Vision Language Models (VLMs) often struggle with culture-specific knowledge, particularly in languages other than English and in underrepresented cultural contexts. To evaluate their understanding of such knowledge, we introduce WorldCuisines, a massive-scale benchmark for multilingual and multicultural, visually grounded language understanding. This benchmark includes a visual question answering (VQA) dataset with text-image pairs across 30 languages and dialects, spanning 9 language families and featuring over 1 million data points, making it the largest multicultural VQA benchmark to date. It includes tasks for identifying dish names and their origins. We provide evaluation datasets in two sizes (12k and 60k instances) alongside a training dataset (1 million instances). Our findings show that while VLMs perform better with correct location context, they struggle with adversarial contexts and predicting specific regional cuisines and languages. To support future research, we release a knowledge base with annotated food entries and images along with the VQA data.

Over 200 million people speak Indonesian, yet the language remains significantly underrepresented in preference-based research for large language models (LLMs). Most existing multilingual datasets are derived from English translations, often resulting in content that lacks cultural and linguistic authenticity. To address this gap, we introduce IndoPref, the first fully human-authored and multi-domain Indonesian preference dataset designed to evaluate the naturalness and quality of LLM-generated text. The dataset contains 522 prompts and yields 4,099 human-annotated pairwise preferences from comparisons across five instruction-tuned LLMs. All annotations are natively written in Indonesian with strong inter-annotator agreement, measured by Krippendorff’s alpha. Our benchmark spans 10 diverse categories, enabling practitioners to identify LLMs’ fine-grained strengths and weaknesses.

Advancements in technology and the increased use of digital data threaten individual privacy, especially in speech containing Personally Identifiable Information (PII). Therefore, systems that can remove or process privacy-sensitive data in speech are needed, particularly for low-resource transcripts. These transcripts are minimally annotated or labeled automatically, which is less precise than human annotation. However, using them can simplify the development of de-identification systems in any language. In this study, we develop and evaluate an efficient speech de-identification system. We create an Indonesian speech dataset containing sensitive private information and design a system with three main components: speech recognition, information extraction, and masking. To enhance performance in low-resource settings, we incorporate transcription data in training, use data augmentation, and apply weakly supervised learning. Our results show that our techniques significantly improve privacy detection performance, with approximately 29% increase in F1 score, 20% in precision, and 30% in recall with minimally labeled data.

Developing dialogue summarization for extremely low-resource languages is a challenging task. We introduce NusaDialogue, a dialogue summarization dataset for three underrepresented languages in the Malayo-Polynesian language family: Minangkabau, Balinese, and Buginese. NusaDialogue covers 17 topics and 185 subtopics, with annotations provided by 73 native speakers. Additionally, we conducted experiments using fine-tuning on a specifically designed medium-sized language model for Indonesian, as well as zero- and few-shot learning on various multilingual large language models (LLMs). The results indicate that, for extremely low-resource languages such as Minangkabau, Balinese, and Buginese, the fine-tuning approach yields significantly higher performance compared to zero- and few-shot prompting, even when applied to LLMs with considerably larger parameter sizes.

Auto-regressive inference of transformers benefit greatly from Key-Value (KV) caching, but can lead to major memory bottlenecks as model size, batch size, and sequence length grow at scale. We introduce Multi-Layer Key-Value (MLKV) sharing, a novel approach extending KV sharing across transformer layers to reduce memory usage beyond what was possible with Multi-Query Attention (MQA) and Grouped-Query Attention (GQA). Evaluations on various NLP benchmarks and inference metrics using uptrained Pythia-160M variants demonstrate that MLKV significantly reduces memory usage with minimal performance loss, reducing KV cache size down to a factor of 6x compared to MQA. These results highlight MLKV’s potential for efficient deployment of transformer models at scale.

Cross-lingual transfer in natural language processing (NLP) models enhances multilingual performance by leveraging shared linguistic knowledge. However, traditional methods that process all data simultaneously often fail to mimic real-world scenarios, leading to challenges like catastrophic forgetting, where fine-tuning on new tasks degrades performance on previously learned ones. Our study explores this issue in multilingual contexts, focusing on linguistic differences affecting representational learning rather than just model parameters. We experiment with 52 languages using LoRA adapters of varying ranks to evaluate non-shared, partially shared, and fully shared parameters. Our aim is to see if parameter sharing through adapters can mitigate forgetting while preserving prior knowledge. We find that languages using non-Latin scripts are more susceptible to catastrophic forgetting, whereas those written in Latin script facilitate more effective cross-lingual transfer.

