Malvina Nissim

Gabriele Sarti, Nils Feldhus, Ludwig Sickert et al.

Past work in natural language processing interpretability focused mainly on popular classification tasks while largely overlooking generation settings, partly due to a lack of dedicated tools. In this work, we introduce Inseq, a Python library to democratize access to interpretability analyses of sequence generation models. Inseq enables intuitive and optimized extraction of models' internal information and feature importance scores for popular decoder-only and encoder-decoder Transformers architectures. We showcase its potential by adopting it to highlight gender biases in machine translation models and locate factual knowledge inside GPT-2. Thanks to its extensible interface supporting cutting-edge techniques such as contrastive feature attribution, Inseq can drive future advances in explainable natural language generation, centralizing good practices and enabling fair and reproducible model evaluations.

Gabriele Sarti, Malvina Nissim

We introduce IT5, the first family of encoder-decoder transformer models pretrained specifically on Italian. We document and perform a thorough cleaning procedure for a large Italian corpus and use it to pretrain four IT5 model sizes. We then introduce the ItaGen benchmark, which includes a broad range of natural language understanding and generation tasks for Italian, and use it to evaluate the performance of IT5 models and multilingual baselines. We find monolingual IT5 models to provide the best scale-to-performance ratio across tested models, consistently outperforming their multilingual counterparts and setting a new state-of-the-art for Italian language generation.

Huiyuan Lai, Antonio Toral, Malvina Nissim

We investigate the potential of ChatGPT as a multidimensional evaluator for the task of \emph{Text Style Transfer}, alongside, and in comparison to, existing automatic metrics as well as human judgements. We focus on a zero-shot setting, i.e. prompting ChatGPT with specific task instructions, and test its performance on three commonly-used dimensions of text style transfer evaluation: style strength, content preservation, and fluency. We perform a comprehensive correlation analysis for two transfer directions (and overall) at different levels. Compared to existing automatic metrics, ChatGPT achieves competitive correlations with human judgments. These preliminary results are expected to provide a first glimpse into the role of large language models in the multidimensional evaluation of stylized text generation.

Huiyuan Lai, Jiali Mao, Antonio Toral et al.

Although text style transfer has witnessed rapid development in recent years, there is as yet no established standard for evaluation, which is performed using several automatic metrics, lacking the possibility of always resorting to human judgement. We focus on the task of formality transfer, and on the three aspects that are usually evaluated: style strength, content preservation, and fluency. To cast light on how such aspects are assessed by common and new metrics, we run a human-based evaluation and perform a rich correlation analysis. We are then able to offer some recommendations on the use of such metrics in formality transfer, also with an eye to their generalisability (or not) to related tasks.

Huiyuan Lai, Antonio Toral, Malvina Nissim

We exploit the pre-trained seq2seq model mBART for multilingual text style transfer. Using machine translated data as well as gold aligned English sentences yields state-of-the-art results in the three target languages we consider. Besides, in view of the general scarcity of parallel data, we propose a modular approach for multilingual formality transfer, which consists of two training strategies that target adaptation to both language and task. Our approach achieves competitive performance without monolingual task-specific parallel data and can be applied to other style transfer tasks as well as to other languages.

Gosse Minnema, Sara Gemelli, Chiara Zanchi et al.

SOCIOFILLMORE is a multilingual tool which helps to bring to the fore the focus or the perspective that a text expresses in depicting an event. Our tool, whose rationale we also support through a large collection of human judgements, is theoretically grounded on frame semantics and cognitive linguistics, and implemented using the LOME frame semantic parser. We describe SOCIOFILLMORE's development and functionalities, show how non-NLP researchers can easily interact with the tool, and present some example case studies which are already incorporated in the system, together with the kind of analysis that can be visualised.

Gosse Minnema, Sara Gemelli, Chiara Zanchi et al.

Different linguistic expressions can conceptualize the same event from different viewpoints by emphasizing certain participants over others. Here, we investigate a case where this has social consequences: how do linguistic expressions of gender-based violence (GBV) influence who we perceive as responsible? We build on previous psycholinguistic research in this area and conduct a large-scale perception survey of GBV descriptions automatically extracted from a corpus of Italian newspapers. We then train regression models that predict the salience of GBV participants with respect to different dimensions of perceived responsibility. Our best model (fine-tuned BERT) shows solid overall performance, with large differences between dimensions and participants: salient _focus_ is more predictable than salient _blame_, and perpetrators' salience is more predictable than victims' salience. Experiments with ridge regression models using different representations show that features based on linguistic theory similarly to word-based features. Overall, we show that different linguistic choices do trigger different perceptions of responsibility, and that such perceptions can be modelled automatically. This work can be a core instrument to raise awareness of the consequences of different perspectivizations in the general public and in news producers alike.

Gabriele Sarti, Grzegorz Chrupała, Malvina Nissim et al.

