Arianna Bisazza

Vera Neplenbroek, Gabriele Sarti, Arianna Bisazza et al.

When large language models (LLMs) are used in high-stakes scenarios, such as legal, medical and financial advice, even a single conversation history is enough to drive differences in outcomes between users. Prior work has demonstrated that this results in outcome disparities between sociodemographic groups, with some groups receiving more advantageous outcomes than others. In this work, we demonstrate that LLMs actually struggle to infer user sociodemographics from a single conversation history and that although there are disparities between sociodemographic groups, they are minimal in magnitude. To investigate what the main driver of these disparities is, we compare user sociodemographics to a range of (psycho)linguistic features of conversations, including conversation topic, emotions, and readability. We find that conversation topics are most predictive of LLM-generated advice within a conversational context, which, to some extent, function as proxies for sociodemographic groups and often affect advice in unpredictable ways. This is cause for concern and highlights the need for future research to better understand and, if needed, mitigate the effect of conversational context on LLM outputs in high-stakes scenarios.

Jaap Jumelet, Abdellah Fourtassi, Akari Haga et al. · mila

We present BabyBabelLM, a multilingual collection of datasets modeling the language a person observes from birth until they acquire a native language. We curate developmentally plausible pretraining data aiming to cover the equivalent of 100M English words of content in each of 45 languages. We compile evaluation suites and train baseline models in each language. BabyBabelLM aims to facilitate multilingual pretraining and cognitive modeling.

Ahmet Üstün, Arianna Bisazza, Gosse Bouma et al.

Massively multilingual models are promising for transfer learning across tasks and languages. However, existing methods are unable to fully leverage training data when it is available in different task-language combinations. To exploit such heterogeneous supervision, we propose Hyper-X, a single hypernetwork that unifies multi-task and multilingual learning with efficient adaptation. This model generates weights for adapter modules conditioned on both tasks and language embeddings. By learning to combine task and language-specific knowledge, our model enables zero-shot transfer for unseen languages and task-language combinations. Our experiments on a diverse set of languages demonstrate that Hyper-X achieves the best or competitive gain when a mixture of multiple resources is available, while being on par with strong baselines in the standard scenario. Hyper-X is also considerably more efficient in terms of parameters and resources compared to methods that train separate adapters. Finally, Hyper-X consistently produces strong results in few-shot scenarios for new languages, showing the versatility of our approach beyond zero-shot transfer.

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.

Jirui Qi, Raquel Fernández, Arianna Bisazza

Multilingual large-scale Pretrained Language Models (PLMs) have been shown to store considerable amounts of factual knowledge, but large variations are observed across languages. With the ultimate goal of ensuring that users with different language backgrounds obtain consistent feedback from the same model, we study the cross-lingual consistency (CLC) of factual knowledge in various multilingual PLMs. To this end, we propose a Ranking-based Consistency (RankC) metric to evaluate knowledge consistency across languages independently from accuracy. Using this metric, we conduct an in-depth analysis of the determining factors for CLC, both at model level and at language-pair level. Among other results, we find that increasing model size leads to higher factual probing accuracy in most languages, but does not improve cross-lingual consistency. Finally, we conduct a case study on CLC when new factual associations are inserted in the PLMs via model editing. Results on a small sample of facts inserted in English reveal a clear pattern whereby the new piece of knowledge transfers only to languages with which English has a high RankC score.

Lukas Edman, Gabriele Sarti, Antonio Toral et al.

Pretrained character-level and byte-level language models have been shown to be competitive with popular subword models across a range of Natural Language Processing (NLP) tasks. However, there has been little research on their effectiveness for neural machine translation (NMT), particularly within the popular pretrain-then-finetune paradigm. This work performs an extensive comparison across multiple languages and experimental conditions of character- and subword-level pretrained models (ByT5 and mT5, respectively) on NMT. We show the effectiveness of character-level modeling in translation, particularly in cases where fine-tuning data is limited. In our analysis, we show how character models' gains in translation quality are reflected in better translations of orthographically similar words and rare words. While evaluating the importance of source texts in driving model predictions, we highlight word-level patterns within ByT5, suggesting an ability to modulate word-level and character-level information during generation. We conclude by assessing the efficiency tradeoff of byte models, suggesting their usage in non-time-critical scenarios to boost translation quality.

Francesca Padovani, Xiulin Yang, Bastian Bunzeck et al.

