Alexander Fraser

Alexandra Chronopoulou, Matthew E. Peters, Alexander Fraser et al. · allen-ai, cmu

Pretrained language models (PLMs) are trained on massive corpora, but often need to specialize to specific domains. A parameter-efficient adaptation method suggests training an adapter for each domain on the task of language modeling. This leads to good in-domain scores but can be impractical for domain- or resource-restricted settings. A solution is to use a related-domain adapter for the novel domain at test time. In this paper, we introduce AdapterSoup, an approach that performs weight-space averaging of adapters trained on different domains. Our approach is embarrassingly parallel: first, we train a set of domain-specific adapters; then, for each novel domain, we determine which adapters should be averaged at test time. We present extensive experiments showing that AdapterSoup consistently improves performance to new domains without extra training. We also explore weight averaging of adapters trained on the same domain with different hyper-parameters, and show that it preserves the performance of a PLM on new domains while obtaining strong in-domain results. We explore various approaches for choosing which adapters to combine, such as text clustering and semantic similarity. We find that using clustering leads to the most competitive results on novel domains.

Alexandra Chronopoulou, Dario Stojanovski, Alexander Fraser · microsoft-research

Large multilingual models trained with self-supervision achieve state-of-the-art results in a wide range of natural language processing tasks. Self-supervised pretrained models are often fine-tuned on parallel data from one or multiple language pairs for machine translation. Multilingual fine-tuning improves performance on low-resource languages but requires modifying the entire model and can be prohibitively expensive. Training a new adapter on each language pair or training a single adapter on all language pairs without updating the pretrained model has been proposed as a parameter-efficient alternative. However, the former does not permit any sharing between languages, while the latter shares parameters for all languages and is susceptible to negative interference. In this paper, we propose training language-family adapters on top of mBART-50 to facilitate cross-lingual transfer. Our approach outperforms related baselines, yielding higher translation scores on average when translating from English to 17 different low-resource languages. We also show that language-family adapters provide an effective method to translate to languages unseen during pretraining.

Marius Gassen, Benjamin Hättasch, Benjamin Hilprecht et al.

Databases for OLTP are often the backbone for applications such as hotel room or cinema ticket booking applications. However, developing a conversational agent (i.e., a chatbot-like interface) to allow end-users to interact with an application using natural language requires both immense amounts of training data and NLP expertise. This motivates CAT, which can be used to easily create conversational agents for transactional databases. The main idea is that, for a given OLTP database, CAT uses weak supervision to synthesize the required training data to train a state-of-the-art conversational agent, allowing users to interact with the OLTP database. Furthermore, CAT provides an out-of-the-box integration of the resulting agent with the database. As a major difference to existing conversational agents, agents synthesized by CAT are data-aware. This means that the agent decides which information should be requested from the user based on the current data distributions in the database, which typically results in markedly more efficient dialogues compared with non-data-aware agents. We publish the code for CAT as open source.

Sophie Henning, William Beluch, Alexander Fraser et al.

Many natural language processing (NLP) tasks are naturally imbalanced, as some target categories occur much more frequently than others in the real world. In such scenarios, current NLP models still tend to perform poorly on less frequent classes. Addressing class imbalance in NLP is an active research topic, yet, finding a good approach for a particular task and imbalance scenario is difficult. With this survey, the first overview on class imbalance in deep-learning based NLP, we provide guidance for NLP researchers and practitioners dealing with imbalanced data. We first discuss various types of controlled and real-world class imbalance. Our survey then covers approaches that have been explicitly proposed for class-imbalanced NLP tasks or, originating in the computer vision community, have been evaluated on them. We organize the methods by whether they are based on sampling, data augmentation, choice of loss function, staged learning, or model design. Finally, we discuss open problems such as dealing with multi-label scenarios, and propose systematic benchmarking and reporting in order to move forward on this problem as a community.

Katharina Hämmerl, Björn Deiseroth, Patrick Schramowski et al.

Pre-trained multilingual language models (PMLMs) are commonly used when dealing with data from multiple languages and cross-lingual transfer. However, PMLMs are trained on varying amounts of data for each language. In practice this means their performance is often much better on English than many other languages. We explore to what extent this also applies to moral norms. Do the models capture moral norms from English and impose them on other languages? Do the models exhibit random and thus potentially harmful beliefs in certain languages? Both these issues could negatively impact cross-lingual transfer and potentially lead to harmful outcomes. In this paper, we (1) apply the MoralDirection framework to multilingual models, comparing results in German, Czech, Arabic, Chinese, and English, (2) analyse model behaviour on filtered parallel subtitles corpora, and (3) apply the models to a Moral Foundations Questionnaire, comparing with human responses from different countries. Our experiments demonstrate that, indeed, PMLMs encode differing moral biases, but these do not necessarily correspond to cultural differences or commonalities in human opinions. We release our code and models.

Silvia Severini, Viktor Hangya, Masoud Jalili Sabet et al.

Bilingual Word Embeddings (BWEs) are one of the cornerstones of cross-lingual transfer of NLP models. They can be built using only monolingual corpora without supervision leading to numerous works focusing on unsupervised BWEs. However, most of the current approaches to build unsupervised BWEs do not compare their results with methods based on easy-to-access cross-lingual signals. In this paper, we argue that such signals should always be considered when developing unsupervised BWE methods. The two approaches we find most effective are: 1) using identical words as seed lexicons (which unsupervised approaches incorrectly assume are not available for orthographically distinct language pairs) and 2) combining such lexicons with pairs extracted by matching romanized versions of words with an edit distance threshold. We experiment on thirteen non-Latin languages (and English) and show that such cheap signals work well and that they outperform using more complex unsupervised methods on distant language pairs such as Chinese, Japanese, Kannada, Tamil, and Thai. In addition, they are even competitive with the use of high-quality lexicons in supervised approaches. Our results show that these training signals should not be neglected when building BWEs, even for distant languages.