Natural Language Explanation (NLE) aims to elucidate the decision-making process by providing detailed, human-friendly explanations in natural language. It helps demystify the decision-making processes of large vision-language models (LVLMs) through the use of language models. While existing methods for creating a Vision Question-Answering with Natural Language Explanation (VQA-NLE) datasets can provide explanations, they heavily rely on human annotations that are time-consuming and costly. In this study, we propose a novel approach that leverages LVLMs to efficiently generate high-quality synthetic VQA-NLE datasets. By evaluating our synthetic data samples, we showcase how advanced prompting techniques can lead to the production of high-quality VQA-NLE data. Our findings indicate that this proposed method achieves up to 20x faster than human annotation, with only a minimal decrease in qualitative metrics, achieving robust quality that is nearly equivalent to human-annotated data. Furthermore, we show that incorporating visual prompts significantly enhances the relevance of text generation. Our study paves the way for a more efficient and robust automated generation of multi-modal NLE data, offering a promising solution to the problem.

Pretrained language models (PLMs) have shown remarkable generalization toward multiple tasks and languages. Nonetheless, the generalization of PLMs towards unseen languages is poor, resulting in significantly worse language performance, or even generating nonsensical responses that are comparable to a random baseline. This limitation has been a longstanding problem of PLMs raising the problem of diversity and equal access to language modeling technology. In this work, we solve this limitation by introducing LinguAlchemy, a regularization technique that incorporates various aspects of languages covering typological, geographical, and phylogenetic constraining the resulting representation of PLMs to better characterize the corresponding linguistics constraints. LinguAlchemy significantly improves the accuracy performance of mBERT and XLM-R on unseen languages by ~18% and ~2%, respectively compared to fully finetuned models and displaying a high degree of unseen language generalization. We further introduce AlchemyScale and AlchemyTune, extension of LinguAlchemy which adjusts the linguistic regularization weights automatically, alleviating the need for hyperparameter search. LinguAlchemy enables better cross-lingual generalization to unseen languages which is vital for better inclusivity and accessibility of PLMs.

Large Language Models (LLMs) demonstrate strong machine translation capabilities on languages they are trained on. However, the impact of factors beyond training data size on translation performance remains a topic of debate, especially concerning languages not directly encountered during training. Our study delves into Llama2’s translation capabilities. By modeling a linear relationship between linguistic feature distances and machine translation scores, we ask ourselves if there are potentially better central languages for LLMs other than English. Our experiments show that the 7B Llama2 model yields above 10 BLEU when translating into all languages it has seen, which rarely happens for languages it has not seen. Most translation improvements into unseen languages come from scaling up the model size rather than instruction tuning or increasing shot count. Furthermore, our correlation analysis reveals that syntactic similarity is not the only linguistic factor that strongly correlates with machine translation scores. Interestingly, we discovered that under specific circumstances, some languages (e.g. Swedish, Catalan), despite having significantly less training data, exhibit comparable correlation levels to English. These insights challenge the prevailing landscape of LLMs, suggesting that models centered around languages other than English could provide a more efficient foundation for multilingual applications.

Southeast Asia (SEA) is a region rich in linguistic diversity and cultural variety, with over 1,300 indigenous languages and a population of 671 million people. However, prevailing AI models suffer from a significant lack of representation of texts, images, and audio datasets from SEA, compromising the quality of AI models for SEA languages. Evaluating models for SEA languages is challenging due to the scarcity of high-quality datasets, compounded by the dominance of English training data, raising concerns about potential cultural misrepresentation. To address these challenges, through a collaborative movement, we introduce SEACrowd, a comprehensive resource center that fills the resource gap by providing standardized corpora in nearly 1,000 SEA languages across three modalities. Through our SEACrowd benchmarks, we assess the quality of AI models on 36 indigenous languages across 13 tasks, offering valuable insights into the current AI landscape in SEA. Furthermore, we propose strategies to facilitate greater AI advancements, maximizing potential utility and resource equity for the future of AI in Southeast Asia.

Indonesia is home to a diverse linguistic landscape, where individuals seamlessly transition between Indonesian, English, and local dialects in their everyday conversations—a phenomenon known as code-switching. Understanding and accommodating this linguistic fluidity is essential, particularly in the development of accurate speech recognition systems. However, tackling code-switching in Indonesian poses a challenge due to the scarcity of paired code-switching data. Thus, this study endeavors to address Indonesian-English code-switching in speech recognition, leveraging unlabeled data and employing a semi-supervised technique known as the machine speech chain. Our findings demonstrate that the machine speech chain method effectively enhances Automatic Speech Recognition (ASR) performance in recognizing code-switching between Indonesian and English, utilizing previously untapped resources of unlabeled data.

Large language models (LLMs) show remarkable human-like capability in various domains and languages. To bridge this quality gap, we introduce Cendol, a collection of Indonesian LLMs encompassing both decoder-only and encoder-decoder architectures across a range of model sizes. We highlight Cendol’s effectiveness across a diverse array of tasks, attaining ~20% improvement, and demonstrate its capability to generalize to unseen tasks and indigenous languages of Indonesia. Furthermore, Cendol models showcase improved human favorability despite their limitations in capturing indigenous knowledge and cultural values in Indonesia. In addition, we discuss the shortcomings of parameter-efficient tunings, such as LoRA, for language adaptation. Alternatively, we propose the usage of vocabulary adaptation to enhance efficiency. Lastly, we evaluate the safety of Cendol and showcase that safety in pre-training in one language such as English is transferable to low-resource languages, such as Indonesian, even without RLHF and safety fine-tuning.

