Global MMLU: Understanding and Addressing Cultural and Linguistic Biases in Multilingual Evaluation

Cultural biases in multilingual datasets pose significant challenges for their effectiveness as global benchmarks. These biases stem not only from language but also from the cultural knowledge required to interpret questions, reducing the practical utility of translated datasets like MMLU. Furthermore, translation often introduces artifacts that can distort the meaning or clarity of questions in the target language. A common practice in multilingual evaluation is to rely on machine-translated evaluation sets, but simply translating a dataset is insufficient to address these challenges. In this work, we trace the impact of both of these issues on multilingual evaluations and ensuing model performances. Our large-scale evaluation of state-of-the-art open and proprietary models illustrates that progress on MMLU depends heavily on learning Western-centric concepts, with 28% of all questions requiring culturally sensitive knowledge. Moreover, for questions requiring geographic knowledge, an astounding 84.9% focus on either North American or European regions. Rankings of model evaluations change depending on whether they are evaluated on the full portion or the subset of questions annotated as culturally sensitive, showing the distortion to model rankings when blindly relying on translated MMLU. We release Global-MMLU, an improved MMLU with evaluation coverage across 42 languages -- with improved overall quality by engaging with compensated professional and community annotators to verify translation quality while also rigorously evaluating cultural biases present in the original dataset. This comprehensive Global-MMLU set also includes designated subsets labeled as culturally sensitive and culturally agnostic to allow for more holistic, complete evaluation.

How good are Large Language Models on African Languages?

Recent advancements in natural language processing have led to the proliferation of large language models (LLMs). These models have been shown to yield good performance, using in-context learning, even on unseen tasks and languages. Additionally, they have been widely adopted as language-model-as-a-service commercial APIs like GPT-4 API. However, their performance on African languages is largely unknown. We present an analysis of three popular large language models (mT0, LLaMa 2, and GPT-4) on five tasks (news topic classification, sentiment classification, machine translation, question answering, and named entity recognition) across 30 African languages, spanning different language families and geographical regions. Our results suggest that all LLMs produce below-par performance on African languages, and there is a large gap in performance compared to high-resource languages like English most tasks. We find that GPT-4 has an average or impressive performance on classification tasks but very poor results on generative tasks like machine translation. Surprisingly, we find that mT0 had the best overall on cross-lingual QA, better than the state-of-the-art supervised model (i.e. fine-tuned mT5) and GPT-4 on African languages. Overall, LLaMa 2 records the worst performance due to its limited multilingual capabilities and English-centric pre-training corpus. In general, our findings present a call-to-action to ensure African languages are well represented in large language models, given their growing popularity.

XTREME-UP: A User-Centric Scarce-Data Benchmark for Under-Represented Languages

Data scarcity is a crucial issue for the development of highly multilingual NLP systems. Yet for many under-represented languages (ULs) -- languages for which NLP re-search is particularly far behind in meeting user needs -- it is feasible to annotate small amounts of data. Motivated by this, we propose XTREME-UP, a benchmark defined by: its focus on the scarce-data scenario rather than zero-shot; its focus on user-centric tasks -- tasks with broad adoption by speakers of high-resource languages; and its focus on under-represented languages where this scarce-data scenario tends to be most realistic. XTREME-UP evaluates the capabilities of language models across 88 under-represented languages over 9 key user-centric technologies including ASR, OCR, MT, and information access tasks that are of general utility. We create new datasets for OCR, autocomplete, semantic parsing, and transliteration, and build on and refine existing datasets for other tasks. XTREME-UP provides methodology for evaluating many modeling scenarios including text-only, multi-modal (vision, audio, and text),supervised parameter tuning, and in-context learning. We evaluate commonly used models on the benchmark. We release all code and scripts to train and evaluate models

MphayaNER: Named Entity Recognition for Tshivenda

Named Entity Recognition (NER) plays a vital role in various Natural Language Processing tasks such as information retrieval, text classification, and question answering. However, NER can be challenging, especially in low-resource languages with limited annotated datasets and tools. This paper adds to the effort of addressing these challenges by introducing MphayaNER, the first Tshivenda NER corpus in the news domain. We establish NER baselines by \textit{fine-tuning} state-of-the-art models on MphayaNER. The study also explores zero-shot transfer between Tshivenda and other related Bantu languages, with chiShona and Kiswahili showing the best results. Augmenting MphayaNER with chiShona data was also found to improve model performance significantly. Both MphayaNER and the baseline models are made publicly available.