CHILDES is a paramount resource for language acquisition studies -- yet computational tools for analyzing its syntactic structure remain limited. Leveraging the recent release of the UD-English-CHILDES treebank with gold-standard Universal Dependencies (UD) annotations, we train a state-of-the-art dependency parser specifically tailored to CHILDES. The parser more accurately captures syntactic patterns in child--adult interactions, outperforming widely used off-the-shelf English parsers, including SpaCy and Stanza. Alongside the parser, we also release a Part-of-Speech tagger and an utterance-level construction tagger, which together form the open-source Syntactic Parsing Toolkit for Child--Adult InTeractions (CAIT). Through a detailed error analysis and a case study tracking the distribution of syntactic constructions across developmental time in CHILDES, we demonstrate the practical utility of the toolkit for large-scale, reproducible research on language acquisition.

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.

Yuchen Lian, Arianna Bisazza, Tessa Verhoef

Artificial learners often behave differently from human learners in the context of neural agent-based simulations of language emergence and change. A common explanation is the lack of appropriate cognitive biases in these learners. However, it has also been proposed that more naturalistic settings of language learning and use could lead to more human-like results. We investigate this latter account focusing on the word-order/case-marking trade-off, a widely attested language universal that has proven particularly hard to simulate. We propose a new Neural-agent Language Learning and Communication framework (NeLLCom) where pairs of speaking and listening agents first learn a miniature language via supervised learning, and then optimize it for communication via reinforcement learning. Following closely the setup of earlier human experiments, we succeed in replicating the trade-off with the new framework without hard-coding specific biases in the agents. We see this as an essential step towards the investigation of language universals with neural learners.

Gabriele Sarti, Arianna Bisazza, Ana Guerberof Arenas et al.

We introduce DivEMT, the first publicly available post-editing study of Neural Machine Translation (NMT) over a typologically diverse set of target languages. Using a strictly controlled setup, 18 professional translators were instructed to translate or post-edit the same set of English documents into Arabic, Dutch, Italian, Turkish, Ukrainian, and Vietnamese. During the process, their edits, keystrokes, editing times and pauses were recorded, enabling an in-depth, cross-lingual evaluation of NMT quality and post-editing effectiveness. Using this new dataset, we assess the impact of two state-of-the-art NMT systems, Google Translate and the multilingual mBART-50 model, on translation productivity. We find that post-editing is consistently faster than translation from scratch. However, the magnitude of productivity gains varies widely across systems and languages, highlighting major disparities in post-editing effectiveness for languages at different degrees of typological relatedness to English, even when controlling for system architecture and training data size. We publicly release the complete dataset including all collected behavioral data, to foster new research on the translation capabilities of NMT systems for typologically diverse languages.

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.

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.

Yuchen Lian, Tessa Verhoef, Arianna Bisazza

Recent advances in computational linguistics include simulating the emergence of human-like languages with interacting neural network agents, starting from sets of random symbols. The recently introduced NeLLCom framework (Lian et al., 2023) allows agents to first learn an artificial language and then use it to communicate, with the aim of studying the emergence of specific linguistics properties. We extend this framework (NeLLCom-X) by introducing more realistic role-alternating agents and group communication in order to investigate the interplay between language learnability, communication pressures, and group size effects. We validate NeLLCom-X by replicating key findings from prior research simulating the emergence of a word-order/case-marking trade-off. Next, we investigate how interaction affects linguistic convergence and emergence of the trade-off. The novel framework facilitates future simulations of diverse linguistic aspects, emphasizing the importance of interaction and group dynamics in language evolution.

Francesca Padovani, Jaap Jumelet, Yevgen Matusevych et al.

Is child-directed language (CDL) optimized to support language learning, and which aspects of linguistic development does it facilitate? We investigate this question using neural language models trained on CDL versus adult-directed language (ADL). We selectively remove syntactic or lexical co-occurrence information from the model training data, and evaluate the impact of these manipulations on verb meaning acquisition. While disrupting syntax impairs learning across all datasets, models trained on CDL and spoken ADL show significantly higher resilience than those trained on written input. Tracking semantic and syntactic performance over training, we observe a semantic-first trajectory, with verb meanings emerging prior to robust syntactic proficiency, an asynchrony most pronounced in the spoken domain, especially CDL. These results suggest that the advantage for verb learning previously attributed to CDL may instead reflect broader properties of the spoken register, rather than a uniquely CDL-specific optimization.

Tianyu Liu, Jirui Qi, Mrinmaya Sachan et al.

