Dietrich Klakow

David Ifeoluwa Adelani, Jesujoba Oluwadara Alabi, Angela Fan et al. · deepmind, mila

Recent advances in the pre-training of language models leverage large-scale datasets to create multilingual models. However, low-resource languages are mostly left out in these datasets. This is primarily because many widely spoken languages are not well represented on the web and therefore excluded from the large-scale crawls used to create datasets. Furthermore, downstream users of these models are restricted to the selection of languages originally chosen for pre-training. This work investigates how to optimally leverage existing pre-trained models to create low-resource translation systems for 16 African languages. We focus on two questions: 1) How can pre-trained models be used for languages not included in the initial pre-training? and 2) How can the resulting translation models effectively transfer to new domains? To answer these questions, we create a new African news corpus covering 16 languages, of which eight languages are not part of any existing evaluation dataset. We demonstrate that the most effective strategy for transferring both to additional languages and to additional domains is to fine-tune large pre-trained models on small quantities of high-quality translation data.

David Ifeoluwa Adelani, Graham Neubig, Sebastian Ruder et al. · mila

African languages are spoken by over a billion people, but are underrepresented in NLP research and development. The challenges impeding progress include the limited availability of annotated datasets, as well as a lack of understanding of the settings where current methods are effective. In this paper, we make progress towards solutions for these challenges, focusing on the task of named entity recognition (NER). We create the largest human-annotated NER dataset for 20 African languages, and we study the behavior of state-of-the-art cross-lingual transfer methods in an Africa-centric setting, demonstrating that the choice of source language significantly affects performance. We show that choosing the best transfer language improves zero-shot F1 scores by an average of 14 points across 20 languages compared to using English. Our results highlight the need for benchmark datasets and models that cover typologically-diverse African languages.

Jesujoba O. Alabi, David Ifeoluwa Adelani, Marius Mosbach et al.

Multilingual pre-trained language models (PLMs) have demonstrated impressive performance on several downstream tasks for both high-resourced and low-resourced languages. However, there is still a large performance drop for languages unseen during pre-training, especially African languages. One of the most effective approaches to adapt to a new language is \textit{language adaptive fine-tuning} (LAFT) -- fine-tuning a multilingual PLM on monolingual texts of a language using the pre-training objective. However, adapting to a target language individually takes a large disk space and limits the cross-lingual transfer abilities of the resulting models because they have been specialized for a single language. In this paper, we perform \textit{multilingual adaptive fine-tuning} on 17 most-resourced African languages and three other high-resource languages widely spoken on the African continent to encourage cross-lingual transfer learning. To further specialize the multilingual PLM, we removed vocabulary tokens from the embedding layer that corresponds to non-African writing scripts before MAFT, thus reducing the model size by around 50%. Our evaluation on two multilingual PLMs (AfriBERTa and XLM-R) and three NLP tasks (NER, news topic classification, and sentiment classification) shows that our approach is competitive to applying LAFT on individual languages while requiring significantly less disk space. Additionally, we show that our adapted PLM also improves the zero-shot cross-lingual transfer abilities of parameter efficient fine-tuning methods.

Vilém Zouhar, Marius Mosbach, Miaoran Zhang et al. · eth-zurich

Recently neural network based approaches to knowledge-intensive NLP tasks, such as question answering, started to rely heavily on the combination of neural retrievers and readers. Retrieval is typically performed over a large textual knowledge base (KB) which requires significant memory and compute resources, especially when scaled up. On HotpotQA we systematically investigate reducing the size of the KB index by means of dimensionality (sparse random projections, PCA, autoencoders) and numerical precision reduction. Our results show that PCA is an easy solution that requires very little data and is only slightly worse than autoencoders, which are less stable. All methods are sensitive to pre- and post-processing and data should always be centered and normalized both before and after dimension reduction. Finally, we show that it is possible to combine PCA with using 1bit per dimension. Overall we achieve (1) 100$\times$ compression with 75%, and (2) 24$\times$ compression with 92% original retrieval performance.

Dawei Zhu, Michael A. Hedderich, Fangzhou Zhai et al.

Incorrect labels in training data occur when human annotators make mistakes or when the data is generated via weak or distant supervision. It has been shown that complex noise-handling techniques - by modeling, cleaning or filtering the noisy instances - are required to prevent models from fitting this label noise. However, we show in this work that, for text classification tasks with modern NLP models like BERT, over a variety of noise types, existing noisehandling methods do not always improve its performance, and may even deteriorate it, suggesting the need for further investigation. We also back our observations with a comprehensive analysis.

Miaoran Zhang, Marius Mosbach, David Ifeoluwa Adelani et al.

Learning semantically meaningful sentence embeddings is an open problem in natural language processing. In this work, we propose a sentence embedding learning approach that exploits both visual and textual information via a multimodal contrastive objective. Through experiments on a variety of semantic textual similarity tasks, we demonstrate that our approach consistently improves the performance across various datasets and pre-trained encoders. In particular, combining a small amount of multimodal data with a large text-only corpus, we improve the state-of-the-art average Spearman's correlation by 1.7%. By analyzing the properties of the textual embedding space, we show that our model excels in aligning semantically similar sentences, providing an explanation for its improved performance.

Vilém Zouhar, Marius Mosbach, Dietrich Klakow · eth-zurich

Although masked language models are highly performant and widely adopted by NLP practitioners, they can not be easily used for autoregressive language modelling (next word prediction and sequence probability estimation). We present an LSTM-based autoregressive language model which uses prefix embeddings (from a pretrained masked language model) via fusion (e.g. concatenation) to obtain a richer context representation for language modelling. We find that fusion helps reliably in lowering the perplexity (16.74 $\rightarrow$ 15.80), which is even preserved after a transfer to a dataset from a different domain than the training data. We also evaluate the best-performing fusion model by correlating its next word surprisal estimates with human reading times. Contradicting our expectation, and despite the improvement in perplexity overall, the correlation remains the same as for the baseline model. Lastly, while we focus on language models pre-trained on text as the sources for the fusion, our approach can be possibly extended to fuse any information represented as a fixed-size vector into an auto-regressive language model. These include e.g. sentence external information retrieved for a knowledge base or representations of multi-modal encoders.

