Simran Khanuja

Alexis Conneau, Min Ma, Simran Khanuja et al. · cmu

We introduce FLEURS, the Few-shot Learning Evaluation of Universal Representations of Speech benchmark. FLEURS is an n-way parallel speech dataset in 102 languages built on top of the machine translation FLoRes-101 benchmark, with approximately 12 hours of speech supervision per language. FLEURS can be used for a variety of speech tasks, including Automatic Speech Recognition (ASR), Speech Language Identification (Speech LangID), Translation and Retrieval. In this paper, we provide baselines for the tasks based on multilingual pre-trained models like mSLAM. The goal of FLEURS is to enable speech technology in more languages and catalyze research in low-resource speech understanding.

Alexis Conneau, Ankur Bapna, Yu Zhang et al. · stanford

We introduce XTREME-S, a new benchmark to evaluate universal cross-lingual speech representations in many languages. XTREME-S covers four task families: speech recognition, classification, speech-to-text translation and retrieval. Covering 102 languages from 10+ language families, 3 different domains and 4 task families, XTREME-S aims to simplify multilingual speech representation evaluation, as well as catalyze research in "universal" speech representation learning. This paper describes the new benchmark and establishes the first speech-only and speech-text baselines using XLS-R and mSLAM on all downstream tasks. We motivate the design choices and detail how to use the benchmark. Datasets and fine-tuning scripts are made easily accessible at https://hf.co/datasets/google/xtreme_s.

Simran Khanuja, Srinivas Gowriraj, Lucio Dery et al.

We consider the task of optimally fine-tuning pre-trained multilingual models, given small amounts of unlabelled target data and an annotation budget. In this paper, we introduce DEMUX, a framework that prescribes the exact data-points to label from vast amounts of unlabelled multilingual data, having unknown degrees of overlap with the target set. Unlike most prior works, our end-to-end framework is language-agnostic, accounts for model representations, and supports multilingual target configurations. Our active learning strategies rely upon distance and uncertainty measures to select task-specific neighbors that are most informative to label, given a model. DeMuX outperforms strong baselines in 84% of the test cases, in the zero-shot setting of disjoint source and target language sets (including multilingual target pools), across three models and four tasks. Notably, in low-budget settings (5-100 examples), we observe gains of up to 8-11 F1 points for token-level tasks, and 2-5 F1 for complex tasks. Our code is released here: https://github.com/simran-khanuja/demux.

Simran Khanuja, Sebastian Ruder, Partha Talukdar

In order for NLP technology to be widely applicable, fair, and useful, it needs to serve a diverse set of speakers across the world's languages, be equitable, i.e., not unduly biased towards any particular language, and be inclusive of all users, particularly in low-resource settings where compute constraints are common. In this paper, we propose an evaluation paradigm that assesses NLP technologies across all three dimensions. While diversity and inclusion have received attention in recent literature, equity is currently unexplored. We propose to address this gap using the Gini coefficient, a well-established metric used for estimating societal wealth inequality. Using our paradigm, we highlight the distressed state of current technologies for Indian (IN) languages (a linguistically large and diverse set, with a varied speaker population), across all three dimensions. To improve upon these metrics, we demonstrate the importance of region-specific choices in model building and dataset creation, and more importantly, propose a novel, generalisable approach to optimal resource allocation during fine-tuning. Finally, we discuss steps to mitigate these biases and encourage the community to employ multi-faceted evaluation when building linguistically diverse and equitable technologies.

Simran Khanuja, Hongbin Liu, Shujian Zhang et al.

LLMs are deployed globally, yet produce responses biased towards cultures with abundant training data. Existing cultural localization approaches such as prompting or post-training alignment are black-box, hard to control, and do not reveal whether failures reflect missing knowledge or poor elicitation. In this paper, we address these gaps using mechanistic interpretability to uncover and manipulate cultural representations in LLMs. Leveraging sparse autoencoders, we identify interpretable features that encode culturally salient information and aggregate them into Cultural Embeddings (CuE). We use CuE both to analyze implicit cultural biases under underspecified prompts and to construct white-box steering interventions. Across multiple models, we show that CuE-based steering increases cultural faithfulness and elicits significantly rarer, long-tail cultural concepts than prompting alone. Notably, CuE-based steering is complementary to black-box localization methods, offering gains when applied on top of prompt-augmented inputs. This also suggests that models do benefit from better elicitation strategies, and don't necessarily lack long-tail knowledge representation, though this varies across cultures. Our results provide both diagnostic insight into cultural representations in LLMs and a controllable method to steer towards desired cultures.

