Dirk Hovy

Paul Röttger, Hannah Rose Kirk, Bertie Vidgen et al. · oxford, stanford

Without proper safeguards, large language models will readily follow malicious instructions and generate toxic content. This risk motivates safety efforts such as red-teaming and large-scale feedback learning, which aim to make models both helpful and harmless. However, there is a tension between these two objectives, since harmlessness requires models to refuse to comply with unsafe prompts, and thus not be helpful. Recent anecdotal evidence suggests that some models may have struck a poor balance, so that even clearly safe prompts are refused if they use similar language to unsafe prompts or mention sensitive topics. In this paper, we introduce a new test suite called XSTest to identify such eXaggerated Safety behaviours in a systematic way. XSTest comprises 250 safe prompts across ten prompt types that well-calibrated models should not refuse to comply with, and 200 unsafe prompts as contrasts that models, for most applications, should refuse. We describe XSTest's creation and composition, and then use the test suite to highlight systematic failure modes in state-of-the-art language models as well as more general challenges in building safer language models.

Federico Bianchi, Stefanie Anja Hills, Patricia Rossini et al. · stanford

Well-annotated data is a prerequisite for good Natural Language Processing models. Too often, though, annotation decisions are governed by optimizing time or annotator agreement. We make a case for nuanced efforts in an interdisciplinary setting for annotating offensive online speech. Detecting offensive content is rapidly becoming one of the most important real-world NLP tasks. However, most datasets use a single binary label, e.g., for hate or incivility, even though each concept is multi-faceted. This modeling choice severely limits nuanced insights, but also performance. We show that a more fine-grained multi-label approach to predicting incivility and hateful or intolerant content addresses both conceptual and performance issues. We release a novel dataset of over 40,000 tweets about immigration from the US and UK, annotated with six labels for different aspects of incivility and intolerance. Our dataset not only allows for a more nuanced understanding of harmful speech online, models trained on it also outperform or match performance on benchmark datasets.

Paul Röttger, Debora Nozza, Federico Bianchi et al. · stanford

Hate speech is a global phenomenon, but most hate speech datasets so far focus on English-language content. This hinders the development of more effective hate speech detection models in hundreds of languages spoken by billions across the world. More data is needed, but annotating hateful content is expensive, time-consuming and potentially harmful to annotators. To mitigate these issues, we explore data-efficient strategies for expanding hate speech detection into under-resourced languages. In a series of experiments with mono- and multilingual models across five non-English languages, we find that 1) a small amount of target-language fine-tuning data is needed to achieve strong performance, 2) the benefits of using more such data decrease exponentially, and 3) initial fine-tuning on readily-available English data can partially substitute target-language data and improve model generalisability. Based on these findings, we formulate actionable recommendations for hate speech detection in low-resource language settings.

Anne Lauscher, Federico Bianchi, Samuel Bowman et al. · stanford

Pre-trained language models (PLMs) have outperformed other NLP models on a wide range of tasks. Opting for a more thorough understanding of their capabilities and inner workings, researchers have established the extend to which they capture lower-level knowledge like grammaticality, and mid-level semantic knowledge like factual understanding. However, there is still little understanding of their knowledge of higher-level aspects of language. In particular, despite the importance of sociodemographic aspects in shaping our language, the questions of whether, where, and how PLMs encode these aspects, e.g., gender or age, is still unexplored. We address this research gap by probing the sociodemographic knowledge of different single-GPU PLMs on multiple English data sets via traditional classifier probing and information-theoretic minimum description length probing. Our results show that PLMs do encode these sociodemographics, and that this knowledge is sometimes spread across the layers of some of the tested PLMs. We further conduct a multilingual analysis and investigate the effect of supplementary training to further explore to what extent, where, and with what amount of pre-training data the knowledge is encoded. Our overall results indicate that sociodemographic knowledge is still a major challenge for NLP. PLMs require large amounts of pre-training data to acquire the knowledge and models that excel in general language understanding do not seem to own more knowledge about these aspects.

Giuseppe Attanasio, Debora Nozza, Federico Bianchi et al. · stanford

Language is constantly changing and evolving, leaving language models to become quickly outdated. Consequently, we should continuously update our models with new data to expose them to new events and facts. However, that requires additional computing, which means new carbon emissions. Do any measurable benefits justify this cost? This paper looks for empirical evidence to support continuous training. We reproduce existing benchmarks and extend them to include additional time periods, models, and tasks. Our results show that the downstream task performance of temporally adapted English models for social media data do not improve over time. Pretrained models without temporal adaptation are actually significantly more effective and efficient. However, we also note a lack of suitable temporal benchmarks. Our findings invite a critical reflection on when and how to temporally adapt language models, accounting for sustainability.

Tommaso Fornaciari, Dirk Hovy, Federico Bianchi · stanford

The most common ways to explore latent document dimensions are topic models and clustering methods. However, topic models have several drawbacks: e.g., they require us to choose the number of latent dimensions a priori, and the results are stochastic. Most clustering methods have the same issues and lack flexibility in various ways, such as not accounting for the influence of different topics on single documents, forcing word-descriptors to belong to a single topic (hard-clustering) or necessarily relying on word representations. We propose PROgressive SImilarity Thresholds - ProSiT, a deterministic and interpretable method, agnostic to the input format, that finds the optimal number of latent dimensions and only has two hyper-parameters, which can be set efficiently via grid search. We compare this method with a wide range of topic models and clustering methods on four benchmark data sets. In most setting, ProSiT matches or outperforms the other methods in terms six metrics of topic coherence and distinctiveness, producing replicable, deterministic results.

Matthias Orlikowski, Paul Röttger, Philipp Cimiano et al.

