Pieter Delobelle

Hi. I'm a postdoctoral researcher at KU Leuven focussing on alignment and fairness of large language models, probably best know for my work on Dutch language models. I am often asked for talks on LLMs and AI and I am open to collaborations, feel free to reach out.

Lines of research


July 10, 2024 Our work on Trans-Tokenization has been accepted at COLM 2024! July 01, 2024 I am visiting the lab of prof. Alan Akbik for the summer! 🥳🇩🇪 July 01, 2024 Our paper on BPE-knockout got accepted at NAACL 2024. June 30, 2024 I got an accepted paper at ICML 2024 from my Apple  internship. June 18, 2024 I will give talks on LLMs to KBC's senior management. June 13, 2024 I will give a course on bias and fairness in NLP. June 11, 2024 I organized an introductory session on Dutch NLP. May 27, 2024 I presented our Dutch LLM: Tweety-7B-Dutch. March 13, 2024 I explained LLMs for law students. February 27, 2024 I gave a talk on translating monolingual LLMs. February 20, 2024 I will give a lecture series on NLP at KU Leuven this semester. View older news

Recent and highlighted research

Posted on July 10, 2024


Language Adaptation of LLMs for Low-Resource NLP

we present a novel cross-lingual vocabulary transfer strategy, trans-tokenization, designed to adapt a high-resource monolingual LLM to a new target language by initializing the token embeddings of the target language using a weighted average of semantically similar token embeddings from the source language. For this, we leverage a translation resource covering both the source and target languages. We validate our method with the Tweeties, a series of trans-tokenized LLMs, and demonstrate their competitive performance on various downstream tasks across a small but diverse set of languages.

Posted on June 10, 2024


Pruning Pre-existing BPE Tokenisers with Backwards-compatible Morphological Semi-supervision

Byte-pair encoding (BPE) has become the default subword tokeniser in language models (LMs), allowing the representation of an infinite space of text with a finite set of units. Yet, BPE training is unsupervised, receiving no explicit information about a language's morphology. This results in a subword vocabulary wherein many units are a concatenation of partial morphemes, preventing their formation as tokens.

Posted on May 13, 2024


A Dutch generative LLM

Most Dutch generative language models start from an English or multilingual model and finetune that, which works well but is not optimal as the tokens are mostly English. We present Tweety-7b-dutch, a Dutch generative language model that is <i>trans-tokenized</i> to use Dutch tokens instead of English. To highlight the benefits of our method, we show that this model outperforms the multilingual and state-of-the-art Dutch generative language models.

Posted on December 27, 2023

Dutch Chat Toolkit

Creating retrieval-augmented chatbots

A lot of NLP technologies are easy to use for beginners, but creating and deploying a chatbot is still a bit tricky. Let's make a Python CLI toolkit to quickly create a chatbot with a web-based user interface.

Posted on November 29, 2023


Updated large and base Dutch BERT models

With RobBERT-2023, we deliver a freshly pre-trained Dutch tokenizer using the latest version of the Dutch OSCAR corpus. This corpus incorporates new high-frequency terms, such as those related to the COVID-19 pandemic, cryptocurrencies, and the ongoing energy crisis, while mitigating the inclusion of previously over-represented terms from adult-oriented content. Unlike the prior versions of RobBERT, which relied on the training methodology of RoBERTa but required a fresh weight initialization, RobBERT-2023 is entirely initialized using the RoBERTa-large model.

Posted on October 19, 2023


Translating language models one token at a time

Training monolingual language models for low and mid-resource languages is made challenging by limited and often inadequate pretraining data. We propose a novel model conversion strategy to address this issue, adapting high-resources monolingual language models to a new target language. We map tokens from the target tokenizer to semantically similar tokens from the source language tokenizer.

Posted on June 28, 2023


Improving Resume Quality Through Co-Creative Tools

Clear and well-written resumes can help jobseekers find better and better-suited jobs. However, many people struggle with writing their resumes, especially if they just entered the job market. Although many tools have been created to help write resumes, an analysis we conducted showed us that these tools focus mainly on layout and only give very limited content-related support. We present a co-creative resume building tool that provides tailored advice to jobseekers based on a comprehensive computational analysis of 444k resumes and the development of a Dutch language model, ResumeRobBERT, to provide contextual suggestions.

Posted on February 07, 2023

How far can it go?

