Branislav Pecher

Matus Tomlein, Branislav Pecher, Jakub Simko et al.

The negative effects of misinformation filter bubbles in adaptive systems have been known to researchers for some time. Several studies investigated, most prominently on YouTube, how fast a user can get into a misinformation filter bubble simply by selecting wrong choices from the items offered. Yet, no studies so far have investigated what it takes to burst the bubble, i.e., revert the bubble enclosure. We present a study in which pre-programmed agents (acting as YouTube users) delve into misinformation filter bubbles by watching misinformation promoting content (for various topics). Then, by watching misinformation debunking content, the agents try to burst the bubbles and reach more balanced recommendation mixes. We recorded the search results and recommendations, which the agents encountered, and analyzed them for the presence of misinformation. Our key finding is that bursting of a filter bubble is possible, albeit it manifests differently from topic to topic. Moreover, we observe that filter bubbles do not truly appear in some situations. We also draw a direct comparison with a previous study. Sadly, we did not find much improvements in misinformation occurrences, despite recent pledges by YouTube.

Ivan Srba, Robert Moro, Matus Tomlein et al.

In this paper, we present results of an auditing study performed over YouTube aimed at investigating how fast a user can get into a misinformation filter bubble, but also what it takes to "burst the bubble", i.e., revert the bubble enclosure. We employ a sock puppet audit methodology, in which pre-programmed agents (acting as YouTube users) delve into misinformation filter bubbles by watching misinformation promoting content. Then they try to burst the bubbles and reach more balanced recommendations by watching misinformation debunking content. We record search results, home page results, and recommendations for the watched videos. Overall, we recorded 17,405 unique videos, out of which we manually annotated 2,914 for the presence of misinformation. The labeled data was used to train a machine learning model classifying videos into three classes (promoting, debunking, neutral) with the accuracy of 0.82. We use the trained model to classify the remaining videos that would not be feasible to annotate manually. Using both the manually and automatically annotated data, we observe the misinformation bubble dynamics for a range of audited topics. Our key finding is that even though filter bubbles do not appear in some situations, when they do, it is possible to burst them by watching misinformation debunking content (albeit it manifests differently from topic to topic). We also observe a sudden decrease of misinformation filter bubble effect when misinformation debunking videos are watched after misinformation promoting videos, suggesting a strong contextuality of recommendations. Finally, when comparing our results with a previous similar study, we do not observe significant improvements in the overall quantity of recommended misinformation content.

Ivan Srba, Branislav Pecher, Matus Tomlein et al.

False information has a significant negative influence on individuals as well as on the whole society. Especially in the current COVID-19 era, we witness an unprecedented growth of medical misinformation. To help tackle this problem with machine learning approaches, we are publishing a feature-rich dataset of approx. 317k medical news articles/blogs and 3.5k fact-checked claims. It also contains 573 manually and more than 51k automatically labelled mappings between claims and articles. Mappings consist of claim presence, i.e., whether a claim is contained in a given article, and article stance towards the claim. We provide several baselines for these two tasks and evaluate them on the manually labelled part of the dataset. The dataset enables a number of additional tasks related to medical misinformation, such as misinformation characterisation studies or studies of misinformation diffusion between sources.

Timo Hromadka, Timotej Smolen, Tomas Remis et al.

This paper presents the best-performing solution to the SemEval 2023 Task 3 on the subtask 3 dedicated to persuasion techniques detection. Due to a high multilingual character of the input data and a large number of 23 predicted labels (causing a lack of labelled data for some language-label combinations), we opted for fine-tuning pre-trained transformer-based language models. Conducting multiple experiments, we find the best configuration, which consists of large multilingual model (XLM-RoBERTa large) trained jointly on all input data, with carefully calibrated confidence thresholds for seen and surprise languages separately. Our final system performed the best on 6 out of 9 languages (including two surprise languages) and achieved highly competitive results on the remaining three languages.

Branislav Pecher, Jan Cegin, Robert Belanec et al.

Large Language Models (LLMs) have demonstrated remarkable multilingual capabilities, making them promising tools in both high- and low-resource languages. One particularly valuable use case is generating synthetic samples that can be used to train smaller models in low-resource scenarios where human-labelled data is scarce. In this work, we investigate whether these synthetic data generation capabilities can serve as a form of distillation, producing smaller models that perform on par with or even better than massive LLMs across languages and tasks. To this end, we use a state-of-the-art multilingual LLM to generate synthetic datasets covering 11 languages and 4 classification tasks. These datasets are then used to train smaller models via fine-tuning or instruction tuning, or as synthetic in-context examples for compact LLMs. Our experiments show that even small amounts of synthetic data enable smaller models to outperform the large generator itself, particularly in low-resource languages. Overall, the results suggest that LLMs are best utilised as generators (teachers) rather than classifiers, producing data that empowers smaller and more efficient multilingual models.

Branislav Pecher, Adrian Bindas, Jan Jakubcik et al.