Xiang Yue, Yueqi Song, Akari Asai et al. · cmu

Despite recent advances in multimodal large language models (MLLMs), their development has predominantly focused on English- and western-centric datasets and tasks, leaving most of the world's languages and diverse cultural contexts underrepresented. This paper introduces Pangea, a multilingual multimodal LLM trained on PangeaIns, a diverse 6M instruction dataset spanning 39 languages. PangeaIns features: 1) high-quality English instructions, 2) carefully machine-translated instructions, and 3) culturally relevant multimodal tasks to ensure cross-cultural coverage. To rigorously assess models' capabilities, we introduce PangeaBench, a holistic evaluation suite encompassing 14 datasets covering 47 languages. Results show that Pangea significantly outperforms existing open-source models in multilingual settings and diverse cultural contexts. Ablation studies further reveal the importance of English data proportions, language popularity, and the number of multimodal training samples on overall performance. We fully open-source our data, code, and trained checkpoints, to facilitate the development of inclusive and robust multilingual MLLMs, promoting equity and accessibility across a broader linguistic and cultural spectrum.

Simran Khanuja, Sathyanarayanan Ramamoorthy, Yueqi Song et al.

Given the rise of multimedia content, human translators increasingly focus on culturally adapting not only words but also other modalities such as images to convey the same meaning. While several applications stand to benefit from this, machine translation systems remain confined to dealing with language in speech and text. In this work, we take a first step towards translating images to make them culturally relevant. First, we build three pipelines comprising state-of-the-art generative models to do the task. Next, we build a two-part evaluation dataset: i) concept: comprising 600 images that are cross-culturally coherent, focusing on a single concept per image, and ii) application: comprising 100 images curated from real-world applications. We conduct a multi-faceted human evaluation of translated images to assess for cultural relevance and meaning preservation. We find that as of today, image-editing models fail at this task, but can be improved by leveraging LLMs and retrievers in the loop. Best pipelines can only translate 5% of images for some countries in the easier concept dataset and no translation is successful for some countries in the application dataset, highlighting the challenging nature of the task. Our code and data is released here: https://github.com/simran-khanuja/image-transcreation.

Jean de Dieu Nyandwi, Yueqi Song, Simran Khanuja et al.

Multimodal Large Language Models excel in high-resource settings, but often misinterpret long-tail cultural entities and underperform in low-resource languages. To address this gap, we propose a data-centric approach that directly grounds MLLMs in cultural knowledge. Leveraging a large scale knowledge graph from Wikidata, we collect images that represent culturally significant entities, and generate synthetic multilingual visual question answering data. The resulting dataset, CulturalGround, comprises 22 million high-quality, culturally-rich VQA pairs spanning 42 countries and 39 languages. We train an open-source MLLM CulturalPangea on CulturalGround, interleaving standard multilingual instruction-tuning data to preserve general abilities. CulturalPangea achieves state-of-the-art performance among open models on various culture-focused multilingual multimodal benchmarks, outperforming prior models by an average of 5.0 without degrading results on mainstream vision-language tasks. Our findings show that our targeted, culturally grounded approach could substantially narrow the cultural gap in MLLMs and offer a practical path towards globally inclusive multimodal systems.

Simran Khanuja, Vivek Iyer, Claire He et al.

Beyond conventional paradigms of translating speech and text, recently, there has been interest in automated transcreation of images to facilitate localization of visual content across different cultures. Attempts to define this as a formal Machine Learning (ML) problem have been impeded by the lack of automatic evaluation mechanisms, with previous work relying solely on human evaluation. In this paper, we seek to close this gap by proposing a suite of automatic evaluation metrics inspired by machine translation (MT) metrics, categorized into: a) Object-based, b) Embedding-based, and c) VLM-based. Drawing on theories from translation studies and real-world transcreation practices, we identify three critical dimensions of image transcreation: cultural relevance, semantic equivalence and visual similarity, and design our metrics to evaluate systems along these axes. Our results show that proprietary VLMs best identify cultural relevance and semantic equivalence, while vision-encoder representations are adept at measuring visual similarity. Meta-evaluation across 7 countries shows our metrics agree strongly with human ratings, with average segment-level correlations ranging from 0.55-0.87. Finally, through a discussion of the merits and demerits of each metric, we offer a robust framework for automated image transcreation evaluation, grounded in both theoretical foundations and practical application. Our code can be found here: https://github.com/simran-khanuja/automatic-eval-img-transcreation.

Sathyanarayanan Ramamoorthy, Vishwa Shah, Simran Khanuja et al.

This paper introduces MERLIN, a novel testbed system for the task of Multilingual Multimodal Entity Linking. The created dataset includes BBC news article titles, paired with corresponding images, in five languages: Hindi, Japanese, Indonesian, Vietnamese, and Tamil, featuring over 7,000 named entity mentions linked to 2,500 unique Wikidata entities. We also include several benchmarks using multilingual and multimodal entity linking methods exploring different language models like LLaMa-2 and Aya-23. Our findings indicate that incorporating visual data improves the accuracy of entity linking, especially for entities where the textual context is ambiguous or insufficient, and particularly for models that do not have strong multilingual abilities. For the work, the dataset, methods are available here at https://github.com/rsathya4802/merlin

Baiqi Li, Zhiqiu Lin, Wenxuan Peng et al.

Vision-language models (VLMs) have made significant progress in recent visual-question-answering (VQA) benchmarks that evaluate complex visio-linguistic reasoning. However, are these models truly effective? In this work, we show that VLMs still struggle with natural images and questions that humans can easily answer, which we term natural adversarial samples. We also find it surprisingly easy to generate these VQA samples from natural image-text corpora using off-the-shelf models like CLIP and ChatGPT. We propose a semi-automated approach to collect a new benchmark, NaturalBench, for reliably evaluating VLMs with 10,000 human-verified VQA samples. Crucially, we adopt a $\textbf{vision-centric}$ design by pairing each question with two images that yield different answers, preventing blind solutions from answering without using the images. This makes NaturalBench more challenging than previous benchmarks that can be solved with commonsense priors. We evaluate 53 state-of-the-art VLMs on NaturalBench, showing that models like LLaVA-OneVision, Cambrian-1, Llama3.2-Vision, Molmo, Qwen2-VL, and even GPT-4o lag 50%-70% behind human performance (over 90%). We analyze why NaturalBench is hard from two angles: (1) Compositionality: Solving NaturalBench requires diverse visio-linguistic skills, including understanding attribute bindings, object relationships, and advanced reasoning like logic and counting. To this end, unlike prior work that uses a single tag per sample, we tag each NaturalBench sample with 1 to 8 skill tags for fine-grained evaluation. (2) Biases: NaturalBench exposes severe biases in VLMs, as models often choose the same answer regardless of the image. Lastly, we apply our benchmark curation method to diverse data sources, including long captions (over 100 words) and non-English languages like Chinese and Hindi, highlighting its potential for dynamic evaluations of VLMs.

Yueqi Song, Simran Khanuja, Graham Neubig

NLP models today strive for supporting multiple languages and modalities, improving accessibility for diverse users. In this paper, we evaluate their multilingual, multimodal capabilities by testing on a visual reasoning task. We observe that proprietary systems like GPT-4V obtain the best performance on this task now, but open models lag in comparison. Surprisingly, GPT-4V exhibits similar performance between English and other languages, indicating the potential for equitable system development across languages. Our analysis on model failures reveals three key aspects that make this task challenging: multilinguality, complex reasoning, and multimodality. To address these challenges, we propose three targeted interventions including a translate-test approach to tackle multilinguality, a visual programming approach to break down complex reasoning, and a method that leverages image captioning to address multimodality. Our interventions achieve the best open performance on this task in a zero-shot setting, boosting open models LLaVA-v1.5-13B by 13.4%, LLaVA-v1.6-34B by 20.3%, and Qwen-VL by 16.7%, while also minorly improving GPT-4V's performance.

Arnav Yayavaram, Siddharth Yayavaram, Simran Khanuja et al.

As text-to-image models become increasingly prevalent, ensuring their equitable performance across diverse cultural contexts is critical. Efforts to mitigate cross-cultural biases have been hampered by trade-offs, including a loss in performance, factual inaccuracies, or offensive outputs. Despite widespread recognition of these challenges, an inability to reliably measure these biases has stalled progress. To address this gap, we introduce CAIRe, an evaluation metric that assesses the degree of cultural relevance of an image, given a user-defined set of labels. Our framework grounds entities and concepts in the image to a knowledge base and uses factual information to give independent graded judgments for each culture label. On a manually curated dataset of culturally salient but rare items built using language models, CAIRe surpasses all baselines by 22% F1 points. Additionally, we construct two datasets for culturally universal concepts, one comprising T2I-generated outputs and another retrieved from naturally occurring data. CAIRe achieves Pearson's correlations of 0.56 and 0.66 with human ratings on these sets, based on a 5-point Likert scale of cultural relevance. This demonstrates its strong alignment with human judgment across diverse image sources.

Anubha Kabra, Emmy Liu, Simran Khanuja et al.