Katharina Hämmerl, Björn Deiseroth, Patrick Schramowski et al.

Massively multilingual sentence representations are trained on large corpora of uncurated data, with a very imbalanced proportion of languages included in the training. This may cause the models to grasp cultural values including moral judgments from the high-resource languages and impose them on the low-resource languages. The lack of data in certain languages can also lead to developing random and thus potentially harmful beliefs. Both these issues can negatively influence zero-shot cross-lingual model transfer and potentially lead to harmful outcomes. Therefore, we aim to (1) detect and quantify these issues by comparing different models in different languages, (2) develop methods for improving undesirable properties of the models. Our initial experiments using the multilingual model XLM-R show that indeed multilingual LMs capture moral norms, even with potentially higher human-agreement than monolingual ones. However, it is not yet clear to what extent these moral norms differ between languages.

Katharina Hämmerl, Jindřich Libovický, Alexander Fraser

Static and contextual multilingual embeddings have complementary strengths. Static embeddings, while less expressive than contextual language models, can be more straightforwardly aligned across multiple languages. We combine the strengths of static and contextual models to improve multilingual representations. We extract static embeddings for 40 languages from XLM-R, validate those embeddings with cross-lingual word retrieval, and then align them using VecMap. This results in high-quality, highly multilingual static embeddings. Then we apply a novel continued pre-training approach to XLM-R, leveraging the high quality alignment of our static embeddings to better align the representation space of XLM-R. We show positive results for multiple complex semantic tasks. We release the static embeddings and the continued pre-training code. Unlike most previous work, our continued pre-training approach does not require parallel text.

Katharina Hämmerl, Alina Fastowski, Jindřich Libovický et al.

Previous work has shown that the representations output by contextual language models are more anisotropic than static type embeddings, and typically display outlier dimensions. This seems to be true for both monolingual and multilingual models, although much less work has been done on the multilingual context. Why these outliers occur and how they affect the representations is still an active area of research. We investigate outlier dimensions and their relationship to anisotropy in multiple pre-trained multilingual language models. We focus on cross-lingual semantic similarity tasks, as these are natural tasks for evaluating multilingual representations. Specifically, we examine sentence representations. Sentence transformers which are fine-tuned on parallel resources (that are not always available) perform better on this task, and we show that their representations are more isotropic. However, we aim to improve multilingual representations in general. We investigate how much of the performance difference can be made up by only transforming the embedding space without fine-tuning, and visualise the resulting spaces. We test different operations: Removing individual outlier dimensions, cluster-based isotropy enhancement, and ZCA whitening. We publish our code for reproducibility.

Songbo Hu, Yinhong Liu, Ej Zhou et al.

Creating spoken dialogue datasets is methodologically challenging, and these challenges are amplified when the goal is to build multilingual, multi-parallel datasets at scale. This work introduces HEALTHDIAL, a large-scale, multilingual, and multi-parallel dataset for developing and evaluating retrieval-augmented generation (RAG)-based spoken dialogue systems. The dataset comprises 6,000 information-seeking dialogues (1,500 per language) grounded in trusted content from the World Health Organization (WHO) and 163 hours of user speech recorded from native speakers of diverse dialects across four official WHO languages: Arabic, Chinese, English, and Spanish. Each speaker is annotated with demographic (e.g., gender, age) and sociolinguistic (e.g., primary language, region of origin) variables. We report benchmark results across key dialogue tasks, which reveal consistent performance disparities across languages, even among high-resource ones. To support future research, we release the dataset, a prototype system, and a toolkit for data collection and system evaluation.

Viktor Hangya, Silvia Severini, Radoslav Ralev et al.

Very low-resource languages, having only a few million tokens worth of data, are not well-supported by multilingual NLP approaches due to poor quality cross-lingual word representations. Recent work showed that good cross-lingual performance can be achieved if a source language is related to the low-resource target language. However, not all language pairs are related. In this paper, we propose to build multilingual word embeddings (MWEs) via a novel language chain-based approach, that incorporates intermediate related languages to bridge the gap between the distant source and target. We build MWEs one language at a time by starting from the resource rich source and sequentially adding each language in the chain till we reach the target. We extend a semi-joint bilingual approach to multiple languages in order to eliminate the main weakness of previous works, i.e., independently trained monolingual embeddings, by anchoring the target language around the multilingual space. We evaluate our method on bilingual lexicon induction for 4 language families, involving 4 very low-resource (<5M tokens) and 4 moderately low-resource (<50M) target languages, showing improved performance in both categories. Additionally, our analysis reveals the importance of good quality embeddings for intermediate languages as well as the importance of leveraging anchor points from all languages in the multilingual space.

Wen Lai, Alexandra Chronopoulou, Alexander Fraser

Multilingual neural machine translation models (MNMT) yield state-of-the-art performance when evaluated on data from a domain and language pair seen at training time. However, when a MNMT model is used to translate under domain shift or to a new language pair, performance drops dramatically. We consider a very challenging scenario: adapting the MNMT model both to a new domain and to a new language pair at the same time. In this paper, we propose $m^4Adapter$ (Multilingual Multi-Domain Adaptation for Machine Translation with a Meta-Adapter), which combines domain and language knowledge using meta-learning with adapters. We present results showing that our approach is a parameter-efficient solution which effectively adapts a model to both a new language pair and a new domain, while outperforming other adapter methods. An ablation study also shows that our approach more effectively transfers domain knowledge across different languages and language information across different domains.