Many NLP tasks exhibit human label variation, where different annotators give different labels to the same texts. This variation is known to depend, at least in part, on the sociodemographics of annotators. Recent research aims to model individual annotator behaviour rather than predicting aggregated labels, and we would expect that sociodemographic information is useful for these models. On the other hand, the ecological fallacy states that aggregate group behaviour, such as the behaviour of the average female annotator, does not necessarily explain individual behaviour. To account for sociodemographics in models of individual annotator behaviour, we introduce group-specific layers to multi-annotator models. In a series of experiments for toxic content detection, we find that explicitly accounting for sociodemographic attributes in this way does not significantly improve model performance. This result shows that individual annotation behaviour depends on much more than just sociodemographics.

Giuseppe Attanasio, Debora Nozza, Dirk Hovy et al.

Natural Language Processing (NLP) models risk overfitting to specific terms in the training data, thereby reducing their performance, fairness, and generalizability. E.g., neural hate speech detection models are strongly influenced by identity terms like gay, or women, resulting in false positives, severe unintended bias, and lower performance. Most mitigation techniques use lists of identity terms or samples from the target domain during training. However, this approach requires a-priori knowledge and introduces further bias if important terms are neglected. Instead, we propose a knowledge-free Entropy-based Attention Regularization (EAR) to discourage overfitting to training-specific terms. An additional objective function penalizes tokens with low self-attention entropy. We fine-tune BERT via EAR: the resulting model matches or exceeds state-of-the-art performance for hate speech classification and bias metrics on three benchmark corpora in English and Italian. EAR also reveals overfitting terms, i.e., terms most likely to induce bias, to help identify their effect on the model, task, and predictions.

Lorenzo Lupo, Oscar Magnusson, Dirk Hovy et al.

Recent advances in large language models (LLMs) like GPT-3.5 and GPT-4 promise automation with better results and less programming, opening up new opportunities for text analysis in political science. In this study, we evaluate LLMs on three original coding tasks involving typical complexities encountered in political science settings: a non-English language, legal and political jargon, and complex labels based on abstract constructs. Along the paper, we propose a practical workflow to optimize the choice of the model and the prompt. We find that the best prompting strategy consists of providing the LLMs with a detailed codebook, as the one provided to human coders. In this setting, an LLM can be as good as or possibly better than a human annotator while being much faster, considerably cheaper, and much easier to scale to large amounts of text. We also provide a comparison of GPT and popular open-source LLMs, discussing the trade-offs in the model's choice. Our software allows LLMs to be easily used as annotators and is publicly available: https://github.com/lorelupo/pappa.

Chia-Chien Hung, Anne Lauscher, Dirk Hovy et al.

Demographic factors (e.g., gender or age) shape our language. Previous work showed that incorporating demographic factors can consistently improve performance for various NLP tasks with traditional NLP models. In this work, we investigate whether these previous findings still hold with state-of-the-art pretrained Transformer-based language models (PLMs). We use three common specialization methods proven effective for incorporating external knowledge into pretrained Transformers (e.g., domain-specific or geographic knowledge). We adapt the language representations for the demographic dimensions of gender and age, using continuous language modeling and dynamic multi-task learning for adaptation, where we couple language modeling objectives with the prediction of demographic classes. Our results, when employing a multilingual PLM, show substantial gains in task performance across four languages (English, German, French, and Danish), which is consistent with the results of previous work. However, controlling for confounding factors - primarily domain and language proficiency of Transformer-based PLMs - shows that downstream performance gains from our demographic adaptation do not actually stem from demographic knowledge. Our results indicate that demographic specialization of PLMs, while holding promise for positive societal impact, still represents an unsolved problem for (modern) NLP.

Eliana Pastor, Alkis Koudounas, Giuseppe Attanasio et al.

Recent advances in eXplainable AI (XAI) have provided new insights into how models for vision, language, and tabular data operate. However, few approaches exist for understanding speech models. Existing work focuses on a few spoken language understanding (SLU) tasks, and explanations are difficult to interpret for most users. We introduce a new approach to explain speech classification models. We generate easy-to-interpret explanations via input perturbation on two information levels. 1) Word-level explanations reveal how each word-related audio segment impacts the outcome. 2) Paralinguistic features (e.g., prosody and background noise) answer the counterfactual: ``What would the model prediction be if we edited the audio signal in this way?'' We validate our approach by explaining two state-of-the-art SLU models on two speech classification tasks in English and Italian. Our findings demonstrate that the explanations are faithful to the model's inner workings and plausible to humans. Our method and findings pave the way for future research on interpreting speech models.

Rishav Hada, Amir Ebrahimi Fard, Sarah Shugars et al.

Increasingly taking place in online spaces, modern political conversations are typically perceived to be unproductively affirming -- siloed in so called ``echo chambers'' of exclusively like-minded discussants. Yet, to date we lack sufficient means to measure viewpoint diversity in conversations. To this end, in this paper, we operationalize two viewpoint metrics proposed for recommender systems and adapt them to the context of social media conversations. This is the first study to apply these two metrics (Representation and Fragmentation) to real world data and to consider the implications for online conversations specifically. We apply these measures to two topics -- daylight savings time (DST), which serves as a control, and the more politically polarized topic of immigration. We find that the diversity scores for both Fragmentation and Representation are lower for immigration than for DST. Further, we find that while pro-immigrant views receive consistent pushback on the platform, anti-immigrant views largely operate within echo chambers. We observe less severe yet similar patterns for DST. Taken together, Representation and Fragmentation paint a meaningful and important new picture of viewpoint diversity.

Tommaso Fornaciari, Luca Luceri, Emilio Ferrara et al.

Detecting misinformation threads is crucial to guarantee a healthy environment on social media. We address the problem using the data set created during the COVID-19 pandemic. It contains cascades of tweets discussing information weakly labeled as reliable or unreliable, based on a previous evaluation of the information source. The models identifying unreliable threads usually rely on textual features. But reliability is not just what is said, but by whom and to whom. We additionally leverage on network information. Following the homophily principle, we hypothesize that users who interact are generally interested in similar topics and spreading similar kind of news, which in turn is generally reliable or not. We test several methods to learn representations of the social interactions within the cascades, combining them with deep neural language models in a Multi-Input (MI) framework. Keeping track of the sequence of the interactions during the time, we improve over previous state-of-the-art models.