Establishing whether language models can use contextual information in a human-plausible way is important to ensure their trustworthiness in real-world settings. However, the questions of when and which parts of the context affect model generations are typically tackled separately, with current plausibility evaluations being practically limited to a handful of artificial benchmarks. To address this, we introduce Plausibility Evaluation of Context Reliance (PECoRe), an end-to-end interpretability framework designed to quantify context usage in language models' generations. Our approach leverages model internals to (i) contrastively identify context-sensitive target tokens in generated texts and (ii) link them to contextual cues justifying their prediction. We use \pecore to quantify the plausibility of context-aware machine translation models, comparing model rationales with human annotations across several discourse-level phenomena. Finally, we apply our method to unannotated model translations to identify context-mediated predictions and highlight instances of (im)plausible context usage throughout generation.

Gosse Minnema, Huiyuan Lai, Benedetta Muscato et al.

Different ways of linguistically expressing the same real-world event can lead to different perceptions of what happened. Previous work has shown that different descriptions of gender-based violence (GBV) influence the reader's perception of who is to blame for the violence, possibly reinforcing stereotypes which see the victim as partly responsible, too. As a contribution to raise awareness on perspective-based writing, and to facilitate access to alternative perspectives, we introduce the novel task of automatically rewriting GBV descriptions as a means to alter the perceived level of responsibility on the perpetrator. We present a quasi-parallel dataset of sentences with low and high perceived responsibility levels for the perpetrator, and experiment with unsupervised (mBART-based), zero-shot and few-shot (GPT3-based) methods for rewriting sentences. We evaluate our models using a questionnaire study and a suite of automatic metrics.

Leonidas Zotos, Hedderik van Rijn, Malvina Nissim

Estimating the difficulty of multiple-choice questions would be great help for educators who must spend substantial time creating and piloting stimuli for their tests, and for learners who want to practice. Supervised approaches to difficulty estimation have yielded to date mixed results. In this contribution we leverage an aspect of generative large models which might be seen as a weakness when answering questions, namely their uncertainty, and exploit it towards exploring correlations between two different metrics of uncertainty, and the actual student response distribution. While we observe some present but weak correlations, we also discover that the models' behaviour is different in the case of correct vs wrong answers, and that correlations differ substantially according to the different question types which are included in our fine-grained, previously unused dataset of 451 questions from a Biopsychology course. In discussing our findings, we also suggest potential avenues to further leverage model uncertainty as an additional proxy for item difficulty.

Huiyuan Lai, Malvina Nissim

Figurative language generation is the task of reformulating a given text in the desired figure of speech while still being faithful to the original context. We take the first step towards multi-figurative language modelling by providing a benchmark for the automatic generation of five common figurative forms in English. We train mFLAG employing a scheme for multi-figurative language pre-training on top of BART, and a mechanism for injecting the target figurative information into the encoder; this enables the generation of text with the target figurative form from another figurative form without parallel figurative-figurative sentence pairs. Our approach outperforms all strong baselines. We also offer some qualitative analysis and reflections on the relationship between the different figures of speech.

Malvina Nissim, Danilo Croce, Viviana Patti et al.

The rapid progress of Large Language Models (LLMs) has transformed natural language processing and broadened its impact across research and society. Yet, systematic evaluation of these models, especially for languages beyond English, remains limited. "Challenging the Abilities of LAnguage Models in ITAlian" (CALAMITA) is a large-scale collaborative benchmarking initiative for Italian, coordinated under the Italian Association for Computational Linguistics. Unlike existing efforts that focus on leaderboards, CALAMITA foregrounds methodology: it federates more than 80 contributors from academia, industry, and the public sector to design, document, and evaluate a diverse collection of tasks, covering linguistic competence, commonsense reasoning, factual consistency, fairness, summarization, translation, and code generation. Through this process, we not only assembled a benchmark of over 20 tasks and almost 100 subtasks, but also established a centralized evaluation pipeline that supports heterogeneous datasets and metrics. We report results for four open-weight LLMs, highlighting systematic strengths and weaknesses across abilities, as well as challenges in task-specific evaluation. Beyond quantitative results, CALAMITA exposes methodological lessons: the necessity of fine-grained, task-representative metrics, the importance of harmonized pipelines, and the benefits and limitations of broad community engagement. CALAMITA is conceived as a rolling benchmark, enabling continuous integration of new tasks and models. This makes it both a resource -- the most comprehensive and diverse benchmark for Italian to date -- and a framework for sustainable, community-driven evaluation. We argue that this combination offers a blueprint for other languages and communities seeking inclusive and rigorous LLM evaluation practices.

Leonidas Zotos, Hedderik van Rijn, Malvina Nissim

When students are unsure of the correct answer to a multiple-choice question (MCQ), guessing is common practice. The availability heuristic, proposed by A. Tversky and D. Kahneman in 1973, suggests that the ease with which relevant instances come to mind, typically operationalised by the mere frequency of exposure, can offer a mental shortcut for problems in which the test-taker does not know the exact answer. Is simply choosing the option that comes most readily to mind a good strategy for answering MCQs? We propose a computational method of assessing the cognitive availability of MCQ options operationalised by concepts' prevalence in large corpora. The key finding, across three large question sets, is that correct answers, independently of the question stem, are significantly more available than incorrect MCQ options. Specifically, using Wikipedia as the retrieval corpus, we find that always selecting the most available option leads to scores 13.5% to 32.9% above the random-guess baseline. We further find that LLM-generated MCQ options show similar patterns of availability compared to expert-created options, despite the LLMs' frequentist nature and their training on large collections of textual data. Our findings suggest that availability should be considered in current and future work when computationally modelling student behaviour.