Mohammad Amine Jradi, Faeze Ghorbanpour, Alexander Fraser

Hate speech annotation is costly, subjective, and prone to annotator disagreement, making large-scale dataset construction challenging. We systematically analyze how well large language models (LLMs) align with human judgments across ten theoretically grounded subjective attributes, such as dehumanization, violence, and sentiment, evaluating both small and large variants of Llama 3.1 and Qwen 2.5. Our analysis reveals a consistent split across all models: behaviorally explicit dimensions (insult, humiliate, attack-defend) correlate strongly with human annotations, while evaluative dimensions (respect, sentiment, hate speech) are systematically inverted. Demographic persona conditioning reduces model confidence without improving alignment. Building on these insights, we propose combining attribute-level LLM predictions via a confidence-weighted Ridge regression to reconstruct continuous hate speech scores from the Measuring Hate Speech corpus, achieving $R^2$ of up to 0.71 and outperforming direct prompting baselines, demonstrating that structured attribute decomposition recovers a richer and more human-aligned signal than end-to-end label prediction alone.

Ruhallah Niazi, Faeze Ghorbanpour, Alexander Fraser

Despite impressive multilingual capabilities, large language models (LLMs) remain poorly evaluated on literary knowledge in non-English languages. We introduce PersLitEval, a benchmark of 4,514 Persian literature multiple-choice questions across eight fine-grained categories spanning spelling, literary devices, grammar, vocabulary, word formation, and conceptual understanding, sourced from materials for the Konkur university entrance examination. We evaluate six LLMs across ten prompting strategies, revealing striking category-level disparities across three tiers of task difficulty: models reach higher accuracy on conceptual similarity tasks but struggle with formal linguistic analysis, with spelling and word formation proving the hardest across all models. Prompting strategy has a significant impact on performance, with explained few-shot examples yielding the best results, particularly on formal linguistic categories. An error analysis identifies three failure modes: semantic comprehension gaps, formal linguistic knowledge gaps, and counting/enumeration errors, suggesting that different categories require different improvement strategies.

Wen Lai, Yingli Shen, Dingnan Jin et al.

Autoregressive language models are widely used for text evaluation, however, their left-to-right factorization introduces positional bias, i.e., early tokens are scored with only leftward context, conflating architectural asymmetry with true text quality. We propose masked reconstruction as an alternative paradigm, where every token is scored using full bidirectional context. We introduce DiffScore, an evaluation framework built on Masked Large Diffusion Language Models. By measuring text recoverability across continuous masking rates, DiffScore eliminates positional bias and naturally establishes an evaluation hierarchy from local fluency to global coherence. We further provide diagnostic tools unavailable to autoregressive frameworks: multi-timestep quality profiles that decompose scores across masking rates, and bidirectional PMI decomposition that disentangles fluency from faithfulness. Experiments across ten benchmarks show that DiffScore consistently outperforms autoregressive baselines in both zero-shot and fine-tuned settings. The code is released at: https://github.com/wenlai-lavine/DiffScore.

Marion Weller-Di Marco, Matthias Huck, Alexander Fraser

Morphologically rich languages pose difficulties to machine translation. Machine translation engines that rely on statistical learning from parallel training data, such as state-of-the-art neural systems, face challenges especially with rich morphology on the output language side. Key challenges of rich target-side morphology in data-driven machine translation include: (1) A large amount of differently inflected word surface forms entails a larger vocabulary and thus data sparsity. (2) Some inflected forms of infrequent terms typically do not appear in the training corpus, which makes closed-vocabulary systems unable to generate these unobserved variants. (3) Linguistic agreement requires the system to correctly match the grammatical categories between inflected word forms in the output sentence, both in terms of target-side morpho-syntactic wellformedness and semantic adequacy with respect to the input. In this paper, we re-investigate two target-side linguistic processing techniques: a lemma-tag strategy and a linguistically informed word segmentation strategy. Our experiments are conducted on a English-German translation task under three training corpus conditions of different magnitudes. We find that a stronger Transformer baseline leaves less room for improvement than a shallow-RNN encoder-decoder model when translating in-domain. However, we find that linguistic modeling of target-side morphology does benefit the Transformer model when the same system is applied to out-of-domain input text. We also successfully apply our approach to English to Czech translation.

Yingli Shen, Wen Lai, Jie Zhou et al.

While LLMs exhibit remarkable fluency, their utility is often compromised by factual hallucinations and a lack of traceable provenance. Existing resources for grounding mitigate this but typically enforce a dichotomy: they offer either structured knowledge without textual context (e.g., knowledge bases) or grounded text with limited scale and linguistic coverage. To bridge this gap, we introduce FactNet, a massive, open-source resource designed to unify 1.7 billion atomic assertions with 3.01 billion auditable evidence pointers derived exclusively from 316 Wikipedia editions. Unlike recent synthetic approaches, FactNet employs a strictly deterministic construction pipeline, ensuring that every evidence unit is recoverable with byte-level precision. Extensive auditing confirms a high grounding precision of 92.1%, even in long-tail languages. Furthermore, we establish FactNet-Bench, a comprehensive evaluation suite for Knowledge Graph Completion, Question Answering, and Fact Checking. FactNet provides the community with a foundational, reproducible resource for training and evaluating trustworthy, verifiable multilingual systems.

Wen Lai, Viktor Hangya, Yingli Shen et al.

