, we make use of structured data scraped from open APIs and evaluate open LLMs (Llama2, Mistral, and Zephyr) on five data-to-text generation tasks, such as generating weather forecasts or product descriptions. Using both GPT-4 and human annotators, we find that the models outputs contain plenty of semantic errors on token-level: on average, more than 80% of outputs contain at least one error.

ACL Anthology arXiv Poster Website Code Demo

WebLINX: Real-World Website Navigation with Multi-Turn Dialogue

ICML 2024 Spotlight paper (3.5% acceptance rate)

Xing Han Lù, Zdeněk Kasner, Siva Reddy

ICML 2024

Automating web navigation – that was the project I was working on during my internship at Mila. We started with the idea that fully autonomous agents controlling a web browser are unsafe and impractical. So instead, we built a dataset for train and evaluating conversational web navigation: giving the model instructions via chat. On our dataset, we recorded demonstrations with expert annotators and evaluated quite a lot of models of various kinds. Since multi-modal models are not pretrained on structured inputs (...such as websites), text-only LLMs turned out to be suprisingly good for the task.

arXiv Poster Website Code

TabGenie: A Toolkit for Table-to-Text Generation

Zdeněk Kasner, Ekaterina Garanina, Ondřej Plátek, Ondřej Dušek

ACL 2023 – Demo Track

We began developing this tool to play with the generative language models in real time but it soon evolved into a swiss knife for table-to-text generation. TabGenie provides interactive data visualization, unified data representation & unified programming interface for more than 15 data-to-text datasets. You can use the web interface as a dataset viewer and a model playground, the programming interface then allows to quickly prototype new models.

ACL Anthology arXiv Poster Code Demo

Mind the Labels: Describing Relations in Knowledge Graphs With Pretrained Models

Zdeněk Kasner, Ioannis Konstas, Ondřej Dušek

EACL 2023

What is a better way to learn the data semantics – memorizing an arbitrary mapping or taking the human-readable data labels into account? On the task of describing a triple (entity_1, relation, entity_2), we show that the models are able to describe previously unseen relations as long as the relation label is meaningful and unambiguous. To put it another way: if you want to train robust data-to-text systems, don't use abbreviations!

ACL Anthology arXiv Poster Code

Neural Pipeline for Zero-Shot Data-to-Text Generation

Zdeněk Kasner, Ondřej Dušek

ACL 2022

To combine the power of pretrained language models with controllability of pipeline approaches, we formulate data-to-text generation as a sequence of trainable text-to-text operations: ordering, aggregation, and paragraph compression. As a welcome side-effect, we get rid of semantically incorrect outputs arising from noisy human-written references. Is NL-only approach to data-to-text generation the way to go?

ACL Anthology arXiv Poster Code

Text-in-Context: Token-Level Error Detection for Table-to-Text Generation

Best submission at Shared Task in Evaluating Accuracy

Zdeněk Kasner, Simon Mille, Ondřej Dušek

INLG 2021

How to automatically detect which parts of the generated text do not correspond to the data? In our submission for the Shared Task in Evaluating Accuracy 2021, we devise a 3-step approach combining a rule-based system with pretrained language models. Our approach was the best out of four submitted metrics!

ACL Anthology Poster Code

Train Hard, Finetune Easy: Multilingual Denoising for RDF-to-Text Generation

2nd place in Russian RDF-to-text generation

Zdeněk Kasner, Ondřej Dušek

INLG 2020

Data == noisy text. That's definitely an overgeneralization, oversimplification... but um, it works! We succesfully generate text from DBPedia data in English and Russian just by finetuning mBART – a pretrained multilingual denoising autoencoder. This is our submission for the WebNLG Challenge 2020, presented at the 3rd Workshop on Natural Language Generation from the Semantic Web.

ACL Anthology

Evaluating Semantic Accuracy of Data-to-Text Generation with Natural Language Inference

INLG 2020 Best Short Paper Award

Ondřej Dušek, Zdeněk Kasner

INLG 2020

Does a text contain all the information from the data? Hard to check manually, even harder to code. But what if we reformulate the question a little bit: can we infer all the data from the text and nothing else? That sounds like natural language inference. And guess what - somebody already pretrained a neural model for that!

ACL Anthology arXiv Code

Data-to-Text Generation with Iterative Text Editing

Zdeněk Kasner, Ondřej Dušek

INLG 2020

Imagine your task is to generate text from data. Which sounds easier: generating text from scratch or joining existing sentences? We propose an approach in which we iteratively join the sentences with a text-editing neural model. Since the model has a limited vocabulary, it has also a limited possibilities of introducing incorrect facts. Moreover, it also turns out that sentence fusion is a quite general task which works on multiple domains.

ACL Anthology arXiv Poster Code

Improving Fluency of Non-Autoregressive Machine Translation

Zdeněk Kasner, Jindřich Libovický, Jindřich Helcl

2020, arXiv

In the follow-up of my master thesis, we improve translation quality of a CTC-based machine translation model. The model is non-autoregressive, i.e. faster but lacking behind autoregressive models in translation quality. To improve the translation quality, we re-score the hypotheses during the beam search decoding with an n-gram language model and several other features.

arXiv Code