Awantee Deshpande, Dana Ruiter, Marius Mosbach et al.

Analyzing ethnic or religious bias is important for improving fairness, accountability, and transparency of natural language processing models. However, many techniques rely on human-compiled lists of bias terms, which are expensive to create and are limited in coverage. In this study, we present a fully data-driven pipeline for generating a knowledge graph (KG) of cultural knowledge and stereotypes. Our resulting KG covers 5 religious groups and 5 nationalities and can easily be extended to include more entities. Our human evaluation shows that the majority (59.2%) of non-singleton entries are coherent and complete stereotypes. We further show that performing intermediate masked language model training on the verbalized KG leads to a higher level of cultural awareness in the model and has the potential to increase classification performance on knowledge-crucial samples on a related task, i.e., hate speech detection.

Badr M. Abdullah, Mohammed Maqsood Shaik, Bernd Möbius et al.

Self-supervised representation learning for speech often involves a quantization step that transforms the acoustic input into discrete units. However, it remains unclear how to characterize the relationship between these discrete units and abstract phonetic categories such as phonemes. In this paper, we develop an information-theoretic framework whereby we represent each phonetic category as a distribution over discrete units. We then apply our framework to two different self-supervised models (namely wav2vec 2.0 and XLSR) and use American English speech as a case study. Our study demonstrates that the entropy of phonetic distributions reflects the variability of the underlying speech sounds, with phonetically similar sounds exhibiting similar distributions. While our study confirms the lack of direct, one-to-one correspondence, we find an intriguing, indirect relationship between phonetic categories and discrete units.

Julius Steuer, Marius Mosbach, Dietrich Klakow

Research on the cognitive plausibility of language models (LMs) has so far mostly concentrated on modelling psycholinguistic response variables such as reading times, gaze durations and N400/P600 EEG signals, while mostly leaving out the dimension of what Mahowald et al. (2023) described as formal and functional linguistic competence, and developmental plausibility. We address this gap by training a series of GPT-like language models of different sizes on the strict version of the BabyLM pretraining corpus, evaluating on the challenge tasks (BLiMP, GLUE, MSGS) and an additional reading time prediction task. We find a positive correlation between LM size and performance on all three challenge tasks, with different preferences for model width and depth in each of the tasks. In contrast, a negative correlation was found between LM size and reading time fit of linear mixed-effects models using LM surprisal as a predictor, with the second-smallest LM achieving the largest log-likelihood reduction over a baseline model without surprisal. This suggests that modelling processing effort and linguistic competence may require an approach different from training GPT-like LMs on a developmentally plausible corpus.

Jesujoba O. Alabi, Xuechen Liu, Dietrich Klakow et al.

In this work, we present AfriHuBERT, an extension of mHuBERT-147, a compact self-supervised learning (SSL) model pretrained on 147 languages. While mHuBERT-147 covered 16 African languages, we expand this to 1,226 through continued pretraining on 10K+ hours of speech data from diverse sources, benefiting an African population of over 600M. We evaluate AfriHuBERT on two key speech tasks, Spoken Language Identification (SLID) and Automatic Speech Recognition (ASR), using the FLEURS benchmark. Our results show a +3.6% F1 score improvement for SLID and a -2.1% average Word Error Rate (WER) reduction for ASR over mHuBERT-147, and demonstrates competitiveness with larger SSL models such as MMS and XEUS. Further analysis shows that ASR models trained on AfriHuBERT exhibit improved cross-corpus generalization and are competitive in extremely low-resource ASR scenarios.

Alexander Blatt, Martin Kocour, Karel Veselý et al.

Air traffic control (ATC) relies on communication via speech between pilot and air-traffic controller (ATCO). The call-sign, as unique identifier for each flight, is used to address a specific pilot by the ATCO. Extracting the call-sign from the communication is a challenge because of the noisy ATC voice channel and the additional noise introduced by the receiver. A low signal-to-noise ratio (SNR) in the speech leads to high word error rate (WER) transcripts. We propose a new call-sign recognition and understanding (CRU) system that addresses this issue. The recognizer is trained to identify call-signs in noisy ATC transcripts and convert them into the standard International Civil Aviation Organization (ICAO) format. By incorporating surveillance information, we can multiply the call-sign accuracy (CSA) up to a factor of four. The introduced data augmentation adds additional performance on high WER transcripts and allows the adaptation of the model to unseen airspaces.

Idris Akinade, Jesujoba Alabi, David Adelani et al.

This paper investigates the performance of massively multilingual neural machine translation (NMT) systems in translating Yorùbá greetings ($\varepsilon$ kú [MASK]), which are a big part of Yorùbá language and culture, into English. To evaluate these models, we present IkiniYorùbá, a Yorùbá-English translation dataset containing some Yorùbá greetings, and sample use cases. We analysed the performance of different multilingual NMT systems including Google and NLLB and show that these models struggle to accurately translate Yorùbá greetings into English. In addition, we trained a Yorùbá-English model by finetuning an existing NMT model on the training split of IkiniYorùbá and this achieved better performance when compared to the pre-trained multilingual NMT models, although they were trained on a large volume of data.

Dawei Zhu, Xiaoyu Shen, Michael A. Hedderich et al.