Gabriele Sarti, Tommaso Caselli, Malvina Nissim et al.

Rebuses are puzzles requiring constrained multi-step reasoning to identify a hidden phrase from a set of images and letters. In this work, we introduce a large collection of verbalized rebuses for the Italian language and use it to assess the rebus-solving capabilities of state-of-the-art large language models. While general-purpose systems such as LLaMA-3 and GPT-4o perform poorly on this task, ad-hoc fine-tuning seems to improve models' performance. However, we find that performance gains from training are largely motivated by memorization. Our results suggest that rebus solving remains a challenging test bed to evaluate large language models' linguistic proficiency and sequential instruction-following skills.

Francesca Padovani, Malvina Nissim

The rapid adoption of Generative AI, including LLM-based chatbots like ChatGPT, has highlighted the need for accessible ways to support public understanding and AI literacy. To address this need, we introduce a game-based, interactive approach in the form of a jigsaw puzzle whose completed image is a comic-based infographic illustrating the workings, capabilities, limitations, and societal implications of these technologies. Each comic sketch also functions as a standalone informational card, providing focused explanations of specific facets of AI use, design, and impact. The visual content was created in a live collaborative session with a professional illustrator and a multidisciplinary group of experts and non experts, combining structured knowledge with informal, exploratory reflections shared during the discussion. By integrating hands-on assembly, visual storytelling, and collaborative interaction, the puzzle provides an engaging and playful tool for exploring the mechanisms, perks, and perils of AI systems in informal learning contexts.

Huiyuan Lai, Malvina Nissim

Large Language Models (LLMs) have shown remarkable performance on complex reasoning tasks, especially when equipped with long chain-of-thought (CoT) reasoning. However, eliciting long CoT typically requires large-scale reinforcement learning (RL) training, while often leading to overthinking with redundant intermediate steps. To improve learning and reasoning efficiency, while preserving or even enhancing performance, we propose TACLer, a model-tailored curriculum reinforcement learning framework that gradually increases the complexity of the data based on the model's proficiency in multi-stage RL training. TACLer features two core components: (i) tailored curriculum learning that determines what knowledge the model lacks and needs to learn in progressive stages; (ii) a hybrid Thinking/NoThinking reasoning paradigm that balances accuracy and efficiency by enabling or disabling the Thinking mode. Our experiments show that TACLer yields a twofold advantage in learning and reasoning: (i) it reduces computational cost, cutting training compute by over 50% compared to long thinking models and reducing inference token usage by over 42% relative to the base model; and (ii) it improves accuracy by over 9% on the base model, consistently outperforming state-of-the-art Nothinking and Thinking baselines across four math datasets with complex problems.

Daniel Scalena, Leonidas Zotos, Elisabetta Fersini et al.

With the rise of reasoning language models and test-time scaling methods as a paradigm for improving model performance, substantial computation is often required to generate multiple candidate sequences from the same prompt. This enables exploration of different reasoning paths toward the correct solution, however, allocates the same compute budget for each prompt. Grounded on the assumption that different prompts carry different degrees of complexity, and thus different computation needs, we propose EAGer, a training-free generation method that leverages model uncertainty through token-wise entropy distribution to reduce redundant computation and concurrently improve overall performance. EAGer allows branching to multiple reasoning paths only in the presence of high-entropy tokens, and then reallocates the saved compute budget to the instances where exploration of alternative paths is most needed. We find that across multiple open-source models on complex reasoning benchmarks such as AIME 2025, EAGer can reallocate the budget without accessing target labels, achieving the best efficiency-performance trade-off in terms of reasoning length and Pass@k. When target labels are accessible, EAGer generates up to 65% fewer tokens (hence saving compute) and achieves up to 37% improvement in Pass@k compared to the Full Parallel Sampling.

Huiyuan Lai, Xiao Zhang, Malvina Nissim

While Large language models (LLMs) have proved able to address some complex reasoning tasks, we also know that they are highly sensitive to input variation, which can lead to different solution paths and final answers. Answer consistency across input variations can thus be taken as a sign of stronger confidence. Leveraging this insight, we introduce a framework, {\em Multidimensional Reasoning Consistency} where, focusing on math problems, models are systematically pushed to diversify solution paths towards a final answer, thereby testing them for answer consistency across multiple input variations. We induce variations in (i) order of shots in prompt, (ii) problem phrasing, and (iii) languages used. Extensive experiments on a large range of open-source state-of-the-art LLMs of various sizes show that reasoning consistency differs by variation dimension, and that by aggregating consistency across dimensions, our framework consistently enhances mathematical reasoning performance on both monolingual dataset GSM8K and multilingual dataset MGSM, especially for smaller models.

