Thailand

Despite the increasing popularity of multilingualism within the NLP community, numerous languages continue to be underrepresented due to the lack of available resources. Our work addresses this gap by introducing experiments on cross-lingual transfer between 158 high-resource (HR) and 31 low-resource (LR) languages. We mainly focus on extremely LR languages, some of which are first presented in research works. Across $158*31$ HR–LR language pairs, we investigate how continued pretraining on different HR languages affects the mT5 model&#39;s performance in representing LR languages in the LM setup. Our findings surprisingly reveal that the optimal language pairs with improved performance do not necessarily align with direct linguistic motivations, with subtoken overlap playing a more crucial role. Our investigation indicates that specific languages tend to be almost universally beneficial for pretraining (\textit{super donors}), while others benefit from pretraining with almost any language (\textit{super recipients}). This pattern recurs in various setups and is unrelated to the linguistic similarity of HR-LR pairs. Furthermore, we perform evaluation on two downstream tasks, part-of-speech (POS) tagging and machine translation (MT), showing how HR pretraining affects LR language performance. We thoroughly explore and discuss the experimental results of this study.

ACL 2024

Super donors and super recipients: Studying cross-lingual transfer between high-resource and low-resource languages

low-resource languages

transfer learning

machine translation

Despite the increasing popularity of multilingualism within the NLP community, numerous languages continue to be underrepresented due to the lack of available resources. Our work addresses this gap by introducing experiments on cross-lingual transfer between 158 high-resource (HR) and 31 low-resource (LR) languages. We mainly focus on extremely LR languages, some of which are first presented in research works. Across $158*31$ HR–LR language pairs, we investigate how continued pretraining on different HR languages affects the mT5 model's performance in representing LR languages in the LM setup. Our findings surprisingly reveal that the optimal language pairs with improved performance do not necessarily align with direct linguistic motivations, with subtoken overlap playing a more crucial role. Our investigation indicates that specific languages tend to be almost universally beneficial for pretraining (\textit{super donors}), while others benefit from pretraining with almost any language (\textit{super recipients}). This pattern recurs in various setups and is unrelated to the linguistic similarity of HR-LR pairs. Furthermore, we perform evaluation on two downstream tasks, part-of-speech (POS) tagging and machine translation (MT), showing how HR pretraining affects LR language performance. We thoroughly explore and discuss the experimental results of this study.

workshop paper

### Welcome!
The 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024) will take place in Bangkok, Thailand from August 11th to 16th, 2024. Our Virtual Poster Sessions will take place online Thursday, August 22, 2024.

You are required to register for this event. **Please register [here](https://2024.aclweb.org/registration). **

If you have already registered, please check your inbox for an email from Underline granting you access to ACL 2024 content.

Please register!

The 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024) will take place in Bangkok, Thailand from August 11th to 16th, 2024. More information will be announced soon.

In Machine Translation, various tokenisers are used to segment inputs before training a model. Despite tokenisation being mostly considered a solved problem for languages such as English, it is still unclear as to how effective different tokenisers are for morphologically rich languages. 
This study aims to explore how different approaches to tokenising Maltese impact machine translation results on the English-Maltese language pair.
We observed that the OPUS-100 dataset has tokenisation inconsistencies in Maltese. We empirically found that training models on the original OPUS-100 dataset led to the generation of sentences with these issues.
We therefore release an updated version of the OPUS-100 parallel English-Maltese dataset, referred to as OPUS-100-Fix, fixing these inconsistencies in Maltese by using the MLRS tokeniser. 
We show that after fixing the inconsistencies in the dataset, results on the fixed test set increase by 2.49 BLEU points over models trained on the original OPUS-100. We also experiment with different tokenisers, including BPE and SentencePiece to find the ideal tokeniser and vocabulary size for our setup, which was shown to be BPE with a vocabulary size of 8,000. Finally, we train different models in both directions for the ENG-MLT language pair using OPUS-100-Fix by training models from scratch as well as fine-tuning other pre-trained models, namely mBART-50 and NLLB, where a finetuned NLLB model performed the best.