Large language models are known to often exhibit inconsistent knowledge. This is particularly problematic in multilingual scenarios, where models are likely to be asked similar questions in different languages, and inconsistent responses can undermine their reliability. In this work, we show that this issue can be mitigated using reinforcement learning with a structured reward function, which leads to an optimal policy with consistent crosslingual responses. We introduce Direct Consistency Optimization (DCO), a DPO-inspired method that requires no explicit reward model and is derived directly from the LLM itself. Comprehensive experiments show that DCO significantly improves crosslingual consistency across diverse LLMs and outperforms existing methods when training with samples of multiple languages, while complementing DPO when gold labels are available. Extra experiments demonstrate the effectiveness of DCO in bilingual settings, significant out-of-domain generalizability, and controllable alignment via direction hyperparameters. Taken together, these results establish DCO as a robust and efficient solution for improving knowledge consistency across languages in multilingual LLMs. All code, training scripts, and evaluation benchmarks are released at https://github.com/Betswish/ConsistencyRL.

Vitalii Hirak, Jaap Jumelet, Arianna Bisazza

Despite major advances in multilingual modeling, large quality disparities persist across languages. Besides the obvious impact of uneven training resources, typological properties have also been proposed to determine the intrinsic difficulty of modeling a language. The existing evidence, however, is mostly based on small monolingual language models or bilingual translation models trained from scratch. We expand on this line of work by analyzing two large pre-trained multilingual translation models, NLLB-200 and Tower+, which are state-of-the-art representatives of encoder-decoder and decoder-only machine translation, respectively. Based on a broad set of languages, we find that target language typology drives translation quality of both models, even after controlling for more trivial factors, such as data resourcedness and writing script. Additionally, languages with certain typological properties benefit more from a wider search of the output space, suggesting that such languages could profit from alternative decoding strategies beyond the standard left-to-right beam search. To facilitate further research in this area, we release a set of fine-grained typological properties for 212 languages of the FLORES+ MT evaluation benchmark.

Jirui Qi, Shan Chen, Zidi Xiong et al.

Recent Large Reasoning Models (LRMs) with thinking traces have shown strong performance on English reasoning tasks. However, their ability to think in other languages is less studied. This capability is as important as answer accuracy for real world applications because users may find the reasoning trace useful for oversight only when it is expressed in their own language. We comprehensively evaluate two leading families of LRMs on our XReasoning benchmark and find that even the most advanced models often revert to English or produce fragmented reasoning in other languages, revealing a substantial gap in multilingual reasoning. Prompt based interventions that force models to reason in the users language improve readability and oversight but reduce answer accuracy, exposing an important trade off. We further show that targeted post training on just 100 examples mitigates this mismatch, though some accuracy loss remains. Our results highlight the limited multilingual reasoning capabilities of current LRMs and outline directions for future work. Code and data are available at https://github.com/Betswish/mCoT-XReasoning.

Xinyi Chen, Baohao Liao, Jirui Qi et al.

Following multiple instructions is a crucial ability for large language models (LLMs). Evaluating this ability comes with significant challenges: (i) limited coherence between multiple instructions, (ii) positional bias where the order of instructions affects model performance, and (iii) a lack of objectively verifiable tasks. To address these issues, we introduce a benchmark designed to evaluate models' abilities to follow multiple instructions through sequential instruction following (SIFo) tasks. In SIFo, the successful completion of multiple instructions is verifiable by examining only the final instruction. Our benchmark evaluates instruction following using four tasks (text modification, question answering, mathematics, and security rules), each assessing different aspects of sequential instruction following. Our evaluation of popular LLMs, both closed-source and open-source, shows that more recent and larger models significantly outperform their older and smaller counterparts on the SIFo tasks, validating the benchmark's effectiveness. All models struggle with following sequences of instructions, hinting at an important lack of robustness of today's language models.

Vera Neplenbroek, Arianna Bisazza, Raquel Fernández

Generative large language models (LLMs) have been shown to exhibit harmful biases and stereotypes. While safety fine-tuning typically takes place in English, if at all, these models are being used by speakers of many different languages. There is existing evidence that the performance of these models is inconsistent across languages and that they discriminate based on demographic factors of the user. Motivated by this, we investigate whether the social stereotypes exhibited by LLMs differ as a function of the language used to prompt them, while controlling for cultural differences and task accuracy. To this end, we present MBBQ (Multilingual Bias Benchmark for Question-answering), a carefully curated version of the English BBQ dataset extended to Dutch, Spanish, and Turkish, which measures stereotypes commonly held across these languages. We further complement MBBQ with a parallel control dataset to measure task performance on the question-answering task independently of bias. Our results based on several open-source and proprietary LLMs confirm that some non-English languages suffer from bias more than English, even when controlling for cultural shifts. Moreover, we observe significant cross-lingual differences in bias behaviour for all except the most accurate models. With the release of MBBQ, we hope to encourage further research on bias in multilingual settings. The dataset and code are available at https://github.com/Veranep/MBBQ.