Malvina Nissim, Viviana Patti, Beatrice Savoldi

As Natural Language Processing (NLP) systems become more pervasive, integrating ethical considerations into NLP education has become essential. However, this presents inherent challenges in curriculum development: the field's rapid evolution from both academia and industry, and the need to foster critical thinking beyond traditional technical training. We introduce our course on Ethical Aspects in NLP and our pedagogical approach, grounded in active learning through interactive sessions, hands-on activities, and "learning by teaching" methods. Over four years, the course has been refined and adapted across different institutions, educational levels, and interdisciplinary backgrounds; it has also yielded many reusable products, both in the form of teaching materials and in the form of actual educational products aimed at diverse audiences, made by the students themselves. By sharing our approach and experience, we hope to provide inspiration for educators seeking to incorporate social impact considerations into their curricula.

Huiyuan Lai, Malvina Nissim

Large language models (LLMs) with Chain-of-thought (CoT) have recently emerged as a powerful technique for eliciting reasoning to improve various downstream tasks. As most research mainly focuses on English, with few explorations in a multilingual context, the question of how reliable this reasoning capability is in different languages is still open. To address it directly, we study multilingual reasoning consistency across multiple languages, using popular open-source LLMs. First, we compile the first large-scale multilingual math reasoning dataset, mCoT-MATH, covering eleven diverse languages. Then, we introduce multilingual CoT instruction tuning to boost reasoning capability across languages, thereby improving model consistency. While existing LLMs show substantial variation across the languages we consider, and especially low performance for lesser resourced languages, our 7B parameter model mCoT achieves impressive consistency across languages, and superior or comparable performance to close- and open-source models even of much larger sizes.

Wietse de Vries, Andreas van Cranenburgh, Arianna Bisazza et al.

The transformer-based pre-trained language model BERT has helped to improve state-of-the-art performance on many natural language processing (NLP) tasks. Using the same architecture and parameters, we developed and evaluated a monolingual Dutch BERT model called BERTje. Compared to the multilingual BERT model, which includes Dutch but is only based on Wikipedia text, BERTje is based on a large and diverse dataset of 2.4 billion tokens. BERTje consistently outperforms the equally-sized multilingual BERT model on downstream NLP tasks (part-of-speech tagging, named-entity recognition, semantic role labeling, and sentiment analysis). Our pre-trained Dutch BERT model is made available at https://github.com/wietsedv/bertje.

Elizaveta Sivak, Paulina Pankowska, Adrienne Mendrik et al.

The social sciences have produced an impressive body of research on determinants of fertility outcomes, or whether and when people have children. However, the strength of these determinants and underlying theories are rarely evaluated on their predictive ability on new data. This prevents us from systematically comparing studies, hindering the evaluation and accumulation of knowledge. In this paper, we present two datasets which can be used to study the predictability of fertility outcomes in the Netherlands. One dataset is based on the LISS panel, a longitudinal survey which includes thousands of variables on a wide range of topics, including individual preferences and values. The other is based on the Dutch register data which lacks attitudinal data but includes detailed information about the life courses of millions of Dutch residents. We provide information about the datasets and the samples, and describe the fertility outcome of interest. We also introduce the fertility prediction data challenge PreFer which is based on these datasets and will start in Spring 2024. We outline the ways in which measuring the predictability of fertility outcomes using these datasets and combining their strengths in the data challenge can advance our understanding of fertility behaviour and computational social science. We further provide details for participants on how to take part in the data challenge.

Leonidas Zotos, Hedderik van Rijn, Malvina Nissim

In an educational setting, an estimate of the difficulty of multiple-choice questions (MCQs), a commonly used strategy to assess learning progress, constitutes very useful information for both teachers and students. Since human assessment is costly from multiple points of view, automatic approaches to MCQ item difficulty estimation are investigated, yielding however mixed success until now. Our approach to this problem takes a different angle from previous work: asking various Large Language Models to tackle the questions included in three different MCQ datasets, we leverage model uncertainty to estimate item difficulty. By using both model uncertainty features as well as textual features in a Random Forest regressor, we show that uncertainty features contribute substantially to difficulty prediction, where difficulty is inversely proportional to the number of students who can correctly answer a question. In addition to showing the value of our approach, we also observe that our model achieves state-of-the-art results on the USMLE and CMCQRD publicly available datasets.

Gabriele Sarti, Vilém Zouhar, Grzegorz Chrupała et al.