Despite the growing variety of languages supported by existing multilingual neural machine translation (MNMT) models, most of the world's languages are still being left behind. We aim to extend large-scale MNMT models to incorporate a new language, enabling translations between this new language and all previously supported languages, even in the challenging scenario where only a parallel corpus between the new language and English is available. Previous methods, such as continued training on parallel data including the new language, often suffer from catastrophic forgetting, which degrades performance on other languages. We propose a novel approach Imit-MNMT which treats this task as an imitation learning problem, a technique widely used in computer vision but less explored in natural language processing. Specifically, we leverage an expert model to generate pseudo-parallel corpora between the new language and the existing languages. We then introduce a data distribution imitation strategy using language-specific weighting, alongside a translation behavior imitation mechanism. Extensive experiments show that our approach significantly improves translation performance between the new and existing languages while mitigating catastrophic forgetting.

Diego Miguel Lozano, Daryna Dementieva, Alexander Fraser

Semantic Textual Similarity (STS) is a crucial component of many Natural Language Processing (NLP) applications. However, existing approaches typically reduce semantic nuances to a single score, limiting interpretability. To address this, we introduce the task of Dissimilar Span Detection (DSD), which aims to identify semantically differing spans between pairs of texts. This can help users understand which particular words or tokens negatively affect the similarity score, or be used to improve performance in STS-dependent downstream tasks. Furthermore, we release a new dataset suitable for the task, the Span Similarity Dataset (SSD), developed through a semi-automated pipeline combining large language models (LLMs) with human verification. We propose and evaluate different baseline methods for DSD, both unsupervised, based on LIME, SHAP, LLMs, and our own method, as well as an additional supervised approach. While LLMs and supervised models achieve the highest performance, overall results remain low, highlighting the complexity of the task. Finally, we set up an additional experiment that shows how DSD can lead to increased performance in the specific task of paraphrase detection.

Sinan Kurtyigit, Sabine Schulte im Walde, Alexander Fraser

Metaphor detection models achieve strong benchmark performance, yet it remains unclear whether this reflects transferable generalization or lexical memorization. To address this, we analyze generalization in metaphor detection through RoBERTa, the shared backbone of many state-of-the-art systems, focusing on English verbs using the VU Amsterdam Metaphor Corpus. We introduce a controlled lexical hold-out setup where all instances of selected target lemmas are strictly excluded from fine-tuning, and compare predictions on these Held-out lemmas against Exposed lemmas (verbs seen during fine-tuning). While the model performs best on Exposed lemmas, it maintains robust performance on Held-out lemmas. Further analysis reveals that sentence context alone is sufficient to match full-model performance on Held-out lemmas, whereas static verb-level embeddings are not. Together, these results suggest that generalization is primarily driven by "learning the cue" (transferable contextual patterns), while "learning the word" (verb-specific memorization) provides an additive boost when lexical exposure is available.

Katharina Hämmerl, Tomasz Limisiewicz, Jindřich Libovický et al.

Previous work has considered token overlap, or even similarity of token distributions, as predictors for multilinguality and cross-lingual knowledge transfer in language models. However, these very literal metrics assign large distances to language pairs with different scripts, which can nevertheless show good cross-linguality. This limits the explanatory strength of token overlap for knowledge transfer between language pairs that use distinct scripts or follow different orthographic conventions. In this paper, we propose subword token alignability as a new way to understand the impact and quality of multilingual tokenisation. In particular, this metric predicts multilinguality much better when scripts are disparate and the overlap of literal tokens is low. We analyse this metric in the context of both encoder and decoder models, look at data size as a potential distractor, and discuss how this insight may be applied to multilingual tokenisation in future work. We recommend our subword token alignability metric for identifying optimal language pairs for cross-lingual transfer, as well as to guide the construction of better multilingual tokenisers in the future. We publish our code and reproducibility details.

Lukas Edman, Helmut Schmid, Alexander Fraser

Large Language Models (LLMs) show remarkable performance on a wide variety of tasks. Most LLMs split text into multi-character tokens and process them as atomic units without direct access to individual characters. This raises the question: To what extent can LLMs learn orthographic information? To answer this, we propose a new benchmark, CUTE, which features a collection of tasks designed to test the orthographic knowledge of LLMs. We evaluate popular LLMs on CUTE, finding that most of them seem to know the spelling of their tokens, yet fail to use this information effectively to manipulate text, calling into question how much of this knowledge is generalizable.

Wen Lai, Alexander Fraser, Ivan Titov

Parameter-efficient fine-tuning (PEFT) methods, such as LoRA, are commonly used to adapt LLMs. However, the effectiveness of standard PEFT methods is limited in low-resource scenarios with only a few hundred examples. Recent advances in interpretability research have inspired the emergence of activation editing (or steering) techniques, which modify the activations of specific model components. Due to their extremely small parameter counts, these methods show promise for small datasets. However, their performance is highly dependent on identifying the correct modules to edit and often lacks stability across different datasets. In this paper, we propose Joint Localization and Activation Editing (JoLA), a method that jointly learns (1) which heads in the Transformer to edit (2) whether the intervention should be additive, multiplicative, or both and (3) the intervention parameters themselves - the vectors applied as additive offsets or multiplicative scalings to the head output. Through evaluations on three benchmarks spanning commonsense reasoning, natural language understanding, and natural language generation, we demonstrate that JoLA consistently outperforms existing methods. The code for the method is released at https://github.com/wenlai-lavine/jola.