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.

Javier Ferrando, Gabriele Sarti, Arianna Bisazza et al.

The rapid progress of research aimed at interpreting the inner workings of advanced language models has highlighted a need for contextualizing the insights gained from years of work in this area. This primer provides a concise technical introduction to the current techniques used to interpret the inner workings of Transformer-based language models, focusing on the generative decoder-only architecture. We conclude by presenting a comprehensive overview of the known internal mechanisms implemented by these models, uncovering connections across popular approaches and active research directions in this area.

Xiulin Yang, Arianna Bisazza, Nathan Schneider et al.

How can children acquire native-level syntax from limited input? According to the Poverty of the Stimulus Hypothesis (PoSH), the linguistic input children receive is insufficient to explain certain generalizations that are robustly learned; innate linguistic constraints, many have argued, are thus necessary to explain language learning. Neural language models, which lack such language-specific constraints in their design, offer a computational test of this longstanding (but controversial) claim. We introduce \poshbench, a training-and-evaluation suite targeting question formation, islands to movement, and other English phenomena at the center of the PoSH arguments. Training Transformer models on 10--50M words of developmentally plausible text, we find indications of generalization on all phenomena even without direct positive evidence -- yet neural models remain less data-efficient and their generalizations are weaker than those of children. We further enhance our models with three recently proposed cognitively motivated inductive biases. We find these biases improve general syntactic competence but not \poshbench performance. Our findings challenge the claim that innate syntax is the only possible route to generalization, while suggesting that human-like data efficiency requires inductive biases beyond those tested here.

Jaap Jumelet, Leonie Weissweiler, Joakim Nivre et al.

We introduce MultiBLiMP 1.0, a massively multilingual benchmark of linguistic minimal pairs, covering 101 languages and 2 types of subject-verb agreement, containing more than 128,000 minimal pairs. Our minimal pairs are created using a fully automated pipeline, leveraging the large-scale linguistic resources of Universal Dependencies and UniMorph. MultiBLiMP 1.0 evaluates abilities of LLMs at an unprecedented multilingual scale, and highlights the shortcomings of the current state-of-the-art in modelling low-resource languages.

Yuqing Zhang, Ecesu Ürker, Tessa Verhoef et al.

Modeling the emergence of human-like lexicons in computational systems has advanced through the use of interacting neural agents, which simulate both learning and communicative pressures. The NeLLCom-Lex framework (Zhang et al., 2025) allows neural agents to develop pragmatic color naming behavior and human-like lexicons through supervised learning (SL) from human data and reinforcement learning (RL) in referential games. Despite these successes, the lexicons that emerge diverge systematically from human color categories, producing highly non-convex regions in color space, which contrast with the convexity typical of human categories. To address this, we introduce two factors, upsampling rare color terms during SL and multi-listener RL interactions, and adopt a convexity measure to quantify geometric coherence. We find that upsampling improves lexical diversity and system-level informativeness of the color lexicon, while many-listener setups promote more convex color categories. The combination of moderate upsampling and multiple listeners produces lexicons most similar to human systems.

Gaofei Shen, Michaela Watkins, Afra Alishahi et al.

Interpretability research has shown that self-supervised Spoken Language Models (SLMs) encode a wide variety of features in human speech from the acoustic, phonetic, phonological, syntactic and semantic levels, to speaker characteristics. The bulk of prior research on representations of phonology has focused on segmental features such as phonemes; the encoding of suprasegmental phonology (such as tone and stress patterns) in SLMs is not yet well understood. Tone is a suprasegmental feature that is present in more than half of the world's languages. This paper aims to analyze the tone encoding capabilities of SLMs, using Mandarin and Vietnamese as case studies. We show that SLMs encode lexical tone to a significant degree even when they are trained on data from non-tonal languages. We further find that SLMs behave similarly to native and non-native human participants in tone and consonant perception studies, but they do not follow the same developmental trajectory.

Ezgi Başar, Francesca Padovani, Jaap Jumelet et al.