Word-level quality estimation (QE) methods aim to detect erroneous spans in machine translations, which can direct and facilitate human post-editing. While the accuracy of word-level QE systems has been assessed extensively, their usability and downstream influence on the speed, quality and editing choices of human post-editing remain understudied. In this study, we investigate the impact of word-level QE on machine translation (MT) post-editing in a realistic setting involving 42 professional post-editors across two translation directions. We compare four error-span highlight modalities, including supervised and uncertainty-based word-level QE methods, for identifying potential errors in the outputs of a state-of-the-art neural MT model. Post-editing effort and productivity are estimated from behavioral logs, while quality improvements are assessed by word- and segment-level human annotation. We find that domain, language and editors' speed are critical factors in determining highlights' effectiveness, with modest differences between human-made and automated QE highlights underlining a gap between accuracy and usability in professional workflows.

Leonidas Zotos, Ivo Pascal de Jong, Matias Valdenegro-Toro et al.

Estimating the difficulty of exam questions is essential for developing good exams, but professors are not always good at this task. We compare various Large Language Model-based methods with three professors in their ability to estimate what percentage of students will give correct answers on True/False exam questions in the areas of Neural Networks and Machine Learning. Our results show that the professors have limited ability to distinguish between easy and difficult questions and that they are outperformed by directly asking Gemini 2.5 to solve this task. Yet, we obtained even better results using uncertainties of the LLMs solving the questions in a supervised learning setting, using only 42 training samples. We conclude that supervised learning using LLM uncertainty can help professors better estimate the difficulty of exam questions, improving the quality of assessment.

Daniela Occhipinti, Marco Guerini, Malvina Nissim

Endowing dialogue agents with persona information has proven to significantly improve the consistency and diversity of their generations. While much focus has been placed on aligning dialogues with provided personas, the adaptation to the interlocutor's profile remains largely underexplored. In this work, we investigate three key aspects: (1) a model's ability to align responses with both the provided persona and the interlocutor's; (2) its robustness when dealing with familiar versus unfamiliar interlocutors and topics, and (3) the impact of additional fine-tuning on specific persona-based dialogues. We evaluate dialogues generated with diverse speaker pairings and topics, framing the evaluation as an author identification task and employing both LLM-as-a-judge and human evaluations. By systematically masking or disclosing information about the interlocutor, we assess its impact on dialogue generation. Results show that access to the interlocutor's persona improves the recognition of the target speaker, while masking it does the opposite. Although models generalise well across topics, they struggle with unfamiliar interlocutors. Finally, we found that in zero-shot settings, LLMs often copy biographical details, facilitating identification but trivialising the task.

Daniel Scalena, Gabriele Sarti, Arianna Bisazza et al.

Large language models have simplified the production of personalized translations reflecting predefined stylistic constraints. However, these systems still struggle when stylistic requirements are implicitly represented by a set of examples, such as texts produced by a specific human translator. In this work, we explore various strategies for personalizing automatically generated translations when few examples are available, with a focus on the challenging domain of literary translation. We begin by determining the feasibility of the task and how style information is encoded within model representations. Then, we evaluate various prompting strategies and inference-time interventions for steering model generations towards a personalized style, with a particular focus on contrastive steering with sparse autoencoder (SAE) latents to identify salient personalization properties. We demonstrate that contrastive SAE steering yields robust style conditioning and translation quality, resulting in higher inference-time computational efficiency than prompting approaches. We further examine the impact of steering on model activations, finding that layers encoding personalization properties are impacted similarly by prompting and SAE steering, suggesting a similar mechanism at play.

Gabriele Sarti, Vilém Zouhar, Malvina Nissim et al.

Word-level quality estimation (WQE) aims to automatically identify fine-grained error spans in machine-translated outputs and has found many uses, including assisting translators during post-editing. Modern WQE techniques are often expensive, involving prompting of large language models or ad-hoc training on large amounts of human-labeled data. In this work, we investigate efficient alternatives exploiting recent advances in language model interpretability and uncertainty quantification to identify translation errors from the inner workings of translation models. In our evaluation spanning 14 metrics across 12 translation directions, we quantify the impact of human label variation on metric performance by using multiple sets of human labels. Our results highlight the untapped potential of unsupervised metrics, the shortcomings of supervised methods when faced with label uncertainty, and the brittleness of single-annotator evaluation practices.

Daniel Scalena, Elisabetta Fersini, Malvina Nissim

Adapting models to a language that was only partially present in the pre-training data requires fine-tuning, which is expensive in terms of both data and computational resources. As an alternative to fine-tuning, we explore the potential of activation steering-based techniques to enhance model performance on Italian tasks. Through our experiments we show that Italian steering (i) can be successfully applied to different models, (ii) achieves performances comparable to, or even better than, fine-tuned models for Italian, and (iii) yields higher quality and consistency in Italian generations. We also discuss the utility of steering and fine-tuning in the contemporary LLM landscape where models are anyway getting high Italian performances even if not explicitly trained in this language.