Wen Lai, Jindřich Libovický, Alexander Fraser

We consider two problems of NMT domain adaptation using meta-learning. First, we want to reach domain robustness, i.e., we want to reach high quality on both domains seen in the training data and unseen domains. Second, we want our systems to be adaptive, i.e., making it possible to finetune systems with just hundreds of in-domain parallel sentences. We study the domain adaptability of meta-learning when improving the domain robustness of the model. In this paper, we propose a novel approach, RMLNMT (Robust Meta-Learning Framework for Neural Machine Translation Domain Adaptation), which improves the robustness of existing meta-learning models. More specifically, we show how to use a domain classifier in curriculum learning and we integrate the word-level domain mixing model into the meta-learning framework with a balanced sampling strategy. Experiments on English$\rightarrow$German and English$\rightarrow$Chinese translation show that RMLNMT improves in terms of both domain robustness and domain adaptability in seen and unseen domains. Our source code is available at https://github.com/lavine-lmu/RMLNMT.

Ryan Cotterell, Thomas Müller, Alexander Fraser et al. · eth-zurich

We present labeled morphological segmentation, an alternative view of morphological processing that unifies several tasks. From an annotation standpoint, we additionally introduce a new hierarchy of morphotactic tagsets. Finally, we develop \modelname, a discriminative morphological segmentation system that, contrary to previous work, explicitly models morphotactics. We show that \textsc{chipmunk} yields improved performance on three tasks for all six languages: (i) morphological segmentation, (ii) stemming and (iii) morphological tag classification. On morphological segmentation, our method shows absolute improvements of 2--6 points $F_1$ over the baseline.

Katharina Hämmerl, Jindřich Libovický, Alexander Fraser

Cross-lingual alignment, the meaningful similarity of representations across languages in multilingual language models, has been an active field of research in recent years. We survey the literature of techniques to improve cross-lingual alignment, providing a taxonomy of methods and summarising insights from throughout the field. We present different understandings of cross-lingual alignment and their limitations. We provide a qualitative summary of results from a large number of surveyed papers. Finally, we discuss how these insights may be applied not only to encoder models, where this topic has been heavily studied, but also to encoder-decoder or even decoder-only models, and argue that an effective trade-off between language-neutral and language-specific information is key.

Felix Friedrich, Katharina Hämmerl, Patrick Schramowski et al.

Text-to-image generation models have recently achieved astonishing results in image quality, flexibility, and text alignment, and are consequently employed in a fast-growing number of applications. Through improvements in multilingual abilities, a larger community now has access to this technology. However, our results show that multilingual models suffer from significant gender biases just as monolingual models do. Furthermore, the natural expectation that multilingual models will provide similar results across languages does not hold up. Instead, there are important differences between languages. We propose a novel benchmark, MAGBIG, intended to foster research on gender bias in multilingual models. We use MAGBIG to investigate the effect of multilingualism on gender bias in T2I models. To this end, we construct multilingual prompts requesting portraits of people with a certain occupation or trait. Our results show that not only do models exhibit strong gender biases but they also behave differently across languages. Furthermore, we investigate prompt engineering strategies, such as indirect, neutral formulations, to mitigate these biases. Unfortunately, these approaches have limited success and result in worse text-to-image alignment. Consequently, we call for more research into diverse representations across languages in image generators, as well as into steerability to address biased model behavior.

Antonia Karamolegkou, Angana Borah, Eunjung Cho et al.

Recent advancements in large language models (LLMs) have unlocked unprecedented possibilities across a range of applications. However, as a community, we believe that the field of Natural Language Processing (NLP) has a growing need to approach deployment with greater intentionality and responsibility. In alignment with the broader vision of AI for Social Good (Tomašev et al., 2020), this paper examines the role of NLP in addressing pressing societal challenges. Through a cross-disciplinary analysis of social goals and emerging risks, we highlight promising research directions and outline challenges that must be addressed to ensure responsible and equitable progress in NLP4SG research.

Yingli Shen, Wen Lai, Shuo Wang et al.

Continued pretraining and instruction tuning on large-scale multilingual data have proven to be effective in scaling large language models (LLMs) to low-resource languages. However, the unaligned nature of such data limits its ability to effectively capture cross-lingual semantics. In contrast, multi-way parallel data, where identical content is aligned across multiple languages, provides stronger cross-lingual consistency and offers greater potential for improving multilingual performance. In this paper, we introduce a large-scale, high-quality multi-way parallel corpus, TED2025, based on TED Talks. The corpus spans 113 languages, with up to 50 languages aligned in parallel, ensuring extensive multilingual coverage. Using this dataset, we investigate best practices for leveraging multi-way parallel data to enhance LLMs, including strategies for continued pretraining, instruction tuning, and the analysis of key influencing factors. Experiments on six multilingual benchmarks show that models trained on multiway parallel data consistently outperform those trained on unaligned multilingual data.

Yingli Shen, Wen Lai, Shuo Wang et al.

The rapid development of multilingual large language models (LLMs) highlights the need for high-quality, diverse, and well-curated multilingual datasets. In this paper, we introduce DCAD-2000 (Data Cleaning as Anomaly Detection), a large-scale multilingual corpus constructed from newly extracted Common Crawl data and existing multilingual sources. DCAD-2000 covers 2,282 languages, 46.72TB of text, and 8.63 billion documents, spanning 155 high- and medium-resource languages and 159 writing scripts. To overcome the limitations of existing data cleaning approaches, which rely on manually designed heuristic thresholds, we reframe data cleaning as an anomaly detection problem. This dynamic filtering paradigm substantially improves data quality by automatically identifying and removing noisy or anomalous content. By fine-tuning LLMs on DCAD-2000, we demonstrate notable improvements in data quality, robustness of the cleaning pipeline, and downstream performance, particularly for low-resource languages across multiple multilingual benchmarks.