Marius Hessenthaler, Emma Strubell, Dirk Hovy et al.

Fairness and environmental impact are important research directions for the sustainable development of artificial intelligence. However, while each topic is an active research area in natural language processing (NLP), there is a surprising lack of research on the interplay between the two fields. This lacuna is highly problematic, since there is increasing evidence that an exclusive focus on fairness can actually hinder environmental sustainability, and vice versa. In this work, we shed light on this crucial intersection in NLP by (1) investigating the efficiency of current fairness approaches through surveying example methods for reducing unfair stereotypical bias from the literature, and (2) evaluating a common technique to reduce energy consumption (and thus environmental impact) of English NLP models, knowledge distillation (KD), for its impact on fairness. In this case study, we evaluate the effect of important KD factors, including layer and dimensionality reduction, with respect to: (a) performance on the distillation task (natural language inference and semantic similarity prediction), and (b) multiple measures and dimensions of stereotypical bias (e.g., gender bias measured via the Word Embedding Association Test). Our results lead us to clarify current assumptions regarding the effect of KD on unfair bias: contrary to other findings, we show that KD can actually decrease model fairness.

Debora Nozza, Dirk Hovy

Work on hate speech has made the consideration of rude and harmful examples in scientific publications inevitable. This raises various problems, such as whether or not to obscure profanities. While science must accurately disclose what it does, the unwarranted spread of hate speech is harmful to readers, and increases its internet frequency. While maintaining publications' professional appearance, obfuscating profanities makes it challenging to evaluate the content, especially for non-native speakers. Surveying 150 ACL papers, we discovered that obfuscation is usually employed for English but not other languages, and even so quite uneven. We discuss the problems with obfuscation and suggest a multilingual community resource called PrOf that has a Python module to standardize profanity obfuscation processes. We believe PrOf can help scientific publication policies to make hate speech work accessible and comparable, irrespective of language.

Flor Miriam Plaza-del-Arco, Debora Nozza, Dirk Hovy

Large Language Models (LLMs) exhibit remarkable text classification capabilities, excelling in zero- and few-shot learning (ZSL and FSL) scenarios. However, since they are trained on different datasets, performance varies widely across tasks between those models. Recent studies emphasize the importance of considering human label variation in data annotation. However, how this human label variation also applies to LLMs remains unexplored. Given this likely model specialization, we ask: Do aggregate LLM labels improve over individual models (as for human annotators)? We evaluate four recent instruction-tuned LLMs as annotators on five subjective tasks across four languages. We use ZSL and FSL setups and label aggregation from human annotation. Aggregations are indeed substantially better than any individual model, benefiting from specialization in diverse tasks or languages. Surprisingly, FSL does not surpass ZSL, as it depends on the quality of the selected examples. However, there seems to be no good information-theoretical strategy to select those. We find that no LLM method rivals even simple supervised models. We also discuss the tradeoffs in accuracy, cost, and moral/ethical considerations between LLM and human annotation.

Chia-Chien Hung, Anne Lauscher, Dirk Hovy et al.

Sociodemographic factors (e.g., gender or age) shape our language. Previous work showed that incorporating specific sociodemographic factors can consistently improve performance for various NLP tasks in traditional NLP models. We investigate whether these previous findings still hold with state-of-the-art pretrained Transformers. We use three common specialization methods proven effective for incorporating external knowledge into pretrained Transformers (e.g., domain-specific or geographic knowledge). We adapt the language representations for the sociodemographic dimensions of gender and age, using continuous language modeling and dynamic multi-task learning for adaptation, where we couple language modeling with the prediction of a sociodemographic class. Our results when employing a multilingual model show substantial performance gains across four languages (English, German, French, and Danish). These findings are in line with the results of previous work and hold promise for successful sociodemographic specialization. However, controlling for confounding factors like domain and language shows that, while sociodemographic adaptation does improve downstream performance, the gains do not always solely stem from sociodemographic knowledge. Our results indicate that sociodemographic specialization, while very important, is still an unresolved problem in NLP.

Flor Miriam Plaza-del-Arco, Amanda Cercas Curry, Susanna Paoli et al.

Emotions play important epistemological and cognitive roles in our lives, revealing our values and guiding our actions. Previous work has shown that LLMs display biases in emotion attribution along gender lines. However, unlike gender, which says little about our values, religion, as a socio-cultural system, prescribes a set of beliefs and values for its followers. Religions, therefore, cultivate certain emotions. Moreover, these rules are explicitly laid out and interpreted by religious leaders. Using emotion attribution, we explore how different religions are represented in LLMs. We find that: Major religions in the US and European countries are represented with more nuance, displaying a more shaded model of their beliefs. Eastern religions like Hinduism and Buddhism are strongly stereotyped. Judaism and Islam are stigmatized -- the models' refusal skyrocket. We ascribe these to cultural bias in LLMs and the scarcity of NLP literature on religion. In the rare instances where religion is discussed, it is often in the context of toxic language, perpetuating the perception of these religions as inherently toxic. This finding underscores the urgent need to address and rectify these biases. Our research underscores the crucial role emotions play in our lives and how our values influence them.

Fabio Pernisi, Dirk Hovy, Paul Röttger

As diverse linguistic communities and users adopt large language models (LLMs), assessing their safety across languages becomes critical. Despite ongoing efforts to make LLMs safe, they can still be made to behave unsafely with jailbreaking, a technique in which models are prompted to act outside their operational guidelines. Research on LLM safety and jailbreaking, however, has so far mostly focused on English, limiting our understanding of LLM safety in other languages. We contribute towards closing this gap by investigating the effectiveness of many-shot jailbreaking, where models are prompted with unsafe demonstrations to induce unsafe behaviour, in Italian. To enable our analysis, we create a new dataset of unsafe Italian question-answer pairs. With this dataset, we identify clear safety vulnerabilities in four families of open-weight LLMs. We find that the models exhibit unsafe behaviors even when prompted with few unsafe demonstrations, and -- more alarmingly -- that this tendency rapidly escalates with more demonstrations.