Shaozhen Shi, Yevgen Matusevych, Malvina Nissim

This study presents our submission to the Strict-Small Track of the 2nd BabyLM Challenge. We use a teacher-student distillation setup with the BabyLLaMa model (Timiryasov and Tastet, 2023) as a backbone. To make the student's learning process more focused, we replace the objective function with a reverse Kullback-Leibler divergence, known to cause mode-seeking (rather than mode-averaging) behaviour in computational learners. We further experiment with having a single teacher (instead of an ensemble of two teachers) and implement additional optimization strategies to improve the distillation process. Our experiments show that under reverse KL divergence, a single-teacher model often outperforms or matches multiple-teacher models across most tasks. Additionally, incorporating advanced optimization techniques further enhances model performance, demonstrating the effectiveness and robustness of our proposed approach. These findings support our idea that "choosy babies need one coach".

Daniel Scalena, Gabriele Sarti, Malvina Nissim

Activation steering methods were shown to be effective in conditioning language model generation by additively intervening over models' intermediate representations. However, the evaluation of these techniques has so far been limited to single conditioning properties and synthetic settings. In this work, we conduct a comprehensive evaluation of various activation steering strategies, highlighting the property-dependent nature of optimal parameters to ensure a robust effect throughout generation. To address this issue, we propose Dynamic Activation Composition, an information-theoretic approach to modulate the steering intensity of one or more properties throughout generation. Our experiments on multi-property steering show that our method successfully maintains high conditioning while minimizing the impact of conditioning on generation fluency.

Daniela Occhipinti, Michele Marchi, Irene Mondella et al.

Automatic methods for generating and gathering linguistic data have proven effective for fine-tuning Language Models (LMs) in languages less resourced than English. Still, while there has been emphasis on data quantity, less attention has been given to its quality. In this work, we investigate the impact of human intervention on machine-generated data when fine-tuning dialogical models. In particular, we study (1) whether post-edited dialogues exhibit higher perceived quality compared to the originals that were automatically generated; (2) whether fine-tuning with post-edited dialogues results in noticeable differences in the generated outputs; and (3) whether post-edited dialogues influence the outcomes when considering the parameter size of the LMs. To this end we created HED-IT, a large-scale dataset where machine-generated dialogues are paired with the version post-edited by humans. Using both the edited and unedited portions of HED-IT, we fine-tuned three different sizes of an LM. Results from both human and automatic evaluation show that the different quality of training data is clearly perceived and it has an impact also on the models trained on such data. Additionally, our findings indicate that larger models are less sensitive to data quality, whereas this has a crucial impact on smaller models. These results enhance our comprehension of the impact of human intervention on training data in the development of high-quality LMs.

Daniel Scalena, Gabriele Sarti, Malvina Nissim et al.

Due to language models' propensity to generate toxic or hateful responses, several techniques were developed to align model generations with users' preferences. Despite the effectiveness of such methods in improving the safety of model interactions, their impact on models' internal processes is still poorly understood. In this work, we apply popular detoxification approaches to several language models and quantify their impact on the resulting models' prompt dependence using feature attribution methods. We evaluate the effectiveness of counter-narrative fine-tuning and compare it with reinforcement learning-driven detoxification, observing differences in prompt reliance between the two methods despite their similar detoxification performances.

Chunliu Wang, Huiyuan Lai, Malvina Nissim et al.

Pre-trained language models (PLMs) have achieved great success in NLP and have recently been used for tasks in computational semantics. However, these tasks do not fully benefit from PLMs since meaning representations are not explicitly included in the pre-training stage. We introduce multilingual pre-trained language-meaning models based on Discourse Representation Structures (DRSs), including meaning representations besides natural language texts in the same model, and design a new strategy to reduce the gap between the pre-training and fine-tuning objectives. Since DRSs are language neutral, cross-lingual transfer learning is adopted to further improve the performance of non-English tasks. Automatic evaluation results show that our approach achieves the best performance on both the multilingual DRS parsing and DRS-to-text generation tasks. Correlation analysis between automatic metrics and human judgements on the generation task further validates the effectiveness of our model. Human inspection reveals that out-of-vocabulary tokens are the main cause of erroneous results.

Huiyuan Lai, Antonio Toral, Malvina Nissim

Figures of speech help people express abstract concepts and evoke stronger emotions than literal expressions, thereby making texts more creative and engaging. Due to its pervasive and fundamental character, figurative language understanding has been addressed in Natural Language Processing, but it's highly understudied in a multilingual setting and when considering more than one figure of speech at the same time. To bridge this gap, we introduce multilingual multi-figurative language modelling, and provide a benchmark for sentence-level figurative language detection, covering three common figures of speech and seven languages. Specifically, we develop a framework for figurative language detection based on template-based prompt learning. In so doing, we unify multiple detection tasks that are interrelated across multiple figures of speech and languages, without requiring task- or language-specific modules. Experimental results show that our framework outperforms several strong baselines and may serve as a blueprint for the joint modelling of other interrelated tasks.