Training deep neural networks (DNNs) under weak supervision has attracted increasing research attention as it can significantly reduce the annotation cost. However, labels from weak supervision can be noisy, and the high capacity of DNNs enables them to easily overfit the label noise, resulting in poor generalization. Recent methods leverage self-training to build noise-resistant models, in which a teacher trained under weak supervision is used to provide highly confident labels for teaching the students. Nevertheless, the teacher derived from such frameworks may have fitted a substantial amount of noise and therefore produce incorrect pseudo-labels with high confidence, leading to severe error propagation. In this work, we propose Meta Self-Refinement (MSR), a noise-resistant learning framework, to effectively combat label noise from weak supervision. Instead of relying on a fixed teacher trained with noisy labels, we encourage the teacher to refine its pseudo-labels. At each training step, MSR performs a meta gradient descent on the current mini-batch to maximize the student performance on a clean validation set. Extensive experimentation on eight NLP benchmarks demonstrates that MSR is robust against label noise in all settings and outperforms state-of-the-art methods by up to 11.4% in accuracy and 9.26% in F1 score.

Lukas Lange, Jannik Strötgen, Heike Adel et al.

The detection and normalization of temporal expressions is an important task and preprocessing step for many applications. However, prior work on normalization is rule-based, which severely limits the applicability in real-world multilingual settings, due to the costly creation of new rules. We propose a novel neural method for normalizing temporal expressions based on masked language modeling. Our multilingual method outperforms prior rule-based systems in many languages, and in particular, for low-resource languages with performance improvements of up to 33 F1 on average compared to the state of the art.

Dana Ruiter, Thomas Kleinbauer, Cristina España-Bonet et al.

Recent research on style transfer takes inspiration from unsupervised neural machine translation (UNMT), learning from large amounts of non-parallel data by exploiting cycle consistency loss, back-translation, and denoising autoencoders. By contrast, the use of self-supervised NMT (SSNMT), which leverages (near) parallel instances hidden in non-parallel data more efficiently than UNMT, has not yet been explored for style transfer. In this paper we present a novel Self-Supervised Style Transfer (3ST) model, which augments SSNMT with UNMT methods in order to identify and efficiently exploit supervisory signals in non-parallel social media posts. We compare 3ST with state-of-the-art (SOTA) style transfer models across civil rephrasing, formality and polarity tasks. We show that 3ST is able to balance the three major objectives (fluency, content preservation, attribute transfer accuracy) the best, outperforming SOTA models on averaged performance across their tested tasks in automatic and human evaluation.

Ali Davody, David Ifeoluwa Adelani, Thomas Kleinbauer et al.

Transferring knowledge from one domain to another is of practical importance for many tasks in natural language processing, especially when the amount of available data in the target domain is limited. In this work, we propose a novel few-shot approach to domain adaptation in the context of Named Entity Recognition (NER). We propose a two-step approach consisting of a variable base module and a template module that leverages the knowledge captured in pre-trained language models with the help of simple descriptive patterns. Our approach is simple yet versatile and can be applied in few-shot and zero-shot settings. Evaluating our lightweight approach across a number of different datasets shows that it can boost the performance of state-of-the-art baselines by 2-5% F1-score.

Vagrant Gautam, Julius Steuer, Eileen Bingert et al.

While measuring bias and robustness in coreference resolution are important goals, such measurements are only as good as the tools we use to measure them. Winogender Schemas (Rudinger et al., 2018) are an influential dataset proposed to evaluate gender bias in coreference resolution, but a closer look reveals issues with the data that compromise its use for reliable evaluation, including treating different pronominal forms as equivalent, violations of template constraints, and typographical errors. We identify these issues and fix them, contributing a new dataset: WinoPron. Using WinoPron, we evaluate two state-of-the-art supervised coreference resolution systems, SpanBERT, and five sizes of FLAN-T5, and demonstrate that accusative pronouns are harder to resolve for all models. We also propose a new method to evaluate pronominal bias in coreference resolution that goes beyond the binary. With this method, we also show that bias characteristics vary not just across pronoun sets (e.g., he vs. she), but also across surface forms of those sets (e.g., him vs. his).

Julius Steuer, Badr Abdullah, Johann-Mattis List et al.

We present a cross-linguistic study that aims to quantify vowel harmony using data-driven computational modeling. Concretely, we define an information-theoretic measure of harmonicity based on the predictability of vowels in a natural language lexicon, which we estimate using phoneme-level language models (PLMs). Prior quantitative studies have relied heavily on inflected word-forms in the analysis of vowel harmony. We instead train our models using cross-linguistically comparable lemma forms with little or no inflection, which enables us to cover more under-studied languages. Training data for our PLMs consists of word lists with a maximum of 1000 entries per language. Despite the fact that the data we employ are substantially smaller than previously used corpora, our experiments demonstrate the neural PLMs capture vowel harmony patterns in a set of languages that exhibit this phenomenon. Our work also demonstrates that word lists are a valuable resource for typological research, and offers new possibilities for future studies on low-resource, under-studied languages.

Badr M. Abdullah, Dietrich Klakow

Acoustic word embeddings (AWEs) are vector representations such that different acoustic exemplars of the same word are projected nearby in the embedding space. In addition to their use in speech technology applications such as spoken term discovery and keyword spotting, AWE models have been adopted as models of spoken-word processing in several cognitively motivated studies and have been shown to exhibit human-like performance in some auditory processing tasks. Nevertheless, the representational geometry of AWEs remains an under-explored topic that has not been studied in the literature. In this paper, we take a closer analytical look at AWEs learned from English speech and study how the choice of the learning objective and the architecture shapes their representational profile. To this end, we employ a set of analytic techniques from machine learning and neuroscience in three different analyses: embedding space uniformity, word discriminability, and representational consistency. Our main findings highlight the prominent role of the learning objective on shaping the representation profile compared to the model architecture.