Donya Rooein, Amanda Cercas Curry, Dirk Hovy

Large language models (LLMs) offer a range of new possibilities, including adapting the text to different audiences and their reading needs. But how well do they adapt? We evaluate the readability of answers generated by four state-of-the-art LLMs (commercial and open-source) to science questions when prompted to target different age groups and education levels. To assess the adaptability of LLMs to diverse audiences, we compare the readability scores of the generated responses against the recommended comprehension level of each age and education group. We find large variations in the readability of the answers by different LLMs. Our results suggest LLM answers need to be better adapted to the intended audience demographics to be more comprehensible. They underline the importance of enhancing the adaptability of LLMs in education settings to cater to diverse age and education levels. Overall, current LLMs have set readability ranges and do not adapt well to different audiences, even when prompted. That limits their potential for educational purposes.

Giuseppe Attanasio, Beatrice Savoldi, Dennis Fucci et al.

Current automatic speech recognition (ASR) models are designed to be used across many languages and tasks without substantial changes. However, this broad language coverage hides performance gaps within languages, for example, across genders. Our study systematically evaluates the performance of two widely used multilingual ASR models on three datasets, encompassing 19 languages from eight language families and two speaking conditions. Our findings reveal clear gender disparities, with the advantaged group varying across languages and models. Surprisingly, those gaps are not explained by acoustic or lexical properties. However, probing internal model states reveals a correlation with gendered performance gap. That is, the easier it is to distinguish speaker gender in a language using probes, the more the gap reduces, favoring female speakers. Our results show that gender disparities persist even in state-of-the-art models. Our findings have implications for the improvement of multilingual ASR systems, underscoring the importance of accessibility to training data and nuanced evaluation to predict and mitigate gender gaps. We release all code and artifacts at https://github.com/g8a9/multilingual-asr-gender-gap.

Paul Röttger, Valentin Hofmann, Valentina Pyatkin et al. · allen-ai, oxford

Much recent work seeks to evaluate values and opinions in large language models (LLMs) using multiple-choice surveys and questionnaires. Most of this work is motivated by concerns around real-world LLM applications. For example, politically-biased LLMs may subtly influence society when they are used by millions of people. Such real-world concerns, however, stand in stark contrast to the artificiality of current evaluations: real users do not typically ask LLMs survey questions. Motivated by this discrepancy, we challenge the prevailing constrained evaluation paradigm for values and opinions in LLMs and explore more realistic unconstrained evaluations. As a case study, we focus on the popular Political Compass Test (PCT). In a systematic review, we find that most prior work using the PCT forces models to comply with the PCT's multiple-choice format. We show that models give substantively different answers when not forced; that answers change depending on how models are forced; and that answers lack paraphrase robustness. Then, we demonstrate that models give different answers yet again in a more realistic open-ended answer setting. We distill these findings into recommendations and open challenges in evaluating values and opinions in LLMs.

Xinpeng Wang, Bolei Ma, Chengzhi Hu et al.

The open-ended nature of language generation makes the evaluation of autoregressive large language models (LLMs) challenging. One common evaluation approach uses multiple-choice questions (MCQ) to limit the response space. The model is then evaluated by ranking the candidate answers by the log probability of the first token prediction. However, first-tokens may not consistently reflect the final response output, due to model's diverse response styles such as starting with "Sure" or refusing to answer. Consequently, MCQ evaluation is not indicative of model behaviour when interacting with users. But by how much? We evaluate how aligned first-token evaluation is with the text output along several dimensions, namely final option choice, refusal rate, choice distribution and robustness under prompt perturbation. Our results show that the two approaches are severely misaligned on all dimensions, reaching mismatch rates over 60%. Models heavily fine-tuned on conversational or safety data are especially impacted. Crucially, models remain misaligned even when we increasingly constrain prompts, i.e., force them to start with an option letter or example template. Our findings i) underscore the importance of inspecting the text output as well and ii) caution against relying solely on first-token evaluation.

Paul Röttger, Fabio Pernisi, Bertie Vidgen et al.

The last two years have seen a rapid growth in concerns around the safety of large language models (LLMs). Researchers and practitioners have met these concerns by creating an abundance of datasets for evaluating and improving LLM safety. However, much of this work has happened in parallel, and with very different goals in mind, ranging from the mitigation of near-term risks around bias and toxic content generation to the assessment of longer-term catastrophic risk potential. This makes it difficult for researchers and practitioners to find the most relevant datasets for their use case, and to identify gaps in dataset coverage that future work may fill. To remedy these issues, we conduct a first systematic review of open datasets for evaluating and improving LLM safety. We review 144 datasets, which we identified through an iterative and community-driven process over the course of several months. We highlight patterns and trends, such as a trend towards fully synthetic datasets, as well as gaps in dataset coverage, such as a clear lack of non-English and naturalistic datasets. We also examine how LLM safety datasets are used in practice -- in LLM release publications and popular LLM benchmarks -- finding that current evaluation practices are highly idiosyncratic and make use of only a small fraction of available datasets. Our contributions are based on SafetyPrompts.com, a living catalogue of open datasets for LLM safety, which we plan to update continuously as the field of LLM safety develops.

Flor Miriam Plaza-del-Arco, Amanda Cercas Curry, Alba Curry et al.

Large language models (LLMs) reflect societal norms and biases, especially about gender. While societal biases and stereotypes have been extensively researched in various NLP applications, there is a surprising gap for emotion analysis. However, emotion and gender are closely linked in societal discourse. E.g., women are often thought of as more empathetic, while men's anger is more socially accepted. To fill this gap, we present the first comprehensive study of gendered emotion attribution in five state-of-the-art LLMs (open- and closed-source). We investigate whether emotions are gendered, and whether these variations are based on societal stereotypes. We prompt the models to adopt a gendered persona and attribute emotions to an event like 'When I had a serious argument with a dear person'. We then analyze the emotions generated by the models in relation to the gender-event pairs. We find that all models consistently exhibit gendered emotions, influenced by gender stereotypes. These findings are in line with established research in psychology and gender studies. Our study sheds light on the complex societal interplay between language, gender, and emotion. The reproduction of emotion stereotypes in LLMs allows us to use those models to study the topic in detail, but raises questions about the predictive use of those same LLMs for emotion applications.