We introduce TurBLiMP, the first Turkish benchmark of linguistic minimal pairs, designed to evaluate the linguistic abilities of monolingual and multilingual language models (LMs). Covering 16 linguistic phenomena with 1000 minimal pairs each, TurBLiMP fills an important gap in linguistic evaluation resources for Turkish. In designing the benchmark, we give extra attention to two properties of Turkish that remain understudied in current syntactic evaluations of LMs, namely word order flexibility and subordination through morphological processes. Our experiments on a wide range of LMs and a newly collected set of human acceptability judgments reveal that even cutting-edge Large LMs still struggle with grammatical phenomena that are not challenging for humans, and may also exhibit different sensitivities to word order and morphological complexity compared to humans.

Francesca Padovani, Jaap Jumelet, Yevgen Matusevych et al.

Seminal work by Huebner et al. (2021) showed that language models (LMs) trained on English Child-Directed Language (CDL) can reach similar syntactic abilities as LMs trained on much larger amounts of adult-directed written text, suggesting that CDL could provide more effective LM training material than the commonly used internet-crawled data. However, the generalizability of these results across languages, model types, and evaluation settings remains unclear. We test this by comparing models trained on CDL vs. Wikipedia across two LM objectives (masked and causal), three languages (English, French, German), and three syntactic minimal-pair benchmarks. Our results on these benchmarks show inconsistent benefits of CDL, which in most cases is outperformed by Wikipedia models. We then identify various shortcomings in previous benchmarks, and introduce a novel testing methodology, FIT-CLAMS, which uses a frequency-controlled design to enable balanced comparisons across training corpora. Through minimal pair evaluations and regression analysis we show that training on CDL does not yield stronger generalizations for acquiring syntax and highlight the importance of controlling for frequency effects when evaluating syntactic ability.

Akari Haga, Akiyo Fukatsu, Miyu Oba et al.

While current large language models have achieved a remarkable success, their data efficiency remains a challenge to overcome. Recently it has been suggested that child-directed speech (CDS) can improve training data efficiency of modern language models based on Transformer neural networks. However, it is not yet understood which specific properties of CDS are effective for training these models. In the context of the BabyLM Challenge, we focus on Variation Sets (VSs), sets of consecutive utterances expressing a similar intent with slightly different words and structures, which are ubiquitous in CDS. To assess the impact of VSs on training data efficiency, we augment CDS data with different proportions of artificial VSs and use these datasets to train an auto-regressive model, GPT-2. We find that the best proportion of VSs depends on the evaluation benchmark: BLiMP and GLUE scores benefit from the presence of VSs, but EWOK scores do not. Additionally, the results vary depending on multiple factors such as the number of epochs and the order of utterance presentation. Taken together, these findings suggest that VSs can have a beneficial influence on language models, while leaving room for further investigation.

Yuchen Lian, Arianna Bisazza, Tessa Verhoef

Differential Case Marking (DCM) refers to the phenomenon where grammatical case marking is applied selectively based on semantic, pragmatic, or other factors. The emergence of DCM has been studied in artificial language learning experiments with human participants, which were specifically aimed at disentangling the effects of learning from those of communication (Smith & Culbertson, 2020). Multi-agent reinforcement learning frameworks based on neural networks have gained significant interest to simulate the emergence of human-like linguistic phenomena. In this study, we employ such a framework in which agents first acquire an artificial language before engaging in communicative interactions, enabling direct comparisons to human result. Using a very generic communication optimization algorithm and neural-network learners that have no prior experience with language or semantic preferences, our results demonstrate that learning alone does not lead to DCM, but when agents communicate, differential use of markers arises. This supports Smith and Culbertson (2020)'s findings that highlight the critical role of communication in shaping DCM and showcases the potential of neural-agent models to complement experimental research on language evolution.

Vera Neplenbroek, Arianna Bisazza, Raquel Fernández

Recent generative large language models (LLMs) show remarkable performance in non-English languages, but when prompted in those languages they tend to express higher harmful social biases and toxicity levels. Prior work has shown that finetuning on specialized datasets can mitigate this behavior, and doing so in English can transfer to other languages. In this work, we investigate the impact of different finetuning methods on the model's bias and toxicity, but also on its ability to produce fluent and diverse text. We reduce biases by finetuning on curated non-harmful text, but find only direct preference optimization to be effective for mitigating toxicity. The mitigation caused by applying these methods in English also transfers to non-English languages. We find evidence that the extent to which transfer takes place can be predicted by the amount of data in a given language present in the model's pretraining data. However, this transfer of bias and toxicity mitigation often comes at the expense of decreased language generation ability in non-English languages, highlighting the importance of developing language-specific bias and toxicity mitigation methods.

