Pieter Delobelle

Xavier Suau, Pieter Delobelle, Katherine Metcalf et al.

An important issue with Large Language Models (LLMs) is their undesired ability to generate toxic language. In this work, we show that the neurons responsible for toxicity can be determined by their power to discriminate toxic sentences, and that toxic language can be mitigated by reducing their activation levels proportionally to this power. We propose AUROC adaptation (AurA), an intervention that can be applied to any pre-trained LLM to mitigate toxicity. As the intervention is proportional to the ability of each neuron to discriminate toxic content, it is free of any model-dependent hyperparameters. We show that AurA can achieve up to $2.2 \times$ reduction in toxicity with only a $0.72$ perplexity increase. We also show that AurA is effective with models of different scale (from 1.5B to 40B parameters), and its effectiveness in mitigating toxic language, while preserving common-sense zero-shot abilities, holds across all scales. AurA can be combined with pre-prompting strategies, boosting its average mitigation potential from $1.28\times$ to $2.35\times$. Moreover, AurA can counteract adversarial pre-prompts that maliciously elicit toxic content, making it an effective method for deploying safer and less toxic models.

Pieter Delobelle, Thomas Winters, Bettina Berendt

Pre-trained large-scale language models such as BERT have gained a lot of attention thanks to their outstanding performance on a wide range of natural language tasks. However, due to their large number of parameters, they are resource-intensive both to deploy and to fine-tune. Researchers have created several methods for distilling language models into smaller ones to increase efficiency, with a small performance trade-off. In this paper, we create several different distilled versions of the state-of-the-art Dutch RobBERT model and call them RobBERTje. The distillations differ in their distillation corpus, namely whether or not they are shuffled and whether they are merged with subsequent sentences. We found that the performance of the models using the shuffled versus non-shuffled datasets is similar for most tasks and that randomly merging subsequent sentences in a corpus creates models that train faster and perform better on tasks with long sequences. Upon comparing distillation architectures, we found that the larger DistilBERT architecture worked significantly better than the Bort hyperparametrization. Interestingly, we also found that the distilled models exhibit less gender-stereotypical bias than its teacher model. Since smaller architectures decrease the time to fine-tune, these models allow for more efficient training and more lightweight deployment of many Dutch downstream language tasks.

Ewoenam Tokpo, Pieter Delobelle, Bettina Berendt et al.

To mitigate gender bias in contextualized language models, different intrinsic mitigation strategies have been proposed, alongside many bias metrics. Considering that the end use of these language models is for downstream tasks like text classification, it is important to understand how these intrinsic bias mitigation strategies actually translate to fairness in downstream tasks and the extent of this. In this work, we design a probe to investigate the effects that some of the major intrinsic gender bias mitigation strategies have on downstream text classification tasks. We discover that instead of resolving gender bias, intrinsic mitigation techniques and metrics are able to hide it in such a way that significant gender information is retained in the embeddings. Furthermore, we show that each mitigation technique is able to hide the bias from some of the intrinsic bias measures but not all, and each intrinsic bias measure can be fooled by some mitigation techniques, but not all. We confirm experimentally, that none of the intrinsic mitigation techniques used without any other fairness intervention is able to consistently impact extrinsic bias. We recommend that intrinsic bias mitigation techniques should be combined with other fairness interventions for downstream tasks.

Pieter Delobelle, Bettina Berendt

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. We found that our distillation method does not negatively affect the downstream performance on most tasks and successfully mitigates stereotyping and representational harms. We demonstrate that FairDistillation can create fairer language models at a considerably lower cost than alternative approaches.

François Remy, Pieter Delobelle, Hayastan Avetisyan et al.

The development of monolingual language models for low and mid-resource languages continues to be hindered by the difficulty in sourcing high-quality training data. In this study, we present a novel cross-lingual vocabulary transfer strategy, trans-tokenization, designed to tackle this challenge and enable more efficient language adaptation. Our approach focuses on adapting a high-resource monolingual LLM to an unseen 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. Additionally, we introduce Hydra LLMs, models with multiple swappable language modeling heads and embedding tables, which further extend the capabilities of our trans-tokenization strategy. By designing a Hydra LLM based on the multilingual model TowerInstruct, we developed a state-of-the-art machine translation model for Tatar, in a zero-shot manner, completely bypassing the need for high-quality parallel data. This breakthrough is particularly significant for low-resource languages like Tatar, where high-quality parallel data is hard to come by. By lowering the data and time requirements for training high-quality models, our trans-tokenization strategy allows for the development of LLMs for a wider range of languages, especially those with limited resources. We hope that our work will inspire further research and collaboration in the field of cross-lingual vocabulary transfer and contribute to the empowerment of languages on a global scale.