On Intrinsic Gender Bias Mitigation for Text Classification

We have designed a probe to investigate the effects of intrinsic gender bias mitigation strategies on downstream text classification tasks. We find that instead of resolving gender bias, these strategies are able to hide it while retaining significant gender information in the embeddings. Based on these findings, we recommend that intrinsic bias mitigation techniques should be combined with other fairness interventions for downstream tasks.

Posted on November 15, 2022


Updating a Dutch Language Model to Account for Evolving Language Use

We update the RobBERT Dutch language model to include new high-frequent tokens present in the latest Dutch OSCAR corpus from 2022. We then pre-train the RobBERT model using this dataset. Our new model is a plug-in replacement for RobBERT and results in a significant performance increase for certain language tasks.

Posted on August 16, 2022


Mitigating Stereotyping in Language Models

Large pre-trained language models are successfully being used in a variety of tasks, across many languages. With this ever-increasing usage, the risk of harmful side effects also rises, for example by reproducing and reinforcing stereotypes. However, detecting and mitigating these harms is difficult to do in general and becomes computationally expensive when tackling multiple languages or when considering different biases. To address this, we present FairDistillation : a cross-lingual method based on knowledge distillation to construct smaller language models while controlling for specific biases.

Posted on December 15, 2021

Measuring Fairness with Biased Rulers

A Survey on Quantifying Biases in Pretrained Language Models

An increasing awareness of biased patterns in natural language processing resources, like BERT, has motivated many metrics to quantify 'bias' and 'fairness'. But comparing the results of different metrics and the works that evaluate with such metrics remains difficult, if not outright impossible. We survey the existing literature on fairness metrics for pretrained language models and experimentally evaluate compatibility, including both biases in language models as in their downstream tasks. We do this by a mixture of traditional literature survey and correlation analysis, as well as by running empirical evaluations. We find that many metrics are not compatible and highly depend on templates, attribute and target seeds and the choice of embeddings.

Posted on April 21, 2021

Attitudes Towards COVID-19 Measures

Measuring Shifts in Belgium Using Multilingual BERT

We classify seven months' worth of Belgian COVID-related Tweets using multilingual BERT and relate them to their governments' COVID measures. We classify Tweets by their stated opinion on Belgian government curfew measures (too strict, ok, too loose). We examine the change in topics discussed and views expressed over time and in reference to dates of related events such as implementation of new measures or COVID-19 related announcements in the media.

Posted on September 14, 2020

Ethical Adversaries

Towards Mitigating Unfairness with Adversarial Machine Learning

We offer a new framework that assists in mitigating unfair representations in the dataset used for training. Our framework relies on adversaries to improve fairness. First, it evaluates a model for unfairness w.r.t. protected attributes and ensures that an adversary cannot guess such attributes for a given outcome, by optimizing the model’s parameters for fairness while limiting utility losses. Second, the framework leverages evasion attacks from adversarial machine learning to perform adversarial retraining with new examples unseen by the model. We evaluated our framework on well-studied datasets in the fairness literature where it can surpass other approaches concerning demographic parity, equality of opportunity and also the model’s utility.

Posted on January 20, 2020


A Dutch RoBERTa-based Language Model

RobBERT is the state-of-the-art Dutch BERT model. It is a large pre-trained general Dutch language model that can be fine-tuned on a given dataset to perform any text classification, regression or token-tagging task. As such, it has been successfully used by many researchers and practitioners for achieving state-of-the-art performance for a wide range of Dutch natural language processing tasks.

Posted on October 03, 2019

Time to Take Emoji Seriously

They Vastly Improve Casual Conversational Models

Graphical emoji are ubiquitous in modern-day online conversations. So is a single thumbs-up emoji able to signify an agreement, without any words. Current state-of-the-art systems are ill-equipped to correctly interpret these emoji, especially in a conversational context. However, in a casual context, the benefits might be high: a better understanding of users’ utterances and more natural, emoji-rich responses.<br/>With this in mind, we modify BERT to fully support emoji, both from the Unicode Standard and custom emoji. This modified BERT is then trained on a corpus of question-answer (QA) tuples with a high number of emoji.

Posted on August 01, 2019

Computational Ad Hominem Detection

Fallacies like the personal attack—also known as the ad hominem attack—are introduced in debates as an easy win, even though they provide no rhetorical contribution. Although their importance in argumentation mining is acknowledged, automated mining and analysis is still lacking. We show TF-IDF approaches are insufficient to detect the ad hominem attack. Therefore we present a machine learning approach for information extraction, which has a recall of 80% for a social media data source. We also demonstrate our approach with an application that uses online learning.