Wietse de Vries, Martijn Wieling, Malvina Nissim

We introduce the Dutch Model Benchmark: DUMB. The benchmark includes a diverse set of datasets for low-, medium- and high-resource tasks. The total set of nine tasks includes four tasks that were previously not available in Dutch. Instead of relying on a mean score across tasks, we propose Relative Error Reduction (RER), which compares the DUMB performance of language models to a strong baseline which can be referred to in the future even when assessing different sets of language models. Through a comparison of 14 pre-trained language models (mono- and multi-lingual, of varying sizes), we assess the internal consistency of the benchmark tasks, as well as the factors that likely enable high performance. Our results indicate that current Dutch monolingual models under-perform and suggest training larger Dutch models with other architectures and pre-training objectives. At present, the highest performance is achieved by DeBERTaV3 (large), XLM-R (large) and mDeBERTaV3 (base). In addition to highlighting best strategies for training larger Dutch models, DUMB will foster further research on Dutch. A public leaderboard is available at https://dumbench.nl.

Anya Belz, Craig Thomson, Ehud Reiter et al.

We report our efforts in identifying a set of previous human evaluations in NLP that would be suitable for a coordinated study examining what makes human evaluations in NLP more/less reproducible. We present our results and findings, which include that just 13\% of papers had (i) sufficiently low barriers to reproduction, and (ii) enough obtainable information, to be considered for reproduction, and that all but one of the experiments we selected for reproduction was discovered to have flaws that made the meaningfulness of conducting a reproduction questionable. As a result, we had to change our coordinated study design from a reproduce approach to a standardise-then-reproduce-twice approach. Our overall (negative) finding that the great majority of human evaluations in NLP is not repeatable and/or not reproducible and/or too flawed to justify reproduction, paints a dire picture, but presents an opportunity for a rethink about how to design and report human evaluations in NLP.

Huiyuan Lai, Antonio Toral, Malvina Nissim

Style transfer aims to rewrite a source text in a different target style while preserving its content. We propose a novel approach to this task that leverages generic resources, and without using any task-specific parallel (source-target) data outperforms existing unsupervised approaches on the two most popular style transfer tasks: formality transfer and polarity swap. In practice, we adopt a multi-step procedure which builds on a generic pre-trained sequence-to-sequence model (BART). First, we strengthen the model's ability to rewrite by further pre-training BART on both an existing collection of generic paraphrases, as well as on synthetic pairs created using a general-purpose lexical resource. Second, through an iterative back-translation approach, we train two models, each in a transfer direction, so that they can provide each other with synthetically generated pairs, dynamically in the training process. Lastly, we let our best reresulting model generate static synthetic pairs to be used in a supervised training regime. Besides methodology and state-of-the-art results, a core contribution of this work is a reflection on the nature of the two tasks we address, and how their differences are highlighted by their response to our approach.

Huiyuan Lai, Antonio Toral, Malvina Nissim

Scarcity of parallel data causes formality style transfer models to have scarce success in preserving content. We show that fine-tuning pre-trained language (GPT-2) and sequence-to-sequence (BART) models boosts content preservation, and that this is possible even with limited amounts of parallel data. Augmenting these models with rewards that target style and content -- the two core aspects of the task -- we achieve a new state-of-the-art.

Wietse de Vries, Martijn Bartelds, Malvina Nissim et al.

For many (minority) languages, the resources needed to train large models are not available. We investigate the performance of zero-shot transfer learning with as little data as possible, and the influence of language similarity in this process. We retrain the lexical layers of four BERT-based models using data from two low-resource target language varieties, while the Transformer layers are independently fine-tuned on a POS-tagging task in the model's source language. By combining the new lexical layers and fine-tuned Transformer layers, we achieve high task performance for both target languages. With high language similarity, 10MB of data appears sufficient to achieve substantial monolingual transfer performance. Monolingual BERT-based models generally achieve higher downstream task performance after retraining the lexical layer than multilingual BERT, even when the target language is included in the multilingual model.

Ludovica Pannitto, Lucia Busso, Claudia Roberta Combei et al.

Although Natural Language Processing (NLP) is at the core of many tools young people use in their everyday life, high school curricula (in Italy) do not include any computational linguistics education. This lack of exposure makes the use of such tools less responsible than it could be and makes choosing computational linguistics as a university degree unlikely. To raise awareness, curiosity, and longer-term interest in young people, we have developed an interactive workshop designed to illustrate the basic principles of NLP and computational linguistics to high school Italian students aged between 13 and 18 years. The workshop takes the form of a game in which participants play the role of machines needing to solve some of the most common problems a computer faces in understanding language: from voice recognition to Markov chains to syntactic parsing. Participants are guided through the workshop with the help of instructors, who present the activities and explain core concepts from computational linguistics. The workshop was presented at numerous outlets in Italy between 2019 and 2021, both face-to-face and online.

Lucio Messina, Lucia Busso, Claudia Roberta Combei et al.

We describe and make available the game-based material developed for a laboratory run at several Italian science festivals to popularize NLP among young students.

Lorenzo De Mattei, Michele Cafagna, Huiyuan Lai et al.

An ongoing debate in the NLG community concerns the best way to evaluate systems, with human evaluation often being considered the most reliable method, compared to corpus-based metrics. However, tasks involving subtle textual differences, such as style transfer, tend to be hard for humans to perform. In this paper, we propose an evaluation method for this task based on purposely-trained classifiers, showing that it better reflects system differences than traditional metrics such as BLEU and ROUGE.