Tianyu Liu, Jirui Qi, Paul He et al.

Recent work suggests that large language models enhanced with retrieval-augmented generation are easily influenced by the order, in which the retrieved documents are presented to the model when solving tasks such as question answering (QA). However, there is no method to date that exploits this phenomenon to improve generation. We fill this gap. In this study, we show that the pointwise mutual information between a context and a question is an effective gauge for language model performance. Importantly, this gauge does not depend on knowing the answer to the question a priori. Through experiments on two question-answering datasets and a variety of large language models, we find evidence for an empirical correlation between answer accuracy and pointwise mutual information. Additionally, we propose two methods that use the pointwise mutual information between a document and a question as a gauge for selecting and constructing prompts that lead to better performance, whose effectiveness we demonstrate through experimentation.

Jirui Qi, Raquel Fernández, Arianna Bisazza

Retrieval-augmented generation (RAG) with large language models (LLMs) has demonstrated strong performance in multilingual question-answering (QA) tasks by leveraging relevant passages retrieved from corpora. In multilingual RAG (mRAG), the retrieved passages can be written in languages other than that of the query entered by the user, making it challenging for LLMs to effectively utilize the provided information. Recent research suggests that retrieving passages from multilingual corpora can improve RAG performance, particularly for low-resource languages. However, the extent to which LLMs can leverage different kinds of multilingual contexts to generate accurate answers, *independently from retrieval quality*, remains understudied. In this paper, we conduct an extensive assessment of LLMs' ability to (i) make consistent use of a relevant passage regardless of its language, (ii) respond in the expected language, and (iii) focus on the relevant passage even when multiple `distracting' passages in different languages are provided in the context. Our experiments with four LLMs across three QA datasets covering a total of 48 languages reveal a surprising ability of LLMs to extract the relevant information from passages in a different language than the query, but a much weaker ability to formulate a full answer in the correct language. Our analysis, based on both accuracy and feature attribution techniques, further shows that distracting passages negatively impact answer quality regardless of their language. However, distractors in the query language exert a slightly stronger influence. Taken together, our findings deepen the understanding of how LLMs utilize context in mRAG systems, providing directions for future improvements.

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.

Francesca Padovani, Bastian Bunzeck, Manar Ali et al.

We investigate whether pre-training exclusively on dialogue data results in formally and functionally apt small language models. Based on this pre-trained llamalogue model, we employ a variety of fine-tuning strategies to enforce "more communicative" text generations by our models. Although our models underperform on most standard BabyLM benchmarks, they excel at dialogue continuation prediction in a minimal pair setting. While PPO fine-tuning has mixed to adversarial effects on our models, DPO fine-tuning further improves their performance on our custom dialogue benchmark.

Yuqing Zhang, Ecesu Ürker, Tessa Verhoef et al.

Lexical semantic change has primarily been investigated with observational and experimental methods; however, observational methods (corpus analysis, distributional semantic modeling) cannot get at causal mechanisms, and experimental paradigms with humans are hard to apply to semantic change due to the extended diachronic processes involved. This work introduces NeLLCom-Lex, a neural-agent framework designed to simulate semantic change by first grounding agents in a real lexical system (e.g. English) and then systematically manipulating their communicative needs. Using a well-established color naming task, we simulate the evolution of a lexical system within a single generation, and study which factors lead agents to: (i) develop human-like naming behavior and lexicons, and (ii) change their behavior and lexicons according to their communicative needs. Our experiments with different supervised and reinforcement learning pipelines show that neural agents trained to 'speak' an existing language can reproduce human-like patterns in color naming to a remarkable extent, supporting the further use of NeLLCom-Lex to elucidate the mechanisms of semantic change.

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.

Vera Neplenbroek, Arianna Bisazza, Raquel Fernández

Generative Large Language Models (LLMs) infer user's demographic information from subtle cues in the conversation -- a phenomenon called implicit personalization. Prior work has shown that such inferences can lead to lower quality responses for users assumed to be from minority groups, even when no demographic information is explicitly provided. In this work, we systematically explore how LLMs respond to stereotypical cues using controlled synthetic conversations, by analyzing the models' latent user representations through both model internals and generated answers to targeted user questions. Our findings reveal that LLMs do infer demographic attributes based on these stereotypical signals, which for a number of groups even persists when the user explicitly identifies with a different demographic group. Finally, we show that this form of stereotype-driven implicit personalization can be effectively mitigated by intervening on the model's internal representations using a trained linear probe to steer them toward the explicitly stated identity. Our results highlight the need for greater transparency and control in how LLMs represent user identity.