Badr M. Abdullah, Bernd Möbius, Dietrich Klakow

Models of acoustic word embeddings (AWEs) learn to map variable-length spoken word segments onto fixed-dimensionality vector representations such that different acoustic exemplars of the same word are projected nearby in the embedding space. In addition to their speech technology applications, AWE models have been shown to predict human performance on a variety of auditory lexical processing tasks. Current AWE models are based on neural networks and trained in a bottom-up approach that integrates acoustic cues to build up a word representation given an acoustic or symbolic supervision signal. Therefore, these models do not leverage or capture high-level lexical knowledge during the learning process. In this paper, we propose a multi-task learning model that incorporates top-down lexical knowledge into the training procedure of AWEs. Our model learns a mapping between the acoustic input and a lexical representation that encodes high-level information such as word semantics in addition to bottom-up form-based supervision. We experiment with three languages and demonstrate that incorporating lexical knowledge improves the embedding space discriminability and encourages the model to better separate lexical categories.

Vagrant Gautam, Miaoran Zhang, Dietrich Klakow

If a question cannot be answered with the available information, robust systems for question answering (QA) should know _not_ to answer. One way to build QA models that do this is with additional training data comprised of unanswerable questions, created either by employing annotators or through automated methods for unanswerable question generation. To show that the model complexity of existing automated approaches is not justified, we examine a simpler data augmentation method for unanswerable question generation in English: performing antonym and entity swaps on answerable questions. Compared to the prior state-of-the-art, data generated with our training-free and lightweight strategy results in better models (+1.6 F1 points on SQuAD 2.0 data with BERT-large), and has higher human-judged relatedness and readability. We quantify the raw benefits of our approach compared to no augmentation across multiple encoder models, using different amounts of generated data, and also on TydiQA-MinSpan data (+9.3 F1 points with BERT-large). Our results establish swaps as a simple but strong baseline for future work.

Dana Ruiter, Liane Reiners, Ashwin Geet D'Sa et al.

Even though hate speech (HS) online has been an important object of research in the last decade, most HS-related corpora over-simplify the phenomenon of hate by attempting to label user comments as "hate" or "neutral". This ignores the complex and subjective nature of HS, which limits the real-life applicability of classifiers trained on these corpora. In this study, we present the M-Phasis corpus, a corpus of ~9k German and French user comments collected from migration-related news articles. It goes beyond the "hate"-"neutral" dichotomy and is instead annotated with 23 features, which in combination become descriptors of various types of speech, ranging from critical comments to implicit and explicit expressions of hate. The annotations are performed by 4 native speakers per language and achieve high (0.77 <= k <= 1) inter-annotator agreements. Besides describing the corpus creation and presenting insights from a content, error and domain analysis, we explore its data characteristics by training several classification baselines.

Florian Dietz, Sophie Fellenz, Dietrich Klakow et al.

Neural Networks are notoriously difficult to inspect. We introduce comgra, an open source python library for use with PyTorch. Comgra extracts data about the internal activations of a model and organizes it in a GUI (graphical user interface). It can show both summary statistics and individual data points, compare early and late stages of training, focus on individual samples of interest, and visualize the flow of the gradient through the network. This makes it possible to inspect the model's behavior from many different angles and save time by rapidly testing different hypotheses without having to rerun it. Comgra has applications for debugging, neural architecture design, and mechanistic interpretability. We publish our library through Python Package Index (PyPI) and provide code, documentation, and tutorials at https://github.com/FlorianDietz/comgra.

Aravind Krishnan, Jesujoba Alabi, Dietrich Klakow

Large pre-trained language models (PLMs) have shown remarkable performance across various natural language understanding (NLU) tasks, particularly in low-resource settings. Nevertheless, their potential in Automatic Speech Recognition (ASR) remains largely unexplored. This study investigates the potential usage of PLMs for language modelling in ASR. We compare the application of large-scale text sampling and probability conversion for approximating GPT-2 into an n-gram model. Furthermore, we introduce a vocabulary-restricted decoding method for random sampling, and evaluate the effects of domain difficulty and data size on the usability of generated text. Our findings across eight domain-specific corpora support the use of sampling-based approximation and show that interpolating with a large sampled corpus improves test perplexity over a baseline trigram by 15%. Our vocabulary-restricted decoding method pushes this improvement further by 5% in domain-specific settings.

Anupama Chingacham, Vera Demberg, Dietrich Klakow

In noisy environments, speech can be hard to understand for humans. Spoken dialog systems can help to enhance the intelligibility of their output, either by modifying the speech synthesis (e.g., imitate Lombard speech) or by optimizing the language generation. We here focus on the second type of approach, by which an intended message is realized with words that are more intelligible in a specific noisy environment. By conducting a speech perception experiment, we created a dataset of 900 paraphrases in babble noise, perceived by native English speakers with normal hearing. We find that careful selection of paraphrases can improve intelligibility by 33% at SNR -5 dB. Our analysis of the data shows that the intelligibility differences between paraphrases are mainly driven by noise-robust acoustic cues. Furthermore, we propose an intelligibility-aware paraphrase ranking model, which outperforms baseline models with a relative improvement of 31.37% at SNR -5 dB.

Dawei Zhu, Michael A. Hedderich, Fangzhou Zhai et al.

For high-resource languages like English, text classification is a well-studied task. The performance of modern NLP models easily achieves an accuracy of more than 90% in many standard datasets for text classification in English (Xie et al., 2019; Yang et al., 2019; Zaheer et al., 2020). However, text classification in low-resource languages is still challenging due to the lack of annotated data. Although methods like weak supervision and crowdsourcing can help ease the annotation bottleneck, the annotations obtained by these methods contain label noise. Models trained with label noise may not generalize well. To this end, a variety of noise-handling techniques have been proposed to alleviate the negative impact caused by the errors in the annotations (for extensive surveys see (Hedderich et al., 2021; Algan & Ulusoy, 2021)). In this work, we experiment with a group of standard noisy-handling methods on text classification tasks with noisy labels. We study both simulated noise and realistic noise induced by weak supervision. Moreover, we find task-adaptive pre-training techniques (Gururangan et al., 2020) are beneficial for learning with noisy labels.