Figurative language permeates human communication, but at the same time is relatively understudied in NLP. Datasets have been created in English to accelerate progress towards measuring and improving figurative language processing in language models (LMs). However, the use of figurative language is an expression of our cultural and societal experiences, making it difficult for these phrases to be universally applicable. In this work, we create a figurative language inference dataset, \datasetname, for seven diverse languages associated with a variety of cultures: Hindi, Indonesian, Javanese, Kannada, Sundanese, Swahili and Yoruba. Our dataset reveals that each language relies on cultural and regional concepts for figurative expressions, with the highest overlap between languages originating from the same region. We assess multilingual LMs' abilities to interpret figurative language in zero-shot and few-shot settings. All languages exhibit a significant deficiency compared to English, with variations in performance reflecting the availability of pre-training and fine-tuning data, emphasizing the need for LMs to be exposed to a broader range of linguistic and cultural variation during training.

Yueqi Song, Catherine Cui, Simran Khanuja et al.

Despite the major advances in NLP, significant disparities in NLP system performance across languages still exist. Arguably, these are due to uneven resource allocation and sub-optimal incentives to work on less resourced languages. To track and further incentivize the global development of equitable language technology, we introduce GlobalBench. Prior multilingual benchmarks are static and have focused on a limited number of tasks and languages. In contrast, GlobalBench is an ever-expanding collection that aims to dynamically track progress on all NLP datasets in all languages. Rather than solely measuring accuracy, GlobalBench also tracks the estimated per-speaker utility and equity of technology across all languages, providing a multi-faceted view of how language technology is serving people of the world. Furthermore, GlobalBench is designed to identify the most under-served languages, and rewards research efforts directed towards those languages. At present, the most under-served languages are the ones with a relatively high population, but nonetheless overlooked by composite multilingual benchmarks (like Punjabi, Portuguese, and Wu Chinese). Currently, GlobalBench covers 966 datasets in 190 languages, and has 1,128 system submissions spanning 62 languages.

Ankur Bapna, Colin Cherry, Yu Zhang et al.

We present mSLAM, a multilingual Speech and LAnguage Model that learns cross-lingual cross-modal representations of speech and text by pre-training jointly on large amounts of unlabeled speech and text in multiple languages. mSLAM combines w2v-BERT pre-training on speech with SpanBERT pre-training on character-level text, along with Connectionist Temporal Classification (CTC) losses on paired speech and transcript data, to learn a single model capable of learning from and representing both speech and text signals in a shared representation space. We evaluate mSLAM on several downstream speech understanding tasks and find that joint pre-training with text improves quality on speech translation, speech intent classification and speech language-ID while being competitive on multilingual ASR, when compared against speech-only pre-training. Our speech translation model demonstrates zero-shot text translation without seeing any text translation data, providing evidence for cross-modal alignment of representations. mSLAM also benefits from multi-modal fine-tuning, further improving the quality of speech translation by directly leveraging text translation data during the fine-tuning process. Our empirical analysis highlights several opportunities and challenges arising from large-scale multimodal pre-training, suggesting directions for future research.

Simran Khanuja, Melvin Johnson, Partha Talukdar

Pre-trained multilingual language models (LMs) have achieved state-of-the-art results in cross-lingual transfer, but they often lead to an inequitable representation of languages due to limited capacity, skewed pre-training data, and sub-optimal vocabularies. This has prompted the creation of an ever-growing pre-trained model universe, where each model is trained on large amounts of language or domain specific data with a carefully curated, linguistically informed vocabulary. However, doing so brings us back full circle and prevents one from leveraging the benefits of multilinguality. To address the gaps at both ends of the spectrum, we propose MergeDistill, a framework to merge pre-trained LMs in a way that can best leverage their assets with minimal dependencies, using task-agnostic knowledge distillation. We demonstrate the applicability of our framework in a practical setting by leveraging pre-existing teacher LMs and training student LMs that perform competitively with or even outperform teacher LMs trained on several orders of magnitude more data and with a fixed model capacity. We also highlight the importance of teacher selection and its impact on student model performance.

Simran Khanuja, Diksha Bansal, Sarvesh Mehtani et al.