Flor Miriam Plaza-del-Arco, Alba Curry, Amanda Cercas Curry et al.

Emotions are a central aspect of communication. Consequently, emotion analysis (EA) is a rapidly growing field in natural language processing (NLP). However, there is no consensus on scope, direction, or methods. In this paper, we conduct a thorough review of 154 relevant NLP publications from the last decade. Based on this review, we address four different questions: (1) How are EA tasks defined in NLP? (2) What are the most prominent emotion frameworks and which emotions are modeled? (3) Is the subjectivity of emotions considered in terms of demographics and cultural factors? and (4) What are the primary NLP applications for EA? We take stock of trends in EA and tasks, emotion frameworks used, existing datasets, methods, and applications. We then discuss four lacunae: (1) the absence of demographic and cultural aspects does not account for the variation in how emotions are perceived, but instead assumes they are universally experienced in the same manner; (2) the poor fit of emotion categories from the two main emotion theories to the task; (3) the lack of standardized EA terminology hinders gap identification, comparison, and future goals; and (4) the absence of interdisciplinary research isolates EA from insights in other fields. Our work will enable more focused research into EA and a more holistic approach to modeling emotions in NLP.

Ilia Kuznetsov, Osama Mohammed Afzal, Koen Dercksen et al.

The number of scientific articles produced every year is growing rapidly. Providing quality control over them is crucial for scientists and, ultimately, for the public good. In modern science, this process is largely delegated to peer review -- a distributed procedure in which each submission is evaluated by several independent experts in the field. Peer review is widely used, yet it is hard, time-consuming, and prone to error. Since the artifacts involved in peer review -- manuscripts, reviews, discussions -- are largely text-based, Natural Language Processing has great potential to improve reviewing. As the emergence of large language models (LLMs) has enabled NLP assistance for many new tasks, the discussion on machine-assisted peer review is picking up the pace. Yet, where exactly is help needed, where can NLP help, and where should it stand aside? The goal of our paper is to provide a foundation for the future efforts in NLP for peer-reviewing assistance. We discuss peer review as a general process, exemplified by reviewing at AI conferences. We detail each step of the process from manuscript submission to camera-ready revision, and discuss the associated challenges and opportunities for NLP assistance, illustrated by existing work. We then turn to the big challenges in NLP for peer review as a whole, including data acquisition and licensing, operationalization and experimentation, and ethical issues. To help consolidate community efforts, we create a companion repository that aggregates key datasets pertaining to peer review. Finally, we issue a detailed call for action for the scientific community, NLP and AI researchers, policymakers, and funding bodies to help bring the research in NLP for peer review forward. We hope that our work will help set the agenda for research in machine-assisted scientific quality control in the age of AI, within the NLP community and beyond.

Diyi Yang, Dirk Hovy, David Jurgens et al.

Language technologies have made enormous progress, especially with the introduction of large language models (LLMs). On traditional tasks such as machine translation and sentiment analysis, these models perform at near-human level. These advances can, however, exacerbate a variety of issues that models have traditionally struggled with, such as bias, evaluation, and risks. In this position paper, we argue that many of these issues share a common core: a lack of awareness of the factors, context, and implications of the social environment in which NLP operates, which we call social awareness. While NLP is getting better at solving the formal linguistic aspects, limited progress has been made in adding the social awareness required for language applications to work in all situations for all users. Integrating social awareness into NLP models will make applications more natural, helpful, and safe, and will open up new possibilities. Thus we argue that substantial challenges remain for NLP to develop social awareness and that we are just at the beginning of a new era for the field.

Paul Röttger, Musashi Hinck, Valentin Hofmann et al. · allen-ai

Large language models (LLMs) are helping millions of users write texts about diverse issues, and in doing so expose users to different ideas and perspectives. This creates concerns about issue bias, where an LLM tends to present just one perspective on a given issue, which in turn may influence how users think about this issue. So far, it has not been possible to measure which issue biases LLMs manifest in real user interactions, making it difficult to address the risks from biased LLMs. Therefore, we create IssueBench: a set of 2.49m realistic English-language prompts to measure issue bias in LLM writing assistance, which we construct based on 3.9k templates (e.g. "write a blog about") and 212 political issues (e.g. "AI regulation") from real user interactions. Using IssueBench, we show that issue biases are common and persistent in 10 state-of-the-art LLMs. We also show that biases are very similar across models, and that all models align more with US Democrat than Republican voter opinion on a subset of issues. IssueBench can easily be adapted to include other issues, templates, or tasks. By enabling robust and realistic measurement, we hope that IssueBench can bring a new quality of evidence to ongoing discussions about LLM biases and how to address them.

Amanda Cercas Curry, Giuseppe Attanasio, Zeerak Talat et al.

Since the foundational work of William Labov on the social stratification of language (Labov, 1964), linguistics has made concentrated efforts to explore the links between sociodemographic characteristics and language production and perception. But while there is strong evidence for socio-demographic characteristics in language, they are infrequently used in Natural Language Processing (NLP). Age and gender are somewhat well represented, but Labov's original target, socioeconomic status, is noticeably absent. And yet it matters. We show empirically that NLP disadvantages less-privileged socioeconomic groups. We annotate a corpus of 95K utterances from movies with social class, ethnicity and geographical language variety and measure the performance of NLP systems on three tasks: language modelling, automatic speech recognition, and grammar error correction. We find significant performance disparities that can be attributed to socioeconomic status as well as ethnicity and geographical differences. With NLP technologies becoming ever more ubiquitous and quotidian, they must accommodate all language varieties to avoid disadvantaging already marginalised groups. We argue for the inclusion of socioeconomic class in future language technologies.

Joachim Baumann, Paul Röttger, Aleksandra Urman et al.