Michael A. Hedderich, Jonas Fischer, Dietrich Klakow et al.

State-of-the-art NLP methods achieve human-like performance on many tasks, but make errors nevertheless. Characterizing these errors in easily interpretable terms gives insight into whether a classifier is prone to making systematic errors, but also gives a way to act and improve the classifier. We propose to discover those patterns of tokens that distinguish correct and erroneous predictions as to obtain global and interpretable descriptions for arbitrary NLP classifiers. We formulate the problem of finding a succinct and non-redundant set of such patterns in terms of the Minimum Description Length principle. Through an extensive set of experiments, we show that our method, Premise, performs well in practice. Unlike existing solutions, it recovers ground truth, even on highly imbalanced data over large vocabularies. In VQA and NER case studies, we confirm that it gives clear and actionable insight into the systematic errors made by NLP classifiers.

Xinghao Chen, Zhijing Sun, Wenjin Guo et al.

Large Language Models (LLMs) excel in reasoning tasks through Chain-of-Thought (CoT) prompting. However, CoT prompting greatly increases computational demands, which has prompted growing interest in distilling CoT capabilities into Small Language Models (SLMs). This study systematically examines the factors influencing CoT distillation, including the choice of granularity, format and teacher model. Through experiments involving four teacher models and seven student models across seven mathematical and commonsense reasoning datasets, we uncover three key findings: (1) Unlike LLMs, SLMs exhibit a non-monotonic relationship with granularity, with stronger models benefiting from finer-grained reasoning and weaker models performing better with simpler CoT supervision; (2) CoT format significantly impacts LLMs but has minimal effect on SLMs, likely due to their reliance on supervised fine-tuning rather than pretraining preferences; (3) Stronger teacher models do NOT always produce better student models, as diversity and complexity in CoT supervision can outweigh accuracy alone. These findings emphasize the need to tailor CoT strategies to specific student model, offering actionable insights for optimizing CoT distillation in SLMs. The code and datasets are available at https://github.com/EIT-NLP/Distilling-CoT-Reasoning.

Edward Bayes, Israel Abebe Azime, Jesujoba O. Alabi et al.

Evaluations of Large Language Models (LLMs) on knowledge-intensive tasks and factual accuracy often focus on high-resource languages primarily because datasets for low-resource languages (LRLs) are scarce. In this paper, we present Uhura -- a new benchmark that focuses on two tasks in six typologically-diverse African languages, created via human translation of existing English benchmarks. The first dataset, Uhura-ARC-Easy, is composed of multiple-choice science questions. The second, Uhura-TruthfulQA, is a safety benchmark testing the truthfulness of models on topics including health, law, finance, and politics. We highlight the challenges creating benchmarks with highly technical content for LRLs and outline mitigation strategies. Our evaluation reveals a significant performance gap between proprietary models such as GPT-4o and o1-preview, and Claude models, and open-source models like Meta's LLaMA and Google's Gemma. Additionally, all models perform better in English than in African languages. These results indicate that LMs struggle with answering scientific questions and are more prone to generating false claims in low-resource African languages. Our findings underscore the necessity for continuous improvement of multilingual LM capabilities in LRL settings to ensure safe and reliable use in real-world contexts. We open-source the Uhura Benchmark and Uhura Platform to foster further research and development in NLP for LRLs.

Atnafu Lambebo Tonja, Israel Abebe Azime, Tadesse Destaw Belay et al.

Large language models (LLMs) have gained popularity recently due to their outstanding performance in various downstream Natural Language Processing (NLP) tasks. However, low-resource languages are still lagging behind current state-of-the-art (SOTA) developments in the field of NLP due to insufficient resources to train LLMs. Ethiopian languages exhibit remarkable linguistic diversity, encompassing a wide array of scripts, and are imbued with profound religious and cultural significance. This paper introduces EthioLLM -- multilingual large language models for five Ethiopian languages (Amharic, Ge'ez, Afan Oromo, Somali, and Tigrinya) and English, and Ethiobenchmark -- a new benchmark dataset for various downstream NLP tasks. We evaluate the performance of these models across five downstream NLP tasks. We open-source our multilingual language models, new benchmark datasets for various downstream tasks, and task-specific fine-tuned language models and discuss the performance of the models. Our dataset and models are available at the https://huggingface.co/EthioNLP repository.

Israel Abebe Azime, Atnafu Lambebo Tonja, Tadesse Destaw Belay et al.

With the rapid development of evaluation datasets to assess LLMs understanding across a wide range of subjects and domains, identifying a suitable language understanding benchmark has become increasingly challenging. In this work, we explore LLM evaluation challenges for low-resource language understanding and introduce \proverbeval, LLM evaluation benchmark for low-resource languages, focusing on low-resource language understanding in culture-specific scenarios. We benchmark various LLMs and explore factors that create variability in the benchmarking process. We observed performance variances of up to 50\%, depending on the order in which answer choices were presented in multiple-choice tasks. Native language proverb descriptions significantly improve tasks such as proverb generation, contributing to improved outcomes. Additionally, monolingual evaluations consistently outperformed their cross-lingual counterparts in generation tasks. We argue that special attention must be given to the order of choices, the choice of prompt language, task variability, and generation tasks when creating LLM evaluation benchmarks. Evaluation data available at https://huggingface.co/datasets/israel/ProverbEval, evaluation code https://github.com/EthioNLP/EthioProverbEval.

Tural Mammadov, Dietrich Klakow, Alexander Koller et al.