Faeze Ghorbanpour, Daryna Dementieva, Alexander Fraser

Despite growing interest in automated hate speech detection, most existing approaches overlook the linguistic diversity of online content. Multilingual instruction-tuned large language models such as LLaMA, Aya, Qwen, and BloomZ offer promising capabilities across languages, but their effectiveness in identifying hate speech through zero-shot and few-shot prompting remains underexplored. This work evaluates LLM prompting-based detection across eight non-English languages, utilizing several prompting techniques and comparing them to fine-tuned encoder models. We show that while zero-shot and few-shot prompting lag behind fine-tuned encoder models on most of the real-world evaluation sets, they achieve better generalization on functional tests for hate speech detection. Our study also reveals that prompt design plays a critical role, with each language often requiring customized prompting techniques to maximize performance.

Prisca Piccirilli, Alexander Fraser, Sabine Schulte im Walde

Metaphor pervades everyday language, allowing speakers to express abstract concepts via concrete domains. While prior work has studied metaphors cognitively and psycholinguistically, large-scale comparisons with literal language remain limited, especially for near-synonymous expressions. We analyze 297 English verb-object pairs (e.g., float idea vs. suggest idea) in ~2M corpus sentences, examining their contextual usage. Using five NLP tools, we extract 2,293 cognitive and linguistic features capturing affective, lexical, syntactic, and discourse-level properties. We address: (i) whether features differ between metaphorical and literal contexts (cross-pair analysis), and (ii) whether individual VO pairs diverge internally (within-pair analysis). Cross-pair results show literal contexts have higher lexical frequency, cohesion, and structural regularity, while metaphorical contexts show greater affective load, imageability, lexical diversity, and constructional specificity. Within-pair analyses reveal substantial heterogeneity, with most pairs showing non-uniform effects. These results suggest no single, consistent distributional pattern that distinguishes metaphorical from literal usage. Instead, differences are largely construction-specific. Overall, large-scale data combined with diverse features provides a fine-grained understanding of metaphor-literal contrasts in VO usage.

Lukas Edman, Alexander Fraser

We describe our strategy for the 2025 edition of the BabyLM Challenge. Our main contribution is that of an improved form of Masked Language Modeling (MLM), which adapts the probabilities of the tokens masked according to the model's ability to predict them. The results show a substantial increase in performance on (Super)GLUE tasks over the standard MLM. We also incorporate sub-token embeddings, finding that this increases the model's morphological generalization capabilities. Our submission beats the baseline in the strict-small track.

Daryna Dementieva, Evgeniya Sukhodolskaya, Alexander Fraser

In the era of social networks and rapid misinformation spread, news analysis remains a critical task. Detecting fake news across multiple languages, particularly beyond English, poses significant challenges. Cross-lingual news comparison offers a promising approach to verify information by leveraging external sources in different languages (Chen and Shu, 2024). However, existing datasets for cross-lingual news analysis (Chen et al., 2022a) were manually curated by journalists and experts, limiting their scalability and adaptability to new languages. In this work, we address this gap by introducing a scalable, explainable crowdsourcing pipeline for cross-lingual news similarity assessment. Using this pipeline, we collected a novel dataset CrossNews-UA of news pairs in Ukrainian as a central language with linguistically and contextually relevant languages-Polish, Russian, and English. Each news pair is annotated for semantic similarity with detailed justifications based on the 4W criteria (Who, What, Where, When). We further tested a range of models, from traditional bag-of-words, Transformer-based architectures to large language models (LLMs). Our results highlight the challenges in multilingual news analysis and offer insights into models performance.

Faeze Ghorbanpour, Alexander Fraser

Large language models (LLMs) increasingly support applications that rely on extended context, from document processing to retrieval-augmented generation. While their long-context capabilities are well studied for reasoning and retrieval, little is known about their behavior in safety-critical scenarios. We evaluate LLMs' sensitivity to harmful content under extended context, varying type (explicit vs. implicit), position (beginning, middle, end), prevalence (0.01-0.50 of the prompt), and context length (600-6000 tokens). Across harmful content categories such as toxic, offensive, and hate speech, with LLaMA-3, Qwen-2.5, and Mistral, we observe similar patterns: performance peaks at moderate harmful prevalence (0.25) but declines when content is very sparse or dominant; recall decreases with increasing context length; harmful sentences at the beginning are generally detected more reliably; and explicit content is more consistently recognized than implicit. These findings provide the first systematic view of how LLMs prioritize and calibrate harmful content in long contexts, highlighting both their emerging strengths and the challenges that remain for safety-critical use.

Lukas Edman, Helmut Schmid, Alexander Fraser

The CUTE benchmark showed that LLMs struggle with character understanding in English. We extend it to more languages with diverse scripts and writing systems, introducing EXECUTE. Our simplified framework allows easy expansion to any language. Tests across multiple LLMs reveal that challenges in other languages are not always on the character level as in English. Some languages show word-level processing issues, some show no issues at all. We also examine sub-character tasks in Chinese, Japanese, and Korean to assess LLMs' understanding of character components.

Daryna Dementieva, Nikolay Babakov, Alexander Fraser

While Ukrainian NLP has seen progress in many texts processing tasks, emotion classification remains an underexplored area with no publicly available benchmark to date. In this work, we introduce EmoBench-UA, the first annotated dataset for emotion detection in Ukrainian texts. Our annotation schema is adapted from the previous English-centric works on emotion detection (Mohammad et al., 2018; Mohammad, 2022) guidelines. The dataset was created through crowdsourcing using the Toloka.ai platform ensuring high-quality of the annotation process. Then, we evaluate a range of approaches on the collected dataset, starting from linguistic-based baselines, synthetic data translated from English, to large language models (LLMs). Our findings highlight the challenges of emotion classification in non-mainstream languages like Ukrainian and emphasize the need for further development of Ukrainian-specific models and training resources.