Gaofei Shen, Hosein Mohebbi, Arianna Bisazza et al.

As the capabilities of large-scale pre-trained models evolve, understanding the determinants of their outputs becomes more important. Feature attribution aims to reveal which parts of the input elements contribute the most to model outputs. In speech processing, the unique characteristics of the input signal make the application of feature attribution methods challenging. We study how factors such as input type and aggregation and perturbation timespan impact the reliability of standard feature attribution methods, and how these factors interact with characteristics of each classification task. We find that standard approaches to feature attribution are generally unreliable when applied to the speech domain, with the exception of word-aligned perturbation methods when applied to word-based classification tasks.

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.

Jirui Qi, Gabriele Sarti, Raquel Fernández et al.

Ensuring the verifiability of model answers is a fundamental challenge for retrieval-augmented generation (RAG) in the question answering (QA) domain. Recently, self-citation prompting was proposed to make large language models (LLMs) generate citations to supporting documents along with their answers. However, self-citing LLMs often struggle to match the required format, refer to non-existent sources, and fail to faithfully reflect LLMs' context usage throughout the generation. In this work, we present MIRAGE --Model Internals-based RAG Explanations -- a plug-and-play approach using model internals for faithful answer attribution in RAG applications. MIRAGE detects context-sensitive answer tokens and pairs them with retrieved documents contributing to their prediction via saliency methods. We evaluate our proposed approach on a multilingual extractive QA dataset, finding high agreement with human answer attribution. On open-ended QA, MIRAGE achieves citation quality and efficiency comparable to self-citation while also allowing for a finer-grained control of attribution parameters. Our qualitative evaluation highlights the faithfulness of MIRAGE's attributions and underscores the promising application of model internals for RAG answer attribution.

Gaofei Shen, Afra Alishahi, Arianna Bisazza et al.

Understanding which information is encoded in deep models of spoken and written language has been the focus of much research in recent years, as it is crucial for debugging and improving these architectures. Most previous work has focused on probing for speaker characteristics, acoustic and phonological information in models of spoken language, and for syntactic information in models of written language. Here we focus on the encoding of syntax in several self-supervised and visually grounded models of spoken language. We employ two complementary probing methods, combined with baselines and reference representations to quantify the degree to which syntactic structure is encoded in the activations of the target models. We show that syntax is captured most prominently in the middle layers of the networks, and more explicitly within models with more parameters.

Arianna Bisazza, Ahmet Üstün, Stephan Sportel

Identifying factors that make certain languages harder to model than others is essential to reach language equality in future Natural Language Processing technologies. Free-order case-marking languages, such as Russian, Latin or Tamil, have proved more challenging than fixed-order languages for the tasks of syntactic parsing and subject-verb agreement prediction. In this work, we investigate whether this class of languages is also more difficult to translate by state-of-the-art Neural Machine Translation models (NMT). Using a variety of synthetic languages and a newly introduced translation challenge set, we find that word order flexibility in the source language only leads to a very small loss of NMT quality, even though the core verb arguments become impossible to disambiguate in sentences without semantic cues. The latter issue is indeed solved by the addition of case marking. However, in medium- and low-resource settings, the overall NMT quality of fixed-order languages remains unmatched.

Yuchen Lian, Arianna Bisazza, Tessa Verhoef

Natural languages display a trade-off among different strategies to convey syntactic structure, such as word order or inflection. This trade-off, however, has not appeared in recent simulations of iterated language learning with neural network agents (Chaabouni et al., 2019b). We re-evaluate this result in light of three factors that play an important role in comparable experiments from the Language Evolution field: (i) speaker bias towards efficient messaging, (ii) non systematic input languages, and (iii) learning bottleneck. Our simulations show that neural agents mainly strive to maintain the utterance type distribution observed during learning, instead of developing a more efficient or systematic language.

Ahmet Üstün, Arianna Bisazza, Gosse Bouma et al.