We introduce Modelizer - a novel framework that, given a black-box program, learns a model from its input/output behavior using neural machine translation algorithms. The resulting model mocks the original program: Given an input, the model predicts the output that would have been produced by the program. However, the model is also reversible - that is, the model can predict the input that would have produced a given output. Finally, the model is differentiable and can be efficiently restricted to predict only a certain aspect of the program behavior. Modelizer uses grammars to synthesize and inputs and unsupervised tokenizers to decompose the resulting outputs, allowing it to learn sequence-to-sequence associations between token streams. Other than input grammars, Modelizer only requires the ability to execute the program. The resulting models are small, requiring fewer than 6.3 million parameters for languages such as Markdown or HTML; and they are accurate, achieving up to 95.4% accuracy and a BLEU score of 0.98 with standard error 0.04 in mocking real-world applications. As it learns from and predicts executions rather than code, Modelizer departs from the LLM-centric research trend, opening new opportunities for program-specific models that are fully tuned towards individual programs. Indeed, we foresee several applications of these models, especially as the output of the program can be any aspect of program behavior. Beyond mocking and predicting program behavior, the models can also synthesize inputs that are likely to produce a particular behavior, such as failures or coverage, thus assisting in program understanding and maintenance.

Hao Yu, Jesujoba O. Alabi, Andiswa Bukula et al.

Slot-filling and intent detection are well-established tasks in Conversational AI. However, current large-scale benchmarks for these tasks often exclude evaluations of low-resource languages and rely on translations from English benchmarks, thereby predominantly reflecting Western-centric concepts. In this paper, we introduce Injongo -- a multicultural, open-source benchmark dataset for 16 African languages with utterances generated by native speakers across diverse domains, including banking, travel, home, and dining. Through extensive experiments, we benchmark the fine-tuning multilingual transformer models and the prompting large language models (LLMs), and show the advantage of leveraging African-cultural utterances over Western-centric utterances for improving cross-lingual transfer from the English language. Experimental results reveal that current LLMs struggle with the slot-filling task, with GPT-4o achieving an average performance of 26 F1-score. In contrast, intent detection performance is notably better, with an average accuracy of 70.6%, though it still falls behind the fine-tuning baselines. Compared to the English language, GPT-4o and fine-tuning baselines perform similarly on intent detection, achieving an accuracy of approximately 81%. Our findings suggest that the performance of LLMs is still behind for many low-resource African languages, and more work is needed to further improve their downstream performance.

Anupama Chingacham, Miaoran Zhang, Vera Demberg et al.

Large Language Models (LLMs) can generate text by transferring style attributes like formality resulting in formal or informal text. However, instructing LLMs to generate text that when spoken, is more intelligible in an acoustically difficult environment, is an under-explored topic. We conduct the first study to evaluate LLMs on a novel task of generating acoustically intelligible paraphrases for better human speech perception in noise. Our experiments in English demonstrated that with standard prompting, LLMs struggle to control the non-textual attribute, i.e., acoustic intelligibility, while efficiently capturing the desired textual attributes like semantic equivalence. To remedy this issue, we propose a simple prompting approach, prompt-and-select, which generates paraphrases by decoupling the desired textual and non-textual attributes in the text generation pipeline. Our approach resulted in a 40% relative improvement in human speech perception, by paraphrasing utterances that are highly distorted in a listening condition with babble noise at a signal-to-noise ratio (SNR) -5 dB. This study reveals the limitation of LLMs in capturing non-textual attributes, and our proposed method showcases the potential of using LLMs for better human speech perception in noise.

Florian Dietz, Dietrich Klakow

We explore the hypothesis that poor compositional generalization in neural networks is caused by difficulties with learning effective routing. To solve this problem, we propose the concept of block-operations, which is based on splitting all activation tensors in the network into uniformly sized blocks and using an inductive bias to encourage modular routing and modification of these blocks. Based on this concept we introduce the Multiplexer, a new architectural component that enhances the Feed Forward Neural Network (FNN). We experimentally confirm that Multiplexers exhibit strong compositional generalization. On both a synthetic and a realistic task our model was able to learn the underlying process behind the task, whereas both FNNs and Transformers were only able to learn heuristic approximations. We propose as future work to use the principles of block-operations to improve other existing architectures.

Yunuo Liu, Dawei Zhu, Zena Al-Khalili et al.

We present PricingLogic, the first benchmark that probes whether Large Language Models(LLMs) can reliably automate tourism-related prices when multiple, overlapping fare rules apply. Travel agencies are eager to offload this error-prone task onto AI systems; however, deploying LLMs without verified reliability could result in significant financial losses and erode customer trust. PricingLogic comprises 300 natural-language questions based on booking requests derived from 42 real-world pricing policies, spanning two levels of difficulty: (i) basic customer-type pricing and (ii)bundled-tour calculations involving interacting discounts. Evaluations of a line of LLMs reveal a steep performance drop on the harder tier,exposing systematic failures in rule interpretation and arithmetic reasoning.These results highlight that, despite their general capabilities, today's LLMs remain unreliable in revenue-critical applications without further safeguards or domain adaptation. Our code and dataset are available at https://github.com/EIT-NLP/PricingLogic.

Zena Al-Khalili, Nick Howell, Dietrich Klakow

Assisting LLMs with code generation improved their performance on mathematical reasoning tasks. However, the evaluation of code-assisted LLMs is generally restricted to execution correctness, lacking a rigorous evaluation of their generated programs. In this work, we bridge this gap by conducting an in-depth analysis of code-assisted LLMs generated programs in response to math reasoning tasks, with a focus on evaluating the soundness of the underlying reasoning processes. For this purpose, we assess the generations of five LLMs, on several math datasets, both manually and automatically, and propose a taxonomy of generated programs based on their logical soundness. Our findings show that the capabilities of models significantly impact the logic implemented to solve the problem. Closed-source LLMs ground their programs in mathematical concepts, whereas open-source models often resort to unsound reasoning, relying on memorized information and exhaustive searches. Furthermore, increasing the difficulty of problems decreases sound generations for all models, revealing a critical shortcoming of LLMs on complex mathematics, contrary to what accuracy metrics suggest. Our work highlights the need for more holistic evaluations of code-assisted LLMs beyond execution accuracy metrics, toward a better understanding of LLMs' limits in the math domain.