Pieter Delobelle, Thomas Winters, Bettina Berendt

Large transformer-based language models, e.g. BERT and GPT-3, outperform previous architectures on most natural language processing tasks. Such language models are first pre-trained on gigantic corpora of text and later used as base-model for finetuning on a particular task. Since the pre-training step is usually not repeated, base models are not up-to-date with the latest information. In this paper, we update RobBERT, a RoBERTa-based state-of-the-art Dutch language model, which was trained in 2019. First, the tokenizer of RobBERT is updated to include new high-frequent tokens present in the latest Dutch OSCAR corpus, e.g. corona-related words. Then we further pre-train the RobBERT model using this dataset. To evaluate if our new model is a plug-in replacement for RobBERT, we introduce two additional criteria based on concept drift of existing tokens and alignment for novel tokens.We found that for certain language tasks this update results in a significant performance increase. These results highlight the benefit of continually updating a language model to account for evolving language use.

François Remy, Pieter Delobelle, Bettina Berendt et al.

Training monolingual language models for low and mid-resource languages is made challenging by limited and often inadequate pretraining data. In this study, we propose a novel model conversion strategy to address this issue, adapting high-resources monolingual language models to a new target language. By generalizing over a word translation dictionary encompassing both the source and target languages, we map tokens from the target tokenizer to semantically similar tokens from the source language tokenizer. This one-to-many token mapping improves tremendously the initialization of the embedding table for the target language. We conduct experiments to convert high-resource models to mid- and low-resource languages, namely Dutch and Frisian. These converted models achieve a new state-of-the-art performance on these languages across all sorts of downstream tasks. By reducing significantly the amount of data and time required for training state-of-the-art models, our novel model conversion strategy has the potential to benefit many languages worldwide.

Rik Adriaensen, Lucas Van Praet, Jessa Bekker et al.

Operationalizing definitions of fairness is difficult in practice, as multiple definitions can be incompatible while each being arguably desirable. Instead, it may be easier to directly describe algorithmic bias through ad-hoc assumptions specific to a particular real-world task, e.g., based on background information on systemic biases in its context. Such assumptions can, in turn, be used to mitigate this bias during training. Yet, a framework for incorporating such assumptions that is simultaneously principled, flexible, and interpretable is currently lacking. Our approach is to formalize bias assumptions as programs in ProbLog, a probabilistic logic programming language that allows for the description of probabilistic causal relationships through logic. Neurosymbolic extensions of ProbLog then allow for easy integration of these assumptions in a neural network's training process. We propose a set of templates to express different types of bias and show the versatility of our approach on synthetic tabular datasets with known biases. Using estimates of the bias distortions present, we also succeed in mitigating algorithmic bias in real-world tabular and image data. We conclude that ProbLog4Fairness outperforms baselines due to its ability to flexibly model the relevant bias assumptions, where other methods typically uphold a fixed bias type or notion of fairness.

Christoph Rauchegger, Sonja Mei Wang, Pieter Delobelle

The FIFA World Cup in Qatar was discussed extensively in the news and on social media. Due to news reports with allegations of human rights violations, there were calls to boycott it. Wearing a OneLove armband was part of a planned protest activity. Controversy around the armband arose when FIFA threatened to sanction captains who wear it. To understand what topics Twitter users Tweeted about and what the opinion of German Twitter users was towards the OneLove armband, we performed an analysis of German Tweets published during the World Cup using in-context learning with LLMs. We validated the labels on human annotations. We found that Twitter users initially discussed the armband's impact, LGBT rights, and politics; after the ban, the conversation shifted towards politics in sports in general, accompanied by a subtle shift in sentiment towards neutrality. Our evaluation serves as a framework for future research to explore the impact of sports activism and evolving public sentiment. This is especially useful in settings where labeling datasets for specific opinions is unfeasible, such as when events are unfolding.

David Hartmann, Manuel Tonneau, Angelie Kraft et al.

The closure of Perspective API at the end of 2026 discards what has functioned as the de facto standard for automated toxicity measurement in NLP, CSS, and LLM evaluation research. We document the structural dependence that the communities built on this single proprietary tool and discuss how this dependence caused epistemic problems that have affected - and will likely continue to affect - collective research efforts. Perspective's model was periodically updated without versioning or disclosure, its annotation structure reflected a single corporate operationalisation of a contested concept, and its scores were used simultaneously as an evaluation target and an evaluation standard. Its closure leaves behind non-updatable benchmarks, irreproducible results, and ultimately a field at risk of perpetuating these issues by turning to closed-source LLMs. We use Perspective's announced termination as an opportunity to call for an independent, valid, adaptable, and reproducible toxicity and hate speech measurement infrastructure, with the technical and governance requirements outlined in this paper.

Matthieu Meeus, Anthony Rathé, François Remy et al.

While Large Language Models (LLMs) have shown remarkable capabilities in natural language understanding and generation, their performance often lags in lower-resource, non-English languages due to biases in the training data. In this work, we explore strategies for adapting the primarily English LLMs (Llama-2 and Llama-3) to Dutch, a language spoken by 30 million people worldwide yet often underrepresented in LLM development. We collect 104GB of Dutch text ($32$B tokens) from various sources to first apply continued pretraining using low-rank adaptation (LoRA), complemented with Dutch posttraining strategies provided by prior work. For Llama-2, we consider using (i) the tokenizer of the original model, and (ii) training a new, Dutch-specific tokenizer combined with embedding reinitialization. We evaluate our adapted models, ChocoLlama-2, both on standard benchmarks and a novel Dutch benchmark, ChocoLlama-Bench. Our results demonstrate that LoRA can effectively scale for language adaptation, and that tokenizer modification with careful weight reinitialization can improve performance. Notably, Llama-3 was released during the course of this project and, upon evaluation, demonstrated superior Dutch capabilities compared to our Dutch-adapted versions of Llama-2. We hence apply the same adaptation technique to Llama-3, using its original tokenizer. While our adaptation methods enhanced Llama-2's Dutch capabilities, we found limited gains when applying the same techniques to Llama-3. This suggests that for ever improving, multilingual foundation models, language adaptation techniques may benefit more from focusing on language-specific posttraining rather than on continued pretraining. We hope this work contributes to the broader understanding of adapting LLMs to lower-resource languages, and to the development of Dutch LLMs in particular.

Pieter Delobelle, Ewoenam Kwaku Tokpo, Toon Calders et al.

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 (i) templates, (ii) attribute and target seeds and (iii) the choice of embeddings. These results indicate that fairness or bias evaluation remains challenging for contextualized language models, if not at least highly subjective. To improve future comparisons and fairness evaluations, we recommend avoiding embedding-based metrics and focusing on fairness evaluations in downstream tasks.

Kristen Scott, Pieter Delobelle, Bettina Berendt

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.

Thomas Winters, Pieter Delobelle

Detecting if a text is humorous is a hard task to do computationally, as it usually requires linguistic and common sense insights. In machine learning, humor detection is usually modeled as a binary classification task, trained to predict if the given text is a joke or another type of text. Rather than using completely different non-humorous texts, we propose using text generation algorithms for imitating the original joke dataset to increase the difficulty for the learning algorithm. We constructed several different joke and non-joke datasets to test the humor detection abilities of different language technologies. In particular, we compare the humor detection capabilities of classic neural network approaches with the state-of-the-art Dutch language model RobBERT. In doing so, we create and compare the first Dutch humor detection systems. We found that while other language models perform well when the non-jokes came from completely different domains, RobBERT was the only one that was able to distinguish jokes from generated negative examples. This performance illustrates the usefulness of using text generation to create negative datasets for humor recognition, and also shows that transformer models are a large step forward in humor detection.

Pieter Delobelle, Paul Temple, Gilles Perrouin et al.

Machine learning is being integrated into a growing number of critical systems with far-reaching impacts on society. Unexpected behaviour and unfair decision processes are coming under increasing scrutiny due to this widespread use and its theoretical considerations. Individuals, as well as organisations, notice, test, and criticize unfair results to hold model designers and deployers accountable. We offer a framework that assists these groups in mitigating unfair representations stemming from the training datasets. Our framework relies on two inter-operating adversaries to improve fairness. First, a model is trained with the goal of preventing the guessing of protected attributes' values while limiting utility losses. This first step optimizes the model's parameters for fairness. Second, the framework leverages evasion attacks from adversarial machine learning to generate new examples that will be misclassified. These new examples are then used to retrain and improve the model in the first step. These two steps are iteratively applied until a significant improvement in fairness is obtained. We evaluated our framework on well-studied datasets in the fairness literature -- including COMPAS -- where it can surpass other approaches concerning demographic parity, equality of opportunity and also the model's utility. We also illustrate our findings on the subtle difficulties when mitigating unfairness and highlight how our framework can assist model designers.

Pieter Delobelle, Thomas Winters, Bettina Berendt

Pre-trained language models have been dominating the field of natural language processing in recent years, and have led to significant performance gains for various complex natural language tasks. One of the most prominent pre-trained language models is BERT, which was released as an English as well as a multilingual version. Although multilingual BERT performs well on many tasks, recent studies show that BERT models trained on a single language significantly outperform the multilingual version. Training a Dutch BERT model thus has a lot of potential for a wide range of Dutch NLP tasks. While previous approaches have used earlier implementations of BERT to train a Dutch version of BERT, we used RoBERTa, a robustly optimized BERT approach, to train a Dutch language model called RobBERT. We measured its performance on various tasks as well as the importance of the fine-tuning dataset size. We also evaluated the importance of language-specific tokenizers and the model's fairness. We found that RobBERT improves state-of-the-art results for various tasks, and especially significantly outperforms other models when dealing with smaller datasets. These results indicate that it is a powerful pre-trained model for a large variety of Dutch language tasks. The pre-trained and fine-tuned models are publicly available to support further downstream Dutch NLP applications.

Pieter Delobelle, Bettina Berendt

Graphical emoji are ubiquitous in modern-day online conversations. So is a single thumbs-up emoji able to signify an agreement, without any words. We argue that the 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. 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, where we're able to increase the 1-of-100 accuracy from 12.7% for the current state-of-the-art to 17.8% for our model with emoji support.