India is a multilingual society with 1369 rationalized languages and dialects being spoken across the country (INDIA, 2011). Of these, the 22 scheduled languages have a staggering total of 1.17 billion speakers and 121 languages have more than 10,000 speakers (INDIA, 2011). India also has the second largest (and an ever growing) digital footprint (Statista, 2020). Despite this, today's state-of-the-art multilingual systems perform suboptimally on Indian (IN) languages. This can be explained by the fact that multilingual language models (LMs) are often trained on 100+ languages together, leading to a small representation of IN languages in their vocabulary and training data. Multilingual LMs are substantially less effective in resource-lean scenarios (Wu and Dredze, 2020; Lauscher et al., 2020), as limited data doesn't help capture the various nuances of a language. One also commonly observes IN language text transliterated to Latin or code-mixed with English, especially in informal settings (for example, on social media platforms) (Rijhwani et al., 2017). This phenomenon is not adequately handled by current state-of-the-art multilingual LMs. To address the aforementioned gaps, we propose MuRIL, a multilingual LM specifically built for IN languages. MuRIL is trained on significantly large amounts of IN text corpora only. We explicitly augment monolingual text corpora with both translated and transliterated document pairs, that serve as supervised cross-lingual signals in training. MuRIL significantly outperforms multilingual BERT (mBERT) on all tasks in the challenging cross-lingual XTREME benchmark (Hu et al., 2020). We also present results on transliterated (native to Latin script) test sets of the chosen datasets and demonstrate the efficacy of MuRIL in handling transliterated data.

Sanket Shah, Satarupa Guha, Simran Khanuja et al.

Spoken Term Detection (STD) is the task of searching for words or phrases within audio, given either text or spoken input as a query. In this work, we use state-of-the-art Hindi, Tamil and Telugu ASR systems cross-lingually for lexical Spoken Term Detection in ten low-resource Indian languages. Since no publicly available dataset exists for Spoken Term Detection in these languages, we create a new dataset using a publicly available TTS dataset. We report a standard metric for STD, Mean Term Weighted Value (MTWV) and show that ASR systems built in languages that are phonetically similar to the target languages have higher accuracy, however, it is also possible to get high MTWV scores for dissimilar languages by using a relaxed phone matching algorithm. We propose a technique to bootstrap the Grapheme-to-Phoneme (g2p) mapping between all the languages under consideration using publicly available resources. Gains are obtained when we combine the output of multiple ASR systems and when we use language-specific Language Models. We show that it is possible to perform STD cross-lingually in a zero-shot manner without the need for any language-specific speech data. We plan to make the STD dataset available for other researchers interested in cross-lingual STD.

Simran Khanuja, Sandipan Dandapat, Anirudh Srinivasan et al.

Code-switching is the use of more than one language in the same conversation or utterance. Recently, multilingual contextual embedding models, trained on multiple monolingual corpora, have shown promising results on cross-lingual and multilingual tasks. We present an evaluation benchmark, GLUECoS, for code-switched languages, that spans several NLP tasks in English-Hindi and English-Spanish. Specifically, our evaluation benchmark includes Language Identification from text, POS tagging, Named Entity Recognition, Sentiment Analysis, Question Answering and a new task for code-switching, Natural Language Inference. We present results on all these tasks using cross-lingual word embedding models and multilingual models. In addition, we fine-tune multilingual models on artificially generated code-switched data. Although multilingual models perform significantly better than cross-lingual models, our results show that in most tasks, across both language pairs, multilingual models fine-tuned on code-switched data perform best, showing that multilingual models can be further optimized for code-switching tasks.

Simran Khanuja, Sandipan Dandapat, Sunayana Sitaram et al.

Natural Language Inference (NLI) is the task of inferring the logical relationship, typically entailment or contradiction, between a premise and hypothesis. Code-mixing is the use of more than one language in the same conversation or utterance, and is prevalent in multilingual communities all over the world. In this paper, we present the first dataset for code-mixed NLI, in which both the premises and hypotheses are in code-mixed Hindi-English. We use data from Hindi movies (Bollywood) as premises, and crowd-source hypotheses from Hindi-English bilinguals. We conduct a pilot annotation study and describe the final annotation protocol based on observations from the pilot. Currently, the data collected consists of 400 premises in the form of code-mixed conversation snippets and 2240 code-mixed hypotheses. We conduct an extensive analysis to infer the linguistic phenomena commonly observed in the dataset obtained. We evaluate the dataset using a standard mBERT-based pipeline for NLI and report results.

Pratik Joshi, Christain Barnes, Sebastin Santy et al.

In this paper, we examine and analyze the challenges associated with developing and introducing language technologies to low-resource language communities. While doing so, we bring to light the successes and failures of past work in this area, challenges being faced in doing so, and what they have achieved. Throughout this paper, we take a problem-facing approach and describe essential factors which the success of such technologies hinges upon. We present the various aspects in a manner which clarify and lay out the different tasks involved, which can aid organizations looking to make an impact in this area. We take the example of Gondi, an extremely-low resource Indian language, to reinforce and complement our discussion.