Marius Mosbach, Maksym Andriushchenko, Dietrich Klakow

Fine-tuning pre-trained transformer-based language models such as BERT has become a common practice dominating leaderboards across various NLP benchmarks. Despite the strong empirical performance of fine-tuned models, fine-tuning is an unstable process: training the same model with multiple random seeds can result in a large variance of the task performance. Previous literature (Devlin et al., 2019; Lee et al., 2020; Dodge et al., 2020) identified two potential reasons for the observed instability: catastrophic forgetting and small size of the fine-tuning datasets. In this paper, we show that both hypotheses fail to explain the fine-tuning instability. We analyze BERT, RoBERTa, and ALBERT, fine-tuned on commonly used datasets from the GLUE benchmark, and show that the observed instability is caused by optimization difficulties that lead to vanishing gradients. Additionally, we show that the remaining variance of the downstream task performance can be attributed to differences in generalization where fine-tuned models with the same training loss exhibit noticeably different test performance. Based on our analysis, we present a simple but strong baseline that makes fine-tuning BERT-based models significantly more stable than the previously proposed approaches. Code to reproduce our results is available online: https://github.com/uds-lsv/bert-stable-fine-tuning.

Tadesse Destaw Belay, Israel Abebe Azime, Abinew Ali Ayele et al.

Large Language Models (LLMs) show promising learning and reasoning abilities. Compared to other NLP tasks, multilingual and multi-label emotion evaluation tasks are under-explored in LLMs. In this paper, we present EthioEmo, a multi-label emotion classification dataset for four Ethiopian languages, namely, Amharic (amh), Afan Oromo (orm), Somali (som), and Tigrinya (tir). We perform extensive experiments with an additional English multi-label emotion dataset from SemEval 2018 Task 1. Our evaluation includes encoder-only, encoder-decoder, and decoder-only language models. We compare zero and few-shot approaches of LLMs to fine-tuning smaller language models. The results show that accurate multi-label emotion classification is still insufficient even for high-resource languages such as English, and there is a large gap between the performance of high-resource and low-resource languages. The results also show varying performance levels depending on the language and model type. EthioEmo is available publicly to further improve the understanding of emotions in language models and how people convey emotions through various languages.

Dawei Zhu, Pinzhen Chen, Miaoran Zhang et al.

Traditionally, success in multilingual machine translation can be attributed to three key factors in training data: large volume, diverse translation directions, and high quality. In the current practice of fine-tuning large language models (LLMs) for translation, we revisit the importance of these factors. We find that LLMs display strong translation capability after being fine-tuned on as few as 32 parallel sentences and that fine-tuning on a single translation direction enables translation in multiple directions. However, the choice of direction is critical: fine-tuning LLMs with only English on the target side can lead to task misinterpretation, which hinders translation into non-English languages. Problems also arise when noisy synthetic data is placed on the target side, especially when the target language is well-represented in LLM pre-training. Yet interestingly, synthesized data in an under-represented language has a less pronounced effect. Our findings suggest that when adapting LLMs to translation, the requirement on data quantity can be eased but careful considerations are still crucial to prevent an LLM from exploiting unintended data biases.

Omer Moussa, Dietrich Klakow, Mariya Toneva · cmu

Speech language models align with human brain responses to natural language to an impressive degree. However, current models rely heavily on low-level speech features, indicating they lack brain-relevant semantics which limits their utility as model organisms of semantic processing in the brain. In this work, we address this limitation by inducing brain-relevant bias directly into the models via fine-tuning with fMRI recordings of people listening to natural stories, a process we name brain-tuning. After testing it on 3 different pretrained model families, we show that brain-tuning not only improves overall alignment with new brain recordings in semantic language regions, but also reduces the reliance on low-level speech features for this alignment. Excitingly, we further show that brain-tuning leads to 1) consistent improvements in performance on a range of downstream tasks and 2) a representational space with increased semantic preference. Our results provide converging evidence, for the first time, that incorporating brain signals into the training of language models improves the models' semantic understanding.

Jesujoba O. Alabi, Marius Mosbach, Matan Eyal et al. · deepmind

We analyze the operation of transformer language adapters, which are small modules trained on top of a frozen language model to adapt its predictions to new target languages. We show that adapted predictions mostly evolve in the source language the model was trained on, while the target language becomes pronounced only in the very last layers of the model. Moreover, the adaptation process is gradual and distributed across layers, where it is possible to skip small groups of adapters without decreasing adaptation performance. Last, we show that adapters operate on top of the model's frozen representation space while largely preserving its structure, rather than on an 'isolated' subspace. Our findings provide a deeper view into the adaptation process of language models to new languages, showcasing the constraints imposed on it by the underlying model and introduces practical implications to enhance its efficiency.

Miaoran Zhang, Vagrant Gautam, Mingyang Wang et al.

In-context learning is a popular inference strategy where large language models solve a task using only a few labeled demonstrations without needing any parameter updates. Although there have been extensive studies on English in-context learning, multilingual in-context learning remains under-explored, and we lack an in-depth understanding of the role of demonstrations in this context. To address this gap, we conduct a multidimensional analysis of multilingual in-context learning, experimenting with 5 models from different model families, 9 datasets covering classification and generation tasks, and 56 typologically diverse languages. Our results reveal that the effectiveness of demonstrations varies significantly across models, tasks, and languages. We also find that strong instruction-following models including Llama 2-Chat, GPT-3.5, and GPT-4 are largely insensitive to the quality of demonstrations. Instead, a carefully crafted template often eliminates the benefits of demonstrations for some tasks and languages altogether. These findings show that the importance of demonstrations might be overestimated. Our work highlights the need for granular evaluation across multiple axes towards a better understanding of in-context learning.

Vagrant Gautam, Eileen Bingert, Dawei Zhu et al.