Constituency parsing is an important task of informing how words are combined to form sentences. While constituency parsing in English has seen significant progress in the last few years, tools for constituency parsing in Indonesian remain few and far between. In this work, we publish ICON (Indonesian CONstituency treebank), the hitherto largest publicly-available manually-annotated benchmark constituency treebank for the Indonesian language with a size of 10,000 sentences and approximately 124,000 constituents and 182,000 tokens, which can support the training of state-of-the-art transformer-based models. We establish strong baselines on the ICON dataset using the Berkeley Neural Parser with transformer-based pre-trained embeddings, with the best performance of 88.85% F1 score coming from our own version of SpanBERT (IndoSpanBERT). We further analyze the predictions made by our best-performing model to reveal certain idiosyncrasies in the Indonesian language that pose challenges for constituency parsing.

Despite being the world’s fourth-most populous country, the development of spoken language technologies in Indonesia still needs improvement. Most automatic speech recognition (ASR) systems that have been developed are still limited to transcribing the exact word-by-word, which, in many cases, consists of ambiguous sentences. In fact, speakers use prosodic characteristics of speech to convey different interpretations, which, unfortunately, these systems often ignore. In this study, we attempt to resolve structurally ambiguous utterances into unambiguous texts in Indonesian using prosodic information. To the best of our knowledge, this might be the first study to address this problem in the ASR context. Our contributions include (1) collecting the Indonesian speech corpus on structurally ambiguous sentences; (2) conducting a survey on how people disambiguate structurally ambiguous sentences presented in both text and speech forms; and (3) constructing an Indonesian ASR and meaning interpretation system by utilizing both cascade and direct approaches to map speech to text, along with two additional prosodic information signals (pause and pitch). The experimental results reveal that it is possible to disambiguate these utterances. In this study, the proposed cascade system, utilizing Mel-spectrograms concatenated with F0 and energy as input, achieved a disambiguation accuracy of 79.6%, while the proposed direct system with the same input yielded an even more impressive disambiguation accuracy of 82.2%.

We present NusaCrowd, a collaborative initiative to collect and unify existing resources for Indonesian languages, including opening access to previously non-public resources. Through this initiative, we have brought together 137 datasets and 118 standardized data loaders. The quality of the datasets has been assessed manually and automatically, and their value is demonstrated through multiple experiments.NusaCrowd’s data collection enables the creation of the first zero-shot benchmarks for natural language understanding and generation in Indonesian and the local languages of Indonesia. Furthermore, NusaCrowd brings the creation of the first multilingual automatic speech recognition benchmark in Indonesian and the local languages of Indonesia. Our work strives to advance natural language processing (NLP) research for languages that are under-represented despite being widely spoken.

Natural language generation (NLG) benchmarks provide an important avenue to measure progress and develop better NLG systems. Unfortunately, the lack of publicly available NLG benchmarks for low-resource languages poses a challenging barrier for building NLG systems that work well for languages with limited amounts of data. Here we introduce IndoNLG, the first benchmark to measure natural language generation (NLG) progress in three low-resource—yet widely spoken—languages of Indonesia: Indonesian, Javanese, and Sundanese. Altogether, these languages are spoken by more than 100 million native speakers, and hence constitute an important use case of NLG systems today. Concretely, IndoNLG covers six tasks: summarization, question answering, chit-chat, and three different pairs of machine translation (MT) tasks. We collate a clean pretraining corpus of Indonesian, Sundanese, and Javanese datasets, Indo4B-Plus, which is used to pretrain our models: IndoBART and IndoGPT. We show that IndoBART and IndoGPT achieve competitive performance on all tasks—despite using only one-fifth the parameters of a larger multilingual model, mBART-large (Liu et al., 2020). This finding emphasizes the importance of pretraining on closely related, localized languages to achieve more efficient learning and faster inference at very low-resource languages like Javanese and Sundanese.

Although Indonesian is known to be the fourth most frequently used language over the internet, the research progress on this language in natural language processing (NLP) is slow-moving due to a lack of available resources. In response, we introduce the first-ever vast resource for training, evaluation, and benchmarking on Indonesian natural language understanding (IndoNLU) tasks. IndoNLU includes twelve tasks, ranging from single sentence classification to pair-sentences sequence labeling with different levels of complexity. The datasets for the tasks lie in different domains and styles to ensure task diversity. We also provide a set of Indonesian pre-trained models (IndoBERT) trained from a large and clean Indonesian dataset (Indo4B) collected from publicly available sources such as social media texts, blogs, news, and websites. We release baseline models for all twelve tasks, as well as the framework for benchmark evaluation, thus enabling everyone to benchmark their system performances.