Wietse de Vries, Malvina Nissim

Large generative language models have been very successful for English, but other languages lag behind, in part due to data and computational limitations. We propose a method that may overcome these problems by adapting existing pre-trained models to new languages. Specifically, we describe the adaptation of English GPT-2 to Italian and Dutch by retraining lexical embeddings without tuning the Transformer layers. As a result, we obtain lexical embeddings for Italian and Dutch that are aligned with the original English lexical embeddings. Additionally, we scale up complexity by transforming relearned lexical embeddings of GPT-2 small to the GPT-2 medium embedding space. This method minimises the amount of training and prevents losing information during adaptation that was learned by GPT-2. English GPT-2 models with relearned lexical embeddings can generate realistic sentences in Italian and Dutch. Though on average these sentences are still identifiable as artificial by humans, they are assessed on par with sentences generated by a GPT-2 model fully trained from scratch.

Gaetana Ruggiero, Albert Gatt, Malvina Nissim

Existing research on Authorship Attribution (AA) focuses on texts for which a lot of data is available (e.g novels), mainly in English. We approach AA via Authorship Verification on short Italian texts in two novel datasets, and analyze the interaction between genre, topic, gender and length. Results show that AV is feasible even with little data, but more evidence helps. Gender and topic can be indicative clues, and if not controlled for, they might overtake more specific aspects of personal style.

Elisa Bassignana, Malvina Nissim, Viviana Patti

As a contribution to personality detection in languages other than English, we rely on distant supervision to create Personal-ITY, a novel corpus of YouTube comments in Italian, where authors are labelled with personality traits. The traits are derived from one of the mainstream personality theories in psychology research, named MBTI. Using personality prediction experiments, we (i) study the task of personality prediction in itself on our corpus as well as on TwiSty, a Twitter dataset also annotated with MBTI labels; (ii) carry out an extensive, in-depth analysis of the features used by the classifier, and view them specifically under the light of the original theory that we used to create the corpus in the first place. We observe that no single model is best at personality detection, and that while some traits are easier than others to detect, and also to match back to theory, for other, less frequent traits the picture is much more blurred.

Elisa Bassignana, Malvina Nissim, Viviana Patti

We present a novel corpus for personality prediction in Italian, containing a larger number of authors and a different genre compared to previously available resources. The corpus is built exploiting Distant Supervision, assigning Myers-Briggs Type Indicator (MBTI) labels to YouTube comments, and can lend itself to a variety of experiments. We report on preliminary experiments on Personal-ITY, which can serve as a baseline for future work, showing that some types are easier to predict than others, and discussing the perks of cross-dataset prediction.

Marion Bartl, Malvina Nissim, Albert Gatt

Contextualized word embeddings have been replacing standard embeddings as the representational knowledge source of choice in NLP systems. Since a variety of biases have previously been found in standard word embeddings, it is crucial to assess biases encoded in their replacements as well. Focusing on BERT (Devlin et al., 2018), we measure gender bias by studying associations between gender-denoting target words and names of professions in English and German, comparing the findings with real-world workforce statistics. We mitigate bias by fine-tuning BERT on the GAP corpus (Webster et al., 2018), after applying Counterfactual Data Substitution (CDS) (Maudslay et al., 2019). We show that our method of measuring bias is appropriate for languages such as English, but not for languages with a rich morphology and gender-marking, such as German. Our results highlight the importance of investigating bias and mitigation techniques cross-linguistically, especially in view of the current emphasis on large-scale, multilingual language models.

Lorenzo De Mattei, Michele Cafagna, Felice Dell'Orletta et al.

In the last few years, pre-trained neural architectures have provided impressive improvements across several NLP tasks. Still, generative language models are available mainly for English. We develop GePpeTto, the first generative language model for Italian, built using the GPT-2 architecture. We provide a thorough analysis of GePpeTto's quality by means of both an automatic and a human-based evaluation. The automatic assessment consists in (i) calculating perplexity across different genres and (ii) a profiling analysis over GePpeTto's writing characteristics. We find that GePpeTto's production is a sort of bonsai version of human production, with shorter but yet complex sentences. Human evaluation is performed over a sentence completion task, where GePpeTto's output is judged as natural more often than not, and much closer to the original human texts than to a simpler language model which we take as baseline.

Wietse de Vries, Andreas van Cranenburgh, Malvina Nissim

Peeking into the inner workings of BERT has shown that its layers resemble the classical NLP pipeline, with progressively more complex tasks being concentrated in later layers. To investigate to what extent these results also hold for a language other than English, we probe a Dutch BERT-based model and the multilingual BERT model for Dutch NLP tasks. In addition, through a deeper analysis of part-of-speech tagging, we show that also within a given task, information is spread over different parts of the network and the pipeline might not be as neat as it seems. Each layer has different specialisations, so that it may be more useful to combine information from different layers, instead of selecting a single one based on the best overall performance.