Robust, faithful and harm-free pronoun use for individuals is an important goal for language model development as their use increases, but prior work tends to study only one or two of these characteristics at a time. To measure progress towards the combined goal, we introduce the task of pronoun fidelity: given a context introducing a co-referring entity and pronoun, the task is to reuse the correct pronoun later. We present RUFF, a carefully-designed dataset of over 5 million instances to measure robust pronoun fidelity in English, and we evaluate 37 model variants from nine popular families, across architectures (encoder-only, decoder-only and encoder-decoder) and scales (11M-70B parameters). When an individual is introduced with a pronoun, models can mostly faithfully reuse this pronoun in the next sentence, but they are significantly worse with she/her/her, singular they and neopronouns. Moreover, models are easily distracted by non-adversarial sentences discussing other people; even one sentence with a distractor pronoun causes accuracy to drop on average by 34 percentage points. Our results show that pronoun fidelity is not robust, in a simple, naturalistic setting where humans achieve nearly 100% accuracy. We encourage researchers to bridge the gaps we find and to carefully evaluate reasoning in settings where superficial repetition might inflate perceptions of model performance.

Mohammed Maqsood Shaik, Dietrich Klakow, Badr M. Abdullah · pku

Pre-trained Transformer-based speech models have shown striking performance when fine-tuned on various downstream tasks such as automatic speech recognition and spoken language identification (SLID). However, the problem of domain mismatch remains a challenge in this area, where the domain of the pre-training data might differ from that of the downstream labeled data used for fine-tuning. In multilingual tasks such as SLID, the pre-trained speech model may not support all the languages in the downstream task. To address this challenge, we propose self-supervised adaptive pre-training (SAPT) to adapt the pre-trained model to the target domain and languages of the downstream task. We apply SAPT to the XLSR-128 model and investigate the effectiveness of this approach for the SLID task. First, we demonstrate that SAPT improves XLSR performance on the FLEURS benchmark with substantial gains up to 40.1% for under-represented languages. Second, we apply SAPT on four different datasets in a few-shot learning setting, showing that our approach improves the sample efficiency of XLSR during fine-tuning. Our experiments provide strong empirical evidence that continual adaptation via self-supervision improves downstream performance for multilingual speech models.

Dawei Zhu, Sony Trenous, Xiaoyu Shen et al.

Recent research has shown that large language models (LLMs) can achieve remarkable translation performance through supervised fine-tuning (SFT) using only a small amount of parallel data. However, SFT simply instructs the model to imitate the reference translations at the token level, making it vulnerable to the noise present in the references. Hence, the assistance from SFT often reaches a plateau once the LLMs have achieved a certain level of translation capability, and further increasing the size of parallel data does not provide additional benefits. To overcome this plateau associated with imitation-based SFT, we propose a preference-based approach built upon the Plackett-Luce model. The objective is to steer LLMs towards a more nuanced understanding of translation preferences from a holistic view, while also being more resilient in the absence of gold translations. We further build a dataset named MAPLE to verify the effectiveness of our approach, which includes multiple translations of varying quality for each source sentence. Extensive experiments demonstrate the superiority of our approach in "breaking the plateau" across diverse LLMs and test settings. Our in-depth analysis underscores the pivotal role of diverse translations and accurate preference scores in the success of our approach.

Miaoran Zhang, Mingyang Wang, Jesujoba O. Alabi et al.

This paper presents our system developed for the SemEval-2024 Task 1: Semantic Textual Relatedness for African and Asian Languages. The shared task aims at measuring the semantic textual relatedness between pairs of sentences, with a focus on a range of under-represented languages. In this work, we propose using machine translation for data augmentation to address the low-resource challenge of limited training data. Moreover, we apply task-adaptive pre-training on unlabeled task data to bridge the gap between pre-training and task adaptation. For model training, we investigate both full fine-tuning and adapter-based tuning, and adopt the adapter framework for effective zero-shot cross-lingual transfer. We achieve competitive results in the shared task: our system performs the best among all ranked teams in both subtask A (supervised learning) and subtask C (cross-lingual transfer).

Jesujoba O. Alabi, Michael A. Hedderich, David Ifeoluwa Adelani et al.

With over 2,000 languages and potentially millions of speakers, Africa represents one of the richest linguistic regions in the world. Yet, this diversity is scarcely reflected in state-of-the-art natural language processing (NLP) systems and large language models (LLMs), which predominantly support a narrow set of high-resource languages. This exclusion not only limits the reach and utility of modern NLP technologies but also risks widening the digital divide across linguistic communities. Nevertheless, NLP research on African languages is active and growing. In recent years, there has been a surge of interest in this area, driven by several factors-including the creation of multilingual language resources, the rise of community-led initiatives, and increased support through funding programs. In this survey, we analyze 884 research papers on NLP for African languages published over the past five years, offering a comprehensive overview of recent progress across core tasks. We identify key trends shaping the field and conclude by outlining promising directions to foster more inclusive and sustainable NLP research for African languages.

Jesujoba O. Alabi, Israel Abebe Azime, Miaoran Zhang et al.

This paper introduces AFRIDOC-MT, a document-level multi-parallel translation dataset covering English and five African languages: Amharic, Hausa, Swahili, Yorùbá, and Zulu. The dataset comprises 334 health and 271 information technology news documents, all human-translated from English to these languages. We conduct document-level translation benchmark experiments by evaluating neural machine translation (NMT) models and large language models (LLMs) for translations between English and these languages, at both the sentence and pseudo-document levels. These outputs are realigned to form complete documents for evaluation. Our results indicate that NLLB-200 achieved the best average performance among the standard NMT models, while GPT-4o outperformed general-purpose LLMs. Fine-tuning selected models led to substantial performance gains, but models trained on sentences struggled to generalize effectively to longer documents. Furthermore, our analysis reveals that some LLMs exhibit issues such as under-generation, repetition of words or phrases, and off-target translations, especially for African languages.