Recent advances in multilingual dependency parsing have brought the idea of a truly universal parser closer to reality. However, cross-language interference and restrained model capacity remain major obstacles. To address this, we propose a novel multilingual task adaptation approach based on contextual parameter generation and adapter modules. This approach enables to learn adapters via language embeddings while sharing model parameters across languages. It also allows for an easy but effective integration of existing linguistic typology features into the parsing network. The resulting parser, UDapter, outperforms strong monolingual and multilingual baselines on the majority of both high-resource and low-resource (zero-shot) languages, showing the success of the proposed adaptation approach. Our in-depth analyses show that soft parameter sharing via typological features is key to this success.

Prajit Dhar, Arianna Bisazza

It is now established that modern neural language models can be successfully trained on multiple languages simultaneously without changes to the underlying architecture. But what kind of knowledge is really shared among languages within these models? Does multilingual training mostly lead to an alignment of the lexical representation spaces or does it also enable the sharing of purely grammatical knowledge? In this paper we dissect different forms of cross-lingual transfer and look for its most determining factors, using a variety of models and probing tasks. We find that exposing our LMs to a related language does not always increase grammatical knowledge in the target language, and that optimal conditions for lexical-semantic transfer may not be optimal for syntactic transfer.

Ke Tran, Arianna Bisazza

We investigate whether off-the-shelf deep bidirectional sentence representations trained on a massively multilingual corpus (multilingual BERT) enable the development of an unsupervised universal dependency parser. This approach only leverages a mix of monolingual corpora in many languages and does not require any translation data making it applicable to low-resource languages. In our experiments we outperform the best CoNLL 2018 language-specific systems in all of the shared task's six truly low-resource languages while using a single system. However, we also find that (i) parsing accuracy still varies dramatically when changing the training languages and (ii) in some target languages zero-shot transfer fails under all tested conditions, raising concerns on the 'universality' of the whole approach.

Ke Tran, Arianna Bisazza, Christof Monz

Recent work has shown that recurrent neural networks (RNNs) can implicitly capture and exploit hierarchical information when trained to solve common natural language processing tasks such as language modeling (Linzen et al., 2016) and neural machine translation (Shi et al., 2016). In contrast, the ability to model structured data with non-recurrent neural networks has received little attention despite their success in many NLP tasks (Gehring et al., 2017; Vaswani et al., 2017). In this work, we compare the two architectures---recurrent versus non-recurrent---with respect to their ability to model hierarchical structure and find that recurrency is indeed important for this purpose.

Marzieh Fadaee, Arianna Bisazza, Christof Monz

Neural Machine Translation (NMT) has been widely used in recent years with significant improvements for many language pairs. Although state-of-the-art NMT systems are generating progressively better translations, idiom translation remains one of the open challenges in this field. Idioms, a category of multiword expressions, are an interesting language phenomenon where the overall meaning of the expression cannot be composed from the meanings of its parts. A first important challenge is the lack of dedicated data sets for learning and evaluating idiom translation. In this paper we address this problem by creating the first large-scale data set for idiom translation. Our data set is automatically extracted from a widely used German-English translation corpus and includes, for each language direction, a targeted evaluation set where all sentences contain idioms and a regular training corpus where sentences including idioms are marked. We release this data set and use it to perform preliminary NMT experiments as the first step towards better idiom translation.

Marlies van der Wees, Arianna Bisazza, Christof Monz

Intelligent selection of training data has proven a successful technique to simultaneously increase training efficiency and translation performance for phrase-based machine translation (PBMT). With the recent increase in popularity of neural machine translation (NMT), we explore in this paper to what extent and how NMT can also benefit from data selection. While state-of-the-art data selection (Axelrod et al., 2011) consistently performs well for PBMT, we show that gains are substantially lower for NMT. Next, we introduce dynamic data selection for NMT, a method in which we vary the selected subset of training data between different training epochs. Our experiments show that the best results are achieved when applying a technique we call gradual fine-tuning, with improvements up to +2.6 BLEU over the original data selection approach and up to +3.1 BLEU over a general baseline.

Marzieh Fadaee, Arianna Bisazza, Christof Monz

Distributed word representations are widely used for modeling words in NLP tasks. Most of the existing models generate one representation per word and do not consider different meanings of a word. We present two approaches to learn multiple topic-sensitive representations per word by using Hierarchical Dirichlet Process. We observe that by modeling topics and integrating topic distributions for each document we obtain representations that are able to distinguish between different meanings of a given word. Our models yield statistically significant improvements for the lexical substitution task indicating that commonly used single word representations, even when combined with contextual information, are insufficient for this task.