Tokenisation in Machine Translation Does Matter: The impact of different tokenisation approaches for Maltese

We propose the first MT models for Innu-Aimun, an Indigenous language in Eastern Canada, in an effort to provide assistance tools for translation and language learning. This project is carried out in collaboration with an Innu community school and involves the participation of students in Innu-Aimun translation, within the framework of a meaningful consideration of Indigenous perspectives.
Our contributions in this paper result from the three initial stages of this project. First, we aim to align bilingual Innu-Aimun/French texts with collaboration and participation of Innu-Aimun locutors. Second, we present the training and evaluation results of the MT models (both statistical and neural) based on these aligned corpora. And third, we collaboratively analyze some of the translations resulting from the MT models.
We also see these developments for Innu-Aimun as a useful case study for answering a larger question: in a context where few aligned bilingual sentences are available for an Indigenous language, can cultural texts such as literature and poetry be used in the development of MT models?

Machine Translation Through Cultural Texts: Can Verses and Prose Help Low-Resource Indigenous Models?

This paper explores the impact of different back-translation approaches on machine translation for Ladin, specifically the Val Badia variant. Given the limited amount of parallel data available for this language (only 18k Ladin--Italian sentence pairs), we investigate the performance of a multilingual neural machine translation model fine-tuned for Ladin--Italian. In addition to the available authentic data, we synthesise further translations by using three different models: a fine-tuned neural model, a rule-based system developed specifically for this language pair, and a large language model. Our experiments show that all approaches achieve comparable translation quality in this low-resource scenario, yet round-trip translations highlight differences in model performance.

Rule-Based, Neural and LLM Back-Translation: Comparative Insights from a Variant of Ladin

African languages are not well-represented in Natural Language Processing (NLP). The main reason is a lack of resources for training models. Low-resource languages, such as Amharic and Ge'ez, cannot benefit from modern NLP methods because of the lack of high-quality datasets. This paper presents AGE, an open-source tripartite alignment of Amharic, Ge'ez, and English parallel dataset. Additionally, we introduced a novel, 1,000 Ge'ez-centered sentences sourced from areas such as news and novels. Furthermore, we developed a model from a multilingual pre-trained language model, which brings 12.29 and 30.66 for English-Ge'ez and Ge'ez to English, respectively, and 9.39 and 12.29 for Amharic-Ge'ez and Ge'ez-Amharic respectively.

AGE: Amharic, Ge’ez and English Parallel Dataset

Using large language models, this paper presents techniques to improve extremely low-resourced indigenous language translations. Our approaches are grounded in the use of (1) the presence of a datastore consisting of a limited number of parallel translation examples, (2) the inherent capabilities of LLMs like GPT-3.5, and (3) a word-level translation dictionary. We harness the potential of LLMs and in-context learning techniques in such a setting for using LLM as universal translators for extremely low-resourced languages. Our methodology hinges on utilizing LLMs as language compilers for selected language pairs, hypothesizing that they could internalize syntactic structures to facilitate accurate translation. We introduce three techniques: KNN-Prompting with Retrieved Prompting Context, Chain-of-Thought Prompting, and Learning-from-Mistakes Prompting, with the last method addressing past errors. The evaluation results suggest that, even with limited corpora, LLMs, when paired with proper prompting, can effectively translate extremely low-resource languages.