Large language models are rapidly transforming social science research by enabling the automation of labor-intensive tasks like data annotation and text analysis. However, LLM outputs vary significantly depending on the implementation choices made by researchers (e.g., model selection or prompting strategy). Such variation can introduce systematic biases and random errors, which propagate to downstream analyses and cause Type I (false positive), Type II (false negative), Type S (wrong sign), or Type M (exaggerated effect) errors. We call this phenomenon where configuration choices lead to incorrect conclusions LLM hacking. We find that intentional LLM hacking is strikingly simple. By replicating 37 data annotation tasks from 21 published social science studies, we show that, with just a handful of prompt paraphrases, virtually anything can be presented as statistically significant. Beyond intentional manipulation, our analysis of 13 million labels from 18 different LLMs across 2361 realistic hypotheses shows that there is also a high risk of accidental LLM hacking, even when following standard research practices. We find incorrect conclusions in approximately 31% of hypotheses for state-of-the-art LLMs, and in half the hypotheses for smaller language models. While higher task performance and stronger general model capabilities reduce LLM hacking risk, even highly accurate models remain susceptible. The risk of LLM hacking decreases as effect sizes increase, indicating the need for more rigorous verification of LLM-based findings near significance thresholds. We analyze 21 mitigation techniques and find that human annotations provide crucial protection against false positives. Common regression estimator correction techniques can restore valid inference but trade off Type I vs. Type II errors. We publish a list of practical recommendations to prevent LLM hacking.

Donya Rooein, Paul Rottger, Anastassia Shaitarova et al.

Using large language models (LLMs) for educational applications like dialogue-based teaching is a hot topic. Effective teaching, however, requires teachers to adapt the difficulty of content and explanations to the education level of their students. Even the best LLMs today struggle to do this well. If we want to improve LLMs on this adaptation task, we need to be able to measure adaptation success reliably. However, current Static metrics for text difficulty, like the Flesch-Kincaid Reading Ease score, are known to be crude and brittle. We, therefore, introduce and evaluate a new set of Prompt-based metrics for text difficulty. Based on a user study, we create Prompt-based metrics as inputs for LLMs. They leverage LLM's general language understanding capabilities to capture more abstract and complex features than Static metrics. Regression experiments show that adding our Prompt-based metrics significantly improves text difficulty classification over Static metrics alone. Our results demonstrate the promise of using LLMs to evaluate text adaptation to different education levels.

Nikita Soni, Niranjan Balasubramanian, H. Andrew Schwartz et al.

Pre-trained language models consider the context of neighboring words and documents but lack any author context of the human generating the text. However, language depends on the author's states, traits, social, situational, and environmental attributes, collectively referred to as human context (Soni et al., 2024). Human-centered natural language processing requires incorporating human context into language models. Currently, two methods exist: pre-training with 1) group-wise attributes (e.g., over-45-year-olds) or 2) individual traits. Group attributes are simple but coarse -- not all 45-year-olds write the same way -- while individual traits allow for more personalized representations, but require more complex modeling and data. It is unclear which approach benefits what tasks. We compare pre-training models with human context via 1) group attributes, 2) individual users, and 3) a combined approach on five user- and document-level tasks. Our results show that there is no best approach, but that human-centered language modeling holds avenues for different methods.

Giuseppe Russo, Debora Nozza, Paul Röttger et al.

People increasingly rely on Large Language Models (LLMs) for moral advice, which may influence humans' decisions. Yet, little is known about how closely LLMs align with human moral judgments. To address this, we introduce the Moral Dilemma Dataset, a benchmark of 1,618 real-world moral dilemmas paired with a distribution of human moral judgments consisting of a binary evaluation and a free-text rationale. We treat this problem as a pluralistic distributional alignment task, comparing the distributions of LLM and human judgments across dilemmas. We find that models reproduce human judgments only under high consensus; alignment deteriorates sharply when human disagreement increases. In parallel, using a 60-value taxonomy built from 3,783 value expressions extracted from rationales, we show that LLMs rely on a narrower set of moral values than humans. These findings reveal a pluralistic moral gap: a mismatch in both the distribution and diversity of values expressed. To close this gap, we introduce Dynamic Moral Profiling (DMP), a Dirichlet-based sampling method that conditions model outputs on human-derived value profiles. DMP improves alignment by 64.3% and enhances value diversity, offering a step toward more pluralistic and human-aligned moral guidance from LLMs.

Paul Röttger, Giuseppe Attanasio, Felix Friedrich et al.

Vision-language models (VLMs), which process image and text inputs, are increasingly integrated into chat assistants and other consumer AI applications. Without proper safeguards, however, VLMs may give harmful advice (e.g. how to self-harm) or encourage unsafe behaviours (e.g. to consume drugs). Despite these clear hazards, little work so far has evaluated VLM safety and the novel risks created by multimodal inputs. To address this gap, we introduce MSTS, a Multimodal Safety Test Suite for VLMs. MSTS comprises 400 test prompts across 40 fine-grained hazard categories. Each test prompt consists of a text and an image that only in combination reveal their full unsafe meaning. With MSTS, we find clear safety issues in several open VLMs. We also find some VLMs to be safe by accident, meaning that they are safe because they fail to understand even simple test prompts. We translate MSTS into ten languages, showing non-English prompts to increase the rate of unsafe model responses. We also show models to be safer when tested with text only rather than multimodal prompts. Finally, we explore the automation of VLM safety assessments, finding even the best safety classifiers to be lacking.

Lorenzo Lupo, Paul Bose, Mahyar Habibi et al.

Social scientists increasingly use demographically stratified social media data to study the attitudes, beliefs, and behavior of the general public. To facilitate such analyses, we construct, validate, and release publicly the representative DADIT dataset of 30M tweets of 20k Italian Twitter users, along with their bios and profile pictures. We enrich the user data with high-quality labels for gender, age, and location. DADIT enables us to train and compare the performance of various state-of-the-art models for the prediction of the gender and age of social media users. In particular, we investigate if tweets contain valuable information for the task, since popular classifiers like M3 don't leverage them. Our best XLM-based classifier improves upon the commonly used competitor M3 by up to 53% F1. Especially for age prediction, classifiers profit from including tweets as features. We also confirm these findings on a German test set.