Social media platforms have become an integral part of everyday life, serving as a primary source of news and information for many users. These platforms increasingly rely on personalised recommendation systems that shape what users see and engage with. While these systems are optimised for engagement, concerns have emerged that they may also drive users toward more polarised perspectives, particularly in contested domains such as politics, climate change, vaccines, and conspiracy theories. In this paper, we present an algorithmic audit of personalisation drift on TikTok in these polarising topics. Using controlled accounts designed to simulate users with interests aligned with or opposed to different polarising topics, we systematically measure the extent to which TikTok steers content exposure toward specific topics and polarities over time. Specifically, we investigated: 1) a preference-aligned drift (showing a strong personalisation towards user interests), 2) a polarisation-topic drift (showing a strong neutralising effect for misinformation-themed topics, and a high preference and reinforcement of interest of US politic topic); and 3) a polarisation-stance drift (showing a preference of oppose stance towards US politics topic and a general reinforcement of users' stance by recommending items aligned with their stance towards polarising topics). Overall, our findings provide evidence that recommendation trajectories differ markedly across topics, with some pathways amplifying polarised viewpoints more strongly than others and offer insights for platform governance, transparency and user awareness.

Branislav Pecher, Michal Spiegel, Robert Belanec et al.

Large language models (LLMs) are widely used as zero-shot and few-shot classifiers, where task behaviour is largely controlled through prompting. A growing number of works have observed that LLMs are sensitive to prompt variations, with small changes leading to large changes in performance. However, in many cases, the investigation of sensitivity is performed using underspecified prompts that provide minimal task instructions and weakly constrain the model's output space. In this work, we argue that a significant portion of the observed prompt sensitivity can be attributed to prompt underspecification. We systematically study and compare the sensitivity of underspecified prompts and prompts that provide specific instructions. Utilising performance analysis, logit analysis, and linear probing, we find that underspecified prompts exhibit higher performance variance and lower logit values for relevant tokens, while instruction-prompts suffer less from such problems. However, linear probing analysis suggests that the effects of prompt underspecification have only a marginal impact on the internal LLM representations, instead emerging in the final layers. Overall, our findings highlight the need for more rigour when investigating and mitigating prompt sensitivity.

Jan Cegin, Branislav Pecher, Jakub Simko et al.

The latest generative large language models (LLMs) have found their application in data augmentation tasks, where small numbers of text samples are LLM-paraphrased and then used to fine-tune downstream models. However, more research is needed to assess how different prompts, seed data selection strategies, filtering methods, or model settings affect the quality of paraphrased data (and downstream models). In this study, we investigate three text diversity incentive methods well established in crowdsourcing: taboo words, hints by previous outlier solutions, and chaining on previous outlier solutions. Using these incentive methods as part of instructions to LLMs augmenting text datasets, we measure their effects on generated texts lexical diversity and downstream model performance. We compare the effects over 5 different LLMs, 6 datasets and 2 downstream models. We show that diversity is most increased by taboo words, but downstream model performance is highest with hints.

Branislav Pecher, Ivan Srba, Maria Bielikova

When solving NLP tasks with limited labelled data, researchers typically either use a general large language model without further update, or use a small number of labelled samples to tune a specialised smaller model. In this work, we answer an important question -- how many labelled samples are required for the specialised small models to outperform general large models, while taking the performance variance into consideration. By observing the behaviour of fine-tuning, instruction-tuning, prompting and in-context learning on 8 language models, we identify such performance break-even points across 8 representative text classification tasks of varying characteristics. We show that the specialised models often need only few samples (on average $100$) to be on par or better than the general ones. At the same time, the number of required labels strongly depends on the dataset or task characteristics, with fine-tuning on binary datasets requiring significantly more samples. When performance variance is taken into consideration, the number of required labels increases on average by $100 - 200\%$. Finally, larger models do not consistently lead to better performance and lower variance, with 4-bit quantisation having negligible impact.

Branislav Pecher, Ivan Srba, Maria Bielikova

While learning with limited labelled data can improve performance when the labels are lacking, it is also sensitive to the effects of uncontrolled randomness introduced by so-called randomness factors (e.g., varying order of data). We propose a method to systematically investigate the effects of randomness factors while taking the interactions between them into consideration. To measure the true effects of an individual randomness factor, our method mitigates the effects of other factors and observes how the performance varies across multiple runs. Applying our method to multiple randomness factors across in-context learning and fine-tuning approaches on 7 representative text classification tasks and meta-learning on 3 tasks, we show that: 1) disregarding interactions between randomness factors in existing works caused inconsistent findings due to incorrect attribution of the effects of randomness factors, such as disproving the consistent sensitivity of in-context learning to sample order even with random sample selection; and 2) besides mutual interactions, the effects of randomness factors, especially sample order, are also dependent on more systematic choices unexplored in existing works, such as number of classes, samples per class or choice of prompt format.

Jan Cegin, Branislav Pecher, Jakub Simko et al.

The generative large language models (LLMs) are increasingly used for data augmentation tasks, where text samples are paraphrased (or generated anew) and then used for classifier fine-tuning. Existing works on augmentation leverage the few-shot scenarios, where samples are given to LLMs as part of prompts, leading to better augmentations. Yet, the samples are mostly selected randomly and a comprehensive overview of the effects of other (more ``informed'') sample selection strategies is lacking. In this work, we compare sample selection strategies existing in few-shot learning literature and investigate their effects in LLM-based textual augmentation. We evaluate this on in-distribution and out-of-distribution classifier performance. Results indicate, that while some ``informed'' selection strategies increase the performance of models, especially for out-of-distribution data, it happens only seldom and with marginal performance increases. Unless further advances are made, a default of random sample selection remains a good option for augmentation practitioners.

Branislav Pecher, Ivan Srba, Maria Bielikova et al.

In few-shot learning, such as meta-learning, few-shot fine-tuning or in-context learning, the limited number of samples used to train a model have a significant impact on the overall success. Although a large number of sample selection strategies exist, their impact on the performance of few-shot learning is not extensively known, as most of them have been so far evaluated in typical supervised settings only. In this paper, we thoroughly investigate the impact of 20 sample selection strategies on the performance of 5 few-shot learning approaches over 8 image and 6 text datasets. In addition, we propose a new method for automatic combination of sample selection strategies (ACSESS) that leverages the strengths and complementary information of the individual strategies. The experimental results show that our method consistently outperforms the individual selection strategies, as well as the recently proposed method for selecting support examples for in-context learning. We also show a strong modality, dataset and approach dependence for the majority of strategies as well as their dependence on the number of shots - demonstrating that the sample selection strategies play a significant role for lower number of shots, but regresses to random selection at higher number of shots.

Robert Belanec, Branislav Pecher, Ivan Srba et al.

Despite the state-of-the-art performance of Large Language Models (LLMs) achieved on many tasks, their massive scale often leads to high computational and environmental costs, limiting their accessibility. Parameter-Efficient Fine-Tuning (PEFT) methods address this challenge by reducing the number of trainable parameters while maintaining strong downstream performance. Despite the advances in PEFT methods, current evaluations remain limited (in terms of evaluated models and datasets) and difficult to reproduce. To bridge this gap, we introduce PEFT-Bench, a unified end-to-end benchmark for evaluating diverse PEFT methods on autoregressive LLMs. We demonstrate its usage across 27 NLP datasets and 7 PEFT methods. To account for different PEFT training and inference factors, we also introduce the PEFT Soft Cost Penalties (PSCP) metric, which takes trainable parameters, inference speed, and training memory usage into account.

Jan Cegin, Branislav Pecher, Ivan Srba et al.

LLMs are powerful generators of synthetic data, which are used for training smaller, specific models. This is especially valuable for low-resource languages, where human-labelled data is scarce but LLMs can still produce high-quality text. However, LLMs differ in how useful their outputs are for training. Selecting the best LLM as a generator is challenging because extrinsic evaluation requires costly human annotations (which are often unavailable for low-resource languages), while intrinsic metrics correlate poorly with downstream performance. We introduce Round robin Synthetic data Evaluation (RoSE), a proxy metric for selecting the best LLM generator without human test sets. RoSE trains a small model on the outputs of a candidate generator (LLM) and then evaluates it on generated synthetic examples from all other candidate LLMs. The final RoSE score is the mean performance of this small model. Across six LLMs, eleven languages, and three tasks (sentiment, topic, intent), RoSE identifies the optimal generator more often than any other intrinsic heuristics. RoSE outperforms intrinsic heuristics and comes within 0.76 percentage points of the optimal generator baseline. This result is measured in terms of downstream performance, obtained by training a small model on the chosen generator's outputs (optimal vs. proxy metric selected) and evaluating it on human-labelled test data. Additionally, RoSE is the only metric to achieve a positive correlation with performance on human test data.

Branislav Pecher, Jan Cegin, Robert Belanec et al.

While fine-tuning of pre-trained language models generally helps to overcome the lack of labelled training samples, it also displays model performance instability. This instability mainly originates from randomness in initialisation or data shuffling. To address this, researchers either modify the training process or augment the available samples, which typically results in increased computational costs. We propose a new mitigation strategy, called Delayed Ensemble with Noisy Interpolation (DENI), that leverages the strengths of ensembling, noise regularisation and model interpolation, while retaining computational efficiency. We compare DENI with 9 representative mitigation strategies across 3 models, 4 tuning strategies and 7 text classification datasets. We show that: 1) DENI outperforms the best performing mitigation strategy (Ensemble), while using only a fraction of its cost; 2) the mitigation strategies are beneficial for parameter-efficient fine-tuning (PEFT) methods, outperforming full fine-tuning in specific cases; and 3) combining DENI with data augmentation often leads to even more effective instability mitigation.