Learning-From-Mistakes Prompting for Indigenous Language Translation

Cross-lingual classification poses a significant challenge in Natural Language Processing (NLP), especially when dealing with languages with scarce training data. This paper delves into the adaptation of ensemble learning to address this challenge, specifically for disaster-related social media texts. Initially, we employ Machine Translation to generate a parallel corpus in the target language to mitigate the issue of data scarcity and foster a robust training environment. Following this, we implement the bagging ensemble technique, integrating multiple classifiers into a cohesive model that demonstrates enhanced performance over individual classifiers. Our experimental results reveal significant improvements in adapting models for Arabic, utilising only English training data and markedly outperforming models intended for linguistically similar languages to English, with our ensemble model achieving an accuracy and F1 score of 0.78 when tested on original Arabic data. This research makes a substantial contribution to the field of cross-lingual classification, establishing a new benchmark for enhancing the effectiveness of language transfer in linguistically challenging scenarios.

Adopting Ensemble Learning for Cross-lingual Classification of Crisis-related Text On Social Media

This paper investigates the finetuning of end-to-end models for bidirectional Estonian-English and Estonian-Russian conversational speech-to-text translation. Due to the limited availability of speech translation data for Estonian, we created additional training data by web scraping and synthesizing data from speech recognition datasets using machine translation. We evaluated three publicly available end-to-end models: Whisper, OWSM 3.1, and SeamlessM4T. Our results indicate that fine-tuning with synthetic data enhances translation accuracy by a large margin, with SeamlessM4T matching or surpassing cascaded speech translation systems that use state-of-the-art speech recognition and machine translation models.

Finetuning End-to-End Models for Estonian Conversational Spoken Language Translation

Assessing the performance of machine translation systems is of critical value, especially to languages with lower resource availability.
Due to the large evaluation effort required by the translation task, studies often compare new systems against single systems or commercial solutions. Consequently, determining the best-performing system for specific languages is often unclear. 
This work benchmarks publicly available translation systems across 4 datasets and 26 languages, including low-resource languages. We consider both effectiveness and efficiency in our evaluation.
Our results are made public through BENG---a FAIR benchmarking platform for Natural Language Generation tasks.

Benchmarking of Low-Resource Machine Translation Systems

The Rosetta Balcanica is an ongoing effort in resource expansion for low-resource Western Balkans languages. This effort focuses on discovering and using accurately translated, officially mapped, and curated parallel language resources and their preparation and use as neural machine translation (NMT) datasets. Some of the guiding principles, practices, and methods employed by Rosetta Balcanica are generalizable and could apply to other low-resource language resource expansion efforts. With this goal in mind, we present our rationale and approach to discovering and using meticulously translated and officially curated low-resource language resources and our use of these resources to develop a parallel ``gold standard'' translation training resource. Secondly, we describe our specific methodology for NMT dataset development from these resources and its publication to a widely-used and accessible repository for natural language processing (\textit{Hugging Face Hub}). Finally, we discuss the trade-offs and limitations of our current approach, and the roadmap for future development and the expansion of the current Rosetta Balcanica language resource.

Rosetta Balcanica: Deriving a "Gold Standard'' Neural Machine Translation (NMT) Parallel Dataset from High-Fidelity Resources for Western Balkan Languages

Large Language Models (LLMs) have demonstrated exceptional performances in a wide range of natural language processing tasks. However, their success does not always extend to machine translation, particularly in challenging scenarios such as translating low-resource languages. This study investigates the multilingual capability of LLMs, with a case study on Irish, an extremely low-resource language, focusing on translation tasks between English and Irish. We propose a dynamic, efficient language adaptation framework for English-centric LLMs, which involves layer-specific adjustments and subsequent fine-tuning for machine translation. Our findings highlight several key insights: (1) different layers in the LLM serve distinct functions such as language understanding and task reasoning, (2) effective translation requires extensive pre-training on both source and target languages, and (3) targeted fine-tuning for machine translation leads to significant improvements of 36.7% for English to Irish and 133.4% for Irish to English compared to the previous state-of-the-art.

Premium content

Super donors and super recipients: Studying cross-lingual transfer between high-resource and low-resource languages

Downloads

Next from ACL 2024

Tokenisation in Machine Translation Does Matter: The impact of different tokenisation approaches for Maltese

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES