Reliable multilingual evaluation is difficult, and culturally appropriate evaluation is even harder to achieve.A common practice to fill this gap is to machine-translate English evaluation sets. However, translation introduces language bias and carries over cultural and regional assumptions from the original questions – often testing knowledge irrelevant to the target audience. In this work, we highlight the extent and impact of these biases and present a multilingual evaluation framework that aims to mitigate them through improved translations and annotation practices.Through a large-scale study involving professional and community translators and annotators, we show that state-of-the-art models excel primarily by learning Western-centric concepts. Notably, we find that model rankings on the full MMLU change when evaluated on a subset of questions explicitly marked as culturally sensitive.We release Global MMLU, a multilingual extension of MMLU across 42 languages, featuring improved translation quality, expanded language coverage, and designated subsets labeled as culturally sensitive and culturally agnostic to enable a more comprehensive and equitable benchmark for evaluating language models across diverse linguistic and cultural contexts.

Dialectal Arabic (DA) varieties are under-served by language technologies, particularly large language models (LLMs). This trend threatens to exacerbate existing social inequalities and limits LLM applications, yet the research community lacks operationalized performance measurements in DA. We present a framework that comprehensively assesses LLMs’ DA modeling capabilities across four dimensions: fidelity, understanding, quality, and diglossia. We evaluate nine LLMs in eight DA varieties and provide practical recommendations. Our evaluation suggests that LLMs do not produce DA as well as they understand it, not because their DA fluency is poor, but because they are reluctant to generate DA. Further analysis suggests that current post-training can contribute to bias against DA, that few-shot examples can overcome this deficiency, and that otherwise no measurable features of input text correlate well with LLM DA performance.

We investigate a surprising limitation of LLMs: their inability to consistently generate text in a user’s desired language. We create the Language Confusion Benchmark (LCB) to evaluate such failures, covering 15 typologically diverse languages with existing and newly-created English and multilingual prompts. We evaluate a range of LLMs on monolingual and cross-lingual generation reflecting practical use cases, finding that Llama Instruct and Mistral models exhibit high degrees of language confusion and even the strongest models fail to consistently respond in the correct language. We observe that base and English-centric instruct models are more prone to language confusion, which is aggravated by complex prompts and high sampling temperatures. We find that language confusion can be partially mitigated via few-shot prompting, multilingual SFT and preference tuning. We release our language confusion benchmark, which serves as a first layer of efficient, scalable multilingual evaluation.

Preference optimization techniques have become a standard final stage for training state-of-art large language models (LLMs). However, despite widespread adoption, the vast majority of work to-date has focused on a small set of high-resource languages like English and Chinese. This captures a small fraction of the languages in the world, but also makes it unclear which aspects of current state-of-the-art research transfer to a multilingual setting. In this work, we perform an exhaustive study to achieve a new state of the art in aligning multilingual LLMs. We introduce a novel, scalable method for generating high-quality multilingual feedback data to balance data coverage. We establish the benefits of cross-lingual transfer and increased dataset size in preference training. Our preference-trained model achieves a 54.4% win-rate against Aya 23 8B, the current state-of-the-art multilingual LLM in its parameter class, and a 69.5% win-rate or higher against widely used models like Gemma, Mistral and Llama 3. As a result of our efforts, we expand the frontier of alignment techniques to 23 languages, covering approximately half of the world’s population.

Quantization techniques are widely used to improve inference speed and deployment of large language models. While a wide body of work examines the impact of quantization on LLMs in English, none have evaluated across languages. We conduct a thorough analysis of quantized multilingual LLMs, focusing on performance across languages and at varying scales. We use automatic benchmarks, LLM-as-a-Judge, and human evaluation, finding that (1) harmful effects of quantization are apparent in human evaluation, which automatic metrics severely underestimate: a 1.7% average drop in Japanese across automatic tasks corresponds to a 16.0% drop reported by human evaluators on realistic prompts; (2) languages are disparately affected by quantization, with non-Latin script languages impacted worst; and (3) challenging tasks like mathematical reasoning degrade fastest. As the ability to serve low-compute models is critical for wide global adoption of NLP technologies, our results urge consideration of multilingual performance as a key evaluation criterion for efficient models.

Prior work shows that it is possible to expand pretrained Masked Language Models (MLMs) to new languages by learning a new set of embeddings, while keeping the transformer body frozen. Despite learning a small subset of parameters, this approach is not compute-efficient, as training the new embeddings requires a full forward and backward pass over the entire model. We propose mini-model adaptation, a compute-efficient alternative that builds a shallow mini-model from a fraction of a large model’s parameters. New language-specific embeddings can then be efficiently trained over the mini-model and plugged into the aligned large model for rapid cross-lingual transfer. We explore two approaches to learn mini-models: MINIJOINT, which jointly pretrains the primary model and the mini-model using a single transformer with a secondary MLM head at a middle layer; and MINIPOST, where we start from a regular pretrained model, build a mini-model by extracting and freezing a few layers, and learn a small number of parameters on top. Experiments on XNLI, MLQA and PAWS-X show that mini-model adaptation matches the performance of the standard approach using up to 2.3x less compute on average.

A popular natural language processing task decades ago, word alignment has been dominated until recently by GIZA++, a statistical method based on the 30-year-old IBM models. New methods that outperform GIZA++ primarily rely on large machine translation models, massively multilingual language models, or supervision from GIZA++ alignments itself. We introduce Embedding-Enhanced GIZA++, and outperform GIZA++ without any of the aforementioned factors. Taking advantage of monolingual embedding spaces of source and target language only, we exceed GIZA++’s performance in every tested scenario for three languages pairs. In the lowest-resource setting, we outperform GIZA++ by 8.5, 10.9, and 12 AER for RoEn, De-En, and En-Fr, respectively. We release our code at www.blind-review.code.

Bilingual lexicons form a critical component of various natural language processing applications, including unsupervised and semisupervised machine translation and crosslingual information retrieval. In this work, we improve bilingual lexicon induction performance across 40 language pairs with a graph-matching method based on optimal transport. The method is especially strong with low amounts of supervision.

The ability to extract high-quality translation dictionaries from monolingual word embedding spaces depends critically on the geometric similarity of the spaces—their degree of “isomorphism.” We address the root-cause of faulty cross-lingual mapping: that word embedding training resulted in the underlying spaces being non-isomorphic. We incorporate global measures of isomorphism directly into the skipgram loss function, successfully increasing the relative isomorphism of trained word embedding spaces and improving their ability to be mapped to a shared cross-lingual space. The result is improved bilingual lexicon induction in general data conditions, under domain mismatch, and with training algorithm dissimilarities. We release IsoVec at https://github.com/kellymarchisio/isovec.

Modern unsupervised machine translation (MT) systems reach reasonable translation quality under clean and controlled data conditions. As the performance gap between supervised and unsupervised MT narrows, it is interesting to ask whether the different training methods result in systematically different output beyond what is visible via quality metrics like adequacy or BLEU. We compare translations from supervised and unsupervised MT systems of similar quality, finding that unsupervised output is more fluent and more structurally different in comparison to human translation than is supervised MT. We then demonstrate a way to combine the benefits of both methods into a single system which results in improved adequacy and fluency as rated by human evaluators. Our results open the door to interesting discussions about how supervised and unsupervised MT might be different yet mutually-beneficial.

Much recent work in bilingual lexicon induction (BLI) views word embeddings as vectors in Euclidean space. As such, BLI is typically solved by finding a linear transformation that maps embeddings to a common space. Alternatively, word embeddings may be understood as nodes in a weighted graph. This framing allows us to examine a node’s graph neighborhood without assuming a linear transform, and exploits new techniques from the graph matching optimization literature. These contrasting approaches have not been compared in BLI so far. In this work, we study the behavior of Euclidean versus graph-based approaches to BLI under differing data conditions and show that they complement each other when combined. We release our code at https://github.com/kellymarchisio/euc-v-graph-bli.

Aimed at generating a seed lexicon for use in downstream natural language tasks and unsupervised methods for bilingual lexicon induction have received much attention in the academic literature recently. While interesting and fully unsupervised settings are unrealistic; small amounts of bilingual data are usually available due to the existence of massively multilingual parallel corpora and or linguists can create small amounts of parallel data. In this work and we demonstrate an effective bootstrapping approach for semi-supervised bilingual lexicon induction that capitalizes upon the complementary strengths of two disparate methods for inducing bilingual lexicons. Whereas statistical methods are highly effective at inducing correct translation pairs for words frequently occurring in a parallel corpus and monolingual embedding spaces have the advantage of having been trained on large amounts of data and and therefore may induce accurate translations for words absent from the small corpus. By combining these relative strengths and our method achieves state-of-the-art results on 3 of 4 language pairs in the challenging VecMap test set using minimal amounts of parallel data and without the need for a translation dictionary. We release our implementation at www.blind-review.code.

Despite the reported success of unsupervised machine translation (MT), the field has yet to examine the conditions under which the methods succeed and fail. We conduct an extensive empirical evaluation using dissimilar language pairs, dissimilar domains, and diverse datasets. We find that performance rapidly deteriorates when source and target corpora are from different domains, and that stochasticity during embedding training can dramatically affect downstream results. We additionally find that unsupervised MT performance declines when source and target languages use different scripts, and observe very poor performance on authentic low-resource language pairs. We advocate for extensive empirical evaluation of unsupervised MT systems to highlight failure points and encourage continued research on the most promising paradigms. We release our preprocessed dataset to encourage evaluations that stress-test systems under multiple data conditions.

We describe the work of Johns Hopkins University for the shared task of news translation organized by the Fourth Conference on Machine Translation (2019). We submitted systems for both directions of the English-German language pair. The systems combine multiple techniques – sampling, filtering, iterative backtranslation, and continued training – previously used to improve performance of neural machine translation models. At submission time, we achieve a BLEU score of 38.1 for De-En and 42.5 for En-De translation directions on newstest2019. Post-submission, the score is 38.4 for De-En and 42.8 for En-De. Various experiments conducted in the process are also described.