Tiancheng Hu, Joachim Baumann, Lorenzo Lupo et al.

Large language model (LLM) simulations of human behavior have the potential to revolutionize the social and behavioral sciences, if and only if they faithfully reflect real human behaviors. Current evaluations are fragmented, based on bespoke tasks and metrics, creating a patchwork of incomparable results. To address this, we introduce SimBench, the first large-scale, standardized benchmark for a robust, reproducible science of LLM simulation. By unifying 20 diverse datasets covering tasks from moral decision-making to economic choice across a large global participant pool, SimBench provides the necessary foundation to ask fundamental questions about when, how, and why LLM simulations succeed or fail. We show that, while even the best LLMs today have limited simulation ability (score: 40.80/100), performance scales log-linearly with model size. Simulation performance is not improved by increased inference-time compute. We demonstrate an alignment-simulation trade-off: instruction-tuning improves performance on low-entropy (consensus) questions but degrades it on high-entropy (diverse) ones. Models particularly struggle when simulating specific demographic groups. Finally, we demonstrate that simulation ability correlates most strongly with deep, knowledge-intensive reasoning (MMLU-Pro, r=0.939). By making progress measurable, we aim to accelerate the development of more faithful LLM simulators.

Jingwei Ni, Yu Fan, Vilém Zouhar et al. · eth-zurich

Variation in human annotation (i.e., disagreements) is common in NLP, often reflecting important information like task subjectivity and sample ambiguity. Modeling this variation is important for applications that are sensitive to such information. Although RLVR-style reasoning (Reinforcement Learning with Verifiable Rewards) has improved Large Language Model (LLM) performance on many tasks, it remains unclear whether such reasoning enables LLMs to capture informative variation in human annotation. In this work, we evaluate the influence of different reasoning settings on LLM disagreement modeling. We systematically evaluate each reasoning setting across model sizes, distribution expression methods, and steering methods, resulting in 60 experimental setups across 3 tasks. Surprisingly, our results show that RLVR-style reasoning degrades performance in disagreement modeling, while naive Chain-of-Thought (CoT) reasoning improves the performance of RLHF LLMs (RL from human feedback). These findings underscore the potential risk of replacing human annotators with reasoning LLMs, especially when disagreements are important.

Donya Rooein, Vilém Zouhar, Debora Nozza et al.

Stories play a pivotal role in human communication, shaping beliefs and morals, particularly in children. As parents increasingly rely on large language models (LLMs) to craft bedtime stories, the presence of cultural and gender stereotypes in these narratives raises significant concerns. To address this issue, we present Biased Tales, a comprehensive dataset designed to analyze how biases influence protagonists' attributes and story elements in LLM-generated stories. Our analysis uncovers striking disparities. When the protagonist is described as a girl (as compared to a boy), appearance-related attributes increase by 55.26%. Stories featuring non-Western children disproportionately emphasize cultural heritage, tradition, and family themes far more than those for Western children. Our findings highlight the role of sociocultural bias in making creative AI use more equitable and diverse.

Chenfei Xiong, Jingwei Ni, Yu Fan et al. · eth-zurich

We introduce Co-DETECT (Collaborative Discovery of Edge cases in TExt ClassificaTion), a novel mixed-initiative annotation framework that integrates human expertise with automatic annotation guided by large language models (LLMs). Co-DETECT starts with an initial, sketch-level codebook and dataset provided by a domain expert, then leverages the LLM to annotate the data and identify edge cases that are not well described by the initial codebook. Specifically, Co-DETECT flags challenging examples, induces high-level, generalizable descriptions of edge cases, and assists user in incorporating edge case handling rules to improve the codebook. This iterative process enables more effective handling of nuanced phenomena through compact, generalizable annotation rules. Extensive user study, qualitative and quantitative analyses prove the effectiveness of Co-DETECT.

Donya Rooein, Dirk Hovy

Open conversations are one of the most engaging forms of teaching. However, creating those conversations in educational software is a complex endeavor, especially if we want to address the needs of different audiences. While language models hold great promise for educational applications, there are substantial challenges in training them to engage in meaningful and effective conversational teaching, especially when considering the diverse needs of various audiences. No official data sets exist for this task to facilitate the training of language models for conversational teaching, considering the diverse needs of various audiences. This paper presents a novel source for facilitating conversational teaching of scientific concepts at various difficulty levels (from preschooler to expert), namely dialogues taken from video transcripts. We analyse this data source in various ways to show that it offers a diverse array of examples that can be used to generate contextually appropriate and natural responses to scientific topics for specific target audiences. It is a freely available valuable resource for training and evaluating conversation models, encompassing organically occurring dialogues. While the raw data is available online, we provide additional metadata for conversational analysis of dialogues at each level in all available videos.

Amanda Cercas Curry, Zeerak Talat, Dirk Hovy

Since Labov's (1964) foundational work on the social stratification of language, linguistics has dedicated concerted efforts towards understanding the relationships between socio-demographic factors and language production and perception. Despite the large body of evidence identifying significant relationships between socio-demographic factors and language production, relatively few of these factors have been investigated in the context of NLP technology. While age and gender are well covered, Labov's initial target, socio-economic class, is largely absent. We survey the existing Natural Language Processing (NLP) literature and find that only 20 papers even mention socio-economic status. However, the majority of those papers do not engage with class beyond collecting information of annotator-demographics. Given this research lacuna, we provide a definition of class that can be operationalised by NLP researchers, and argue for including socio-economic class in future language technologies.

Cesar Gonzalez-Gutierrez, Dirk Hovy

Prompting is a common approach for leveraging LMs in zero-shot settings. However, the underlying mechanisms that enable LMs to perform diverse tasks without task-specific supervision remain poorly understood. Studying the relationship between prompting and the quality of internal representations can shed light on how pre-trained embeddings may support in-context task solving. In this empirical study, we conduct a series of probing experiments on prompt embeddings, analyzing various combinations of prompt templates for zero-shot classification. Our findings show that while prompting affects the quality of representations, these changes do not consistently correlate with the relevance of the prompts to the target task. This result challenges the assumption that more relevant prompts necessarily lead to better representations. We further analyze potential factors that may contribute to this unexpected behavior.