AI4D -- African Language Program

Advances in speech and language technologies enable tools such as voice-search, text-to-speech, speech recognition and machine translation. These are however only available for high resource languages like English, French or Chinese. Without foundational digital resources for African languages, which are considered low-resource in the digital context, these advanced tools remain out of reach. This work details the AI4D - African Language Program, a 3-part project that 1) incentivised the crowd-sourcing, collection and curation of language datasets through an online quantitative and qualitative challenge, 2) supported research fellows for a period of 3-4 months to create datasets annotated for NLP tasks, and 3) hosted competitive Machine Learning challenges on the basis of these datasets. Key outcomes of the work so far include 1) the creation of 9+ open source, African language datasets annotated for a variety of ML tasks, and 2) the creation of baseline models for these datasets through hosting of competitive ML challenges.

MENYO-20k: A Multi-domain English-Yorùbá Corpus for Machine Translation and Domain Adaptation

Massively multilingual machine translation (MT) has shown impressive capabilities, including zero and few-shot translation between low-resource language pairs. However, these models are often evaluated on high-resource languages with the assumption that they generalize to low-resource ones. The difficulty of evaluating MT models on low-resource pairs is often due the lack of standardized evaluation datasets. In this paper, we present MENYO-20k, the first multi-domain parallel corpus for the low-resource Yor\`ub\'a--English (yo--en) language pair with standardized train-test splits for benchmarking. We provide several neural MT (NMT) benchmarks on this dataset and compare to the performance of popular pre-trained (massively multilingual) MT models, showing that, in almost all cases, our simple benchmarks outperform the pre-trained MT models. A major gain of BLEU $+9.9$ and $+8.6$ (en2yo) is achieved in comparison to Facebook's M2M-100 and Google multilingual NMT respectively when we use MENYO-20k to fine-tune generic models.

Improving Yorùbá Diacritic Restoration

Yor\`ub\'a is a widely spoken West African language with a writing system rich in orthographic and tonal diacritics. They provide morphological information, are crucial for lexical disambiguation, pronunciation and are vital for any computational Speech or Natural Language Processing tasks. However diacritic marks are commonly excluded from electronic texts due to limited device and application support as well as general education on proper usage. We report on recent efforts at dataset cultivation. By aggregating and improving disparate texts from the web and various personal libraries, we were able to significantly grow our clean Yor\`ub\'a dataset from a majority Bibilical text corpora with three sources to millions of tokens from over a dozen sources. We evaluate updated diacritic restoration models on a new, general purpose, public-domain Yor\`ub\'a evaluation dataset of modern journalistic news text, selected to be multi-purpose and reflecting contemporary usage. All pre-trained models, datasets and source-code have been released as an open-source project to advance efforts on Yor\`ub\'a language technology.

Massive vs. Curated Word Embeddings for Low-Resourced Languages. The Case of Yorùbá and Twi

The success of several architectures to learn semantic representations from unannotated text and the availability of these kind of texts in online multilingual resources such as Wikipedia has facilitated the massive and automatic creation of resources for multiple languages. The evaluation of such resources is usually done for the high-resourced languages, where one has a smorgasbord of tasks and test sets to evaluate on. For low-resourced languages, the evaluation is more difficult and normally ignored, with the hope that the impressive capability of deep learning architectures to learn (multilingual) representations in the high-resourced setting holds in the low-resourced setting too. In this paper we focus on two African languages, Yor\`ub\'a and Twi, and compare the word embeddings obtained in this way, with word embeddings obtained from curated corpora and a language-dependent processing. We analyse the noise in the publicly available corpora, collect high quality and noisy data for the two languages and quantify the improvements that depend not only on the amount of data but on the quality too. We also use different architectures that learn word representations both from surface forms and characters to further exploit all the available information which showed to be important for these languages. For the evaluation, we manually translate the wordsim-353 word pairs dataset from English into Yor\`ub\'a and Twi. As output of the work, we provide corpora, embeddings and the test suits for both languages.