Faeze Ghorbanpour, Daryna Dementieva, Alexander Fraser

Considering the importance of detecting hateful language, labeled hate speech data is expensive and time-consuming to collect, particularly for low-resource languages. Prior work has demonstrated the effectiveness of cross-lingual transfer learning and data augmentation in improving performance on tasks with limited labeled data. To develop an efficient and scalable cross-lingual transfer learning approach, we leverage nearest-neighbor retrieval to augment minimal labeled data in the target language, thereby enhancing detection performance. Specifically, we assume access to a small set of labeled training instances in the target language and use these to retrieve the most relevant labeled examples from a large multilingual hate speech detection pool. We evaluate our approach on eight languages and demonstrate that it consistently outperforms models trained solely on the target language data. Furthermore, in most cases, our method surpasses the current state-of-the-art. Notably, our approach is highly data-efficient, retrieving as small as 200 instances in some cases while maintaining superior performance. Moreover, it is scalable, as the retrieval pool can be easily expanded, and the method can be readily adapted to new languages and tasks. We also apply maximum marginal relevance to mitigate redundancy and filter out highly similar retrieved instances, resulting in improvements in some languages.

Lukas Edman, Lisa Bylinina, Faeze Ghorbanpour et al.

This paper describes a linguistically-motivated approach to the 2024 edition of the BabyLM Challenge (Warstadt et al. 2023). Rather than pursuing a first language learning (L1) paradigm, we approach the challenge from a second language (L2) learning perspective. In L2 learning, there is a stronger focus on learning explicit linguistic information, such as grammatical notions, definitions of words or different ways of expressing a meaning. This makes L2 learning potentially more efficient and concise. We approximate this using data from Wiktionary, grammar examples either generated by an LLM or sourced from grammar books, and paraphrase data. We find that explicit information about word meaning (in our case, Wiktionary) does not boost model performance, while grammatical information can give a small improvement. The most impactful data ingredient is sentence paraphrases, with our two best models being trained on 1) a mix of paraphrase data and data from the BabyLM pretraining dataset, and 2) exclusively paraphrase data.

Wen Lai, Mohsen Mesgar, Alexander Fraser

To democratize large language models (LLMs) to most natural languages, it is imperative to make these models capable of understanding and generating texts in many languages, in particular low-resource ones. While recent multilingual LLMs demonstrate remarkable performance in such capabilities, these LLMs still support a limited number of human languages due to the lack of training data for low-resource languages. Moreover, these LLMs are not yet aligned with human preference for downstream tasks, which is crucial for the success of LLMs in English. In this paper, we introduce xLLaMA-100 and xBLOOM-100 (collectively xLLMs-100), which scale the multilingual capabilities of LLaMA and BLOOM to 100 languages. To do so, we construct two datasets: a multilingual instruction dataset including 100 languages, which represents the largest language coverage to date, and a cross-lingual human feedback dataset encompassing 30 languages. We perform multilingual instruction tuning on the constructed instruction data and further align the LLMs with human feedback using the DPO algorithm on our cross-lingual human feedback dataset. We evaluate the multilingual understanding and generating capabilities of xLLMs-100 on five multilingual benchmarks. Experimental results show that xLLMs-100 consistently outperforms its peers across the benchmarks by considerable margins, defining a new state-of-the-art multilingual LLM that supports 100 languages.

Yihong Liu, Alexandra Chronopoulou, Hinrich Schütze et al.

Although unsupervised neural machine translation (UNMT) has achieved success in many language pairs, the copying problem, i.e., directly copying some parts of the input sentence as the translation, is common among distant language pairs, especially when low-resource languages are involved. We find this issue is closely related to an unexpected copying behavior during online back-translation (BT). In this work, we propose a simple but effective training schedule that incorporates a language discriminator loss. The loss imposes constraints on the intermediate translation so that the translation is in the desired language. By conducting extensive experiments on different language pairs, including similar and distant, high and low-resource languages, we find that our method alleviates the copying problem, thus improving the translation performance on low-resource languages.

Viktor Hangya, Alexander Fraser

Due to the broad range of social media platforms, the requirements of abusive language detection systems are varied and ever-changing. Already a large set of annotated corpora with different properties and label sets were created, such as hate or misogyny detection, but the form and targets of abusive speech are constantly evolving. Since, the annotation of new corpora is expensive, in this work we leverage datasets we already have, covering a wide range of tasks related to abusive language detection. Our goal is to build models cheaply for a new target label set and/or language, using only a few training examples of the target domain. We propose a two-step approach: first we train our model in a multitask fashion. We then carry out few-shot adaptation to the target requirements. Our experiments show that using already existing datasets and only a few-shots of the target task the performance of models improve both monolingually and across languages. Our analysis also shows that our models acquire a general understanding of abusive language, since they improve the prediction of labels which are present only in the target dataset and can benefit from knowledge about labels which are not directly used for the target task.

Wen Lai, Alexandra Chronopoulou, Alexander Fraser

Despite advances in multilingual neural machine translation (MNMT), we argue that there are still two major challenges in this area: data imbalance and representation degeneration. The data imbalance problem refers to the imbalance in the amount of parallel corpora for all language pairs, especially for long-tail languages (i.e., very low-resource languages). The representation degeneration problem refers to the problem of encoded tokens tending to appear only in a small subspace of the full space available to the MNMT model. To solve these two issues, we propose Bi-ACL, a framework that uses only target-side monolingual data and a bilingual dictionary to improve the performance of the MNMT model. We define two modules, named bidirectional autoencoder and bidirectional contrastive learning, which we combine with an online constrained beam search and a curriculum learning sampling strategy. Extensive experiments show that our proposed method is more effective both in long-tail languages and in high-resource languages. We also demonstrate that our approach is capable of transferring knowledge between domains and languages in zero-shot scenarios.