Flor Miriam Plaza-del-Arco, Paul Röttger, Nino Scherrer et al.

Large language models (LLMs) are increasingly integrated into our daily lives and personalized. However, LLM personalization might also increase unintended side effects. Recent work suggests that persona prompting can lead models to falsely refuse user requests. However, no work has fully quantified the extent of this issue. To address this gap, we measure the impact of 15 sociodemographic personas (based on gender, race, religion, and disability) on false refusal. To control for other factors, we also test 16 different models, 3 tasks (Natural Language Inference, politeness, and offensiveness classification), and nine prompt paraphrases. We propose a Monte Carlo-based method to quantify this issue in a sample-efficient manner. Our results show that as models become more capable, personas impact the refusal rate less and less. Certain sociodemographic personas increase false refusal in some models, which suggests underlying biases in the alignment strategies or safety mechanisms. However, we find that the model choice and task significantly influence false refusals, especially in sensitive content tasks. Our findings suggest that persona effects have been overestimated, and might be due to other factors.

Donya Rooein, Flor Miriam Plaza-del-Arco, Debora Nozza et al.

Given Farsi's speaker base of over 127 million people and the growing availability of digital text, including more than 1.3 million articles on Wikipedia, it is considered a middle-resource language. However, this label quickly crumbles when the situation is examined more closely. We focus on three subjective tasks (Sentiment Analysis, Emotion Analysis, and Toxicity Detection) and find significant challenges in data availability and quality, despite the overall increase in data availability. We review 110 publications on subjective tasks in Farsi and observe a lack of publicly available datasets. Furthermore, existing datasets often lack essential demographic factors, such as age and gender, that are crucial for accurately modeling subjectivity in language. When evaluating prediction models using the few available datasets, the results are highly unstable across both datasets and models. Our findings indicate that the volume of data is insufficient to significantly improve a language's prospects in NLP.

Pedro Henrique Luz de Araujo, Paul Röttger, Dirk Hovy et al.

Expert persona prompting -- assigning roles such as expert in math to language models -- is widely used for task improvement. However, prior work shows mixed results on its effectiveness, and does not consider when and why personas should improve performance. We analyze the literature on persona prompting for task improvement and distill three desiderata: 1) performance advantage of expert personas, 2) robustness to irrelevant persona attributes, and 3) fidelity to persona attributes. We then evaluate 9 state-of-the-art LLMs across 27 tasks with respect to these desiderata. We find that expert personas usually lead to positive or non-significant performance changes. Surprisingly, models are highly sensitive to irrelevant persona details, with performance drops of almost 30 percentage points. In terms of fidelity, we find that while higher education, specialization, and domain-relatedness can boost performance, their effects are often inconsistent or negligible across tasks. We propose mitigation strategies to improve robustness -- but find they only work for the largest, most capable models. Our findings underscore the need for more careful persona design and for evaluation schemes that reflect the intended effects of persona usage.

Anne Lauscher, Debora Nozza, Archie Crowley et al.

As 3rd-person pronoun usage shifts to include novel forms, e.g., neopronouns, we need more research on identity-inclusive NLP. Exclusion is particularly harmful in one of the most popular NLP applications, machine translation (MT). Wrong pronoun translations can discriminate against marginalized groups, e.g., non-binary individuals (Dev et al., 2021). In this ``reality check'', we study how three commercial MT systems translate 3rd-person pronouns. Concretely, we compare the translations of gendered vs. gender-neutral pronouns from English to five other languages (Danish, Farsi, French, German, Italian), and vice versa, from Danish to English. Our error analysis shows that the presence of a gender-neutral pronoun often leads to grammatical and semantic translation errors. Similarly, gender neutrality is often not preserved. By surveying the opinions of affected native speakers from diverse languages, we provide recommendations to address the issue in future MT research.

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.

Anne Lauscher, Archie Crowley, Dirk Hovy

The world of pronouns is changing. From a closed class of words with few members to a much more open set of terms to reflect identities. However, Natural Language Processing (NLP) is barely reflecting this linguistic shift, even though recent work outlined the harms of gender-exclusive language technology. Particularly problematic is the current modeling 3rd person pronouns, as it largely ignores various phenomena like neopronouns, i.e., pronoun sets that are novel and not (yet) widely established. This omission contributes to the discrimination of marginalized and underrepresented groups, e.g., non-binary individuals. However, other identity-expression phenomena beyond gender are also ignored by current NLP technology. In this paper, we provide an overview of 3rd person pronoun issues for NLP. Based on our observations and ethical considerations, we define a series of desiderata for modeling pronouns in language technology. We evaluate existing and novel modeling approaches w.r.t. these desiderata qualitatively, and quantify the impact of a more discrimination-free approach on established benchmark data.

Federico Bianchi, Vincenzo Cutrona, Dirk Hovy

Twitter data have become essential to Natural Language Processing (NLP) and social science research, driving various scientific discoveries in recent years. However, the textual data alone are often not enough to conduct studies: especially social scientists need more variables to perform their analysis and control for various factors. How we augment this information, such as users' location, age, or tweet sentiment, has ramifications for anonymity and reproducibility, and requires dedicated effort. This paper describes Twitter-Demographer, a simple, flow-based tool to enrich Twitter data with additional information about tweets and users. Twitter-Demographer is aimed at NLP practitioners and (computational) social scientists who want to enrich their datasets with aggregated information, facilitating reproducibility, and providing algorithmic privacy-by-design measures for pseudo-anonymity. We discuss our design choices, inspired by the flow-based programming paradigm, to use black-box components that can easily be chained together and extended. We also analyze the ethical issues related to the use of this tool, and the built-in measures to facilitate pseudo-anonymity.