Irina Bigoulaeva, Viktor Hangya, Iryna Gurevych et al.

The goal of hate speech detection is to filter negative online content aiming at certain groups of people. Due to the easy accessibility of social media platforms it is crucial to protect everyone which requires building hate speech detection systems for a wide range of languages. However, the available labeled hate speech datasets are limited making it problematic to build systems for many languages. In this paper we focus on cross-lingual transfer learning to support hate speech detection in low-resource languages. We leverage cross-lingual word embeddings to train our neural network systems on the source language and apply it to the target language, which lacks labeled examples, and show that good performance can be achieved. We then incorporate unlabeled target language data for further model improvements by bootstrapping labels using an ensemble of different model architectures. Furthermore, we investigate the issue of label imbalance of hate speech datasets, since the high ratio of non-hate examples compared to hate examples often leads to low model performance. We test simple data undersampling and oversampling techniques and show their effectiveness.

Jindřich Libovický, Helmut Schmid, Alexander Fraser

We present a literature and empirical survey that critically assesses the state of the art in character-level modeling for machine translation (MT). Despite evidence in the literature that character-level systems are comparable with subword systems, they are virtually never used in competitive setups in WMT competitions. We empirically show that even with recent modeling innovations in character-level natural language processing, character-level MT systems still struggle to match their subword-based counterparts. Character-level MT systems show neither better domain robustness, nor better morphological generalization, despite being often so motivated. However, we are able to show robustness towards source side noise and that translation quality does not degrade with increasing beam size at decoding time.

Jindřich Libovický, Alexander Fraser

We propose the neural string edit distance model for string-pair matching and string transduction based on learnable string edit distance. We modify the original expectation-maximization learned edit distance algorithm into a differentiable loss function, allowing us to integrate it into a neural network providing a contextual representation of the input. We evaluate on cognate detection, transliteration, and grapheme-to-phoneme conversion, and show that we can trade off between performance and interpretability in a single framework. Using contextual representations, which are difficult to interpret, we match the performance of state-of-the-art string-pair matching models. Using static embeddings and a slightly different loss function, we force interpretability, at the expense of an accuracy drop.

Alexandra Chronopoulou, Dario Stojanovski, Alexander Fraser

Successful methods for unsupervised neural machine translation (UNMT) employ crosslingual pretraining via self-supervision, often in the form of a masked language modeling or a sequence generation task, which requires the model to align the lexical- and high-level representations of the two languages. While cross-lingual pretraining works for similar languages with abundant corpora, it performs poorly in low-resource and distant languages. Previous research has shown that this is because the representations are not sufficiently aligned. In this paper, we enhance the bilingual masked language model pretraining with lexical-level information by using type-level cross-lingual subword embeddings. Empirical results demonstrate improved performance both on UNMT (up to 4.5 BLEU) and bilingual lexicon induction using our method compared to a UNMT baseline.

Alexandra Chronopoulou, Dario Stojanovski, Viktor Hangya et al.

This paper describes the submission of LMU Munich to the WMT 2020 unsupervised shared task, in two language directions, German<->Upper Sorbian. Our core unsupervised neural machine translation (UNMT) system follows the strategy of Chronopoulou et al. (2020), using a monolingual pretrained language generation model (on German) and fine-tuning it on both German and Upper Sorbian, before initializing a UNMT model, which is trained with online backtranslation. Pseudo-parallel data obtained from an unsupervised statistical machine translation (USMT) system is used to fine-tune the UNMT model. We also apply BPE-Dropout to the low resource (Upper Sorbian) data to obtain a more robust system. We additionally experiment with residual adapters and find them useful in the Upper Sorbian->German direction. We explore sampling during backtranslation and curriculum learning to use SMT translations in a more principled way. Finally, we ensemble our best-performing systems and reach a BLEU score of 32.4 on German->Upper Sorbian and 35.2 on Upper Sorbian->German.

Tobias Eder, Viktor Hangya, Alexander Fraser

Good quality monolingual word embeddings (MWEs) can be built for languages which have large amounts of unlabeled text. MWEs can be aligned to bilingual spaces using only a few thousand word translation pairs. For low resource languages training MWEs monolingually results in MWEs of poor quality, and thus poor bilingual word embeddings (BWEs) as well. This paper proposes a new approach for building BWEs in which the vector space of the high resource source language is used as a starting point for training an embedding space for the low resource target language. By using the source vectors as anchors the vector spaces are automatically aligned during training. We experiment on English-German, English-Hiligaynon and English-Macedonian. We show that our approach results not only in improved BWEs and bilingual lexicon induction performance, but also in improved target language MWE quality as measured using monolingual word similarity.

Alexandra Chronopoulou, Dario Stojanovski, Alexander Fraser

Using a language model (LM) pretrained on two languages with large monolingual data in order to initialize an unsupervised neural machine translation (UNMT) system yields state-of-the-art results. When limited data is available for one language, however, this method leads to poor translations. We present an effective approach that reuses an LM that is pretrained only on the high-resource language. The monolingual LM is fine-tuned on both languages and is then used to initialize a UNMT model. To reuse the pretrained LM, we have to modify its predefined vocabulary, to account for the new language. We therefore propose a novel vocabulary extension method. Our approach, RE-LM, outperforms a competitive cross-lingual pretraining model (XLM) in English-Macedonian (En-Mk) and English-Albanian (En-Sq), yielding more than +8.3 BLEU points for all four translation directions.