United States

This paper presents SETU-ADAPT&#39;s submissions to the WMT 2024 Biomedical Shared Task, where we participated for the language pairs English-to-French and English-to-German. Our approach focused on fine-tuning Large Language Models (LLMs), using in-domain and synthetic data, employing different data retrieval strategies. We introduce a novel MT framework, involving three autonomous agents: a Translator Agent, an Evaluator Agent and a Reviewer Agent. We present our findings and report the quality of the outputs.

EMNLP 2024

The SETU-ADAPT Submission for WMT 24 Biomedical Shared Task

llms

agents

data-augmentation

## Welcome to EMNLP 2024! 
We are excited to welcome you to one of the most prominent conferences in the field of Natural Language Processing. This year, EMNLP 2024 is being held in a hybrid format,
offering both virtual and in-person participation in beautiful Miami. Due to a record-breaking number of submissions, we've expanded the total number of accepted papers to accommodate more cutting-edge research from around the globe.
### [Conference Handbook](https://drive.google.com/file/d/1WPROgxjLAC96AJL7Ugy0tEnYm7dkrbHt/view?usp=sharing)

You are required to register for this event. **Please register [here](https://2024.emnlp.org/registration/).** The EMNLP 2024 event page on Underline will be open to public one week prior to the event.

Please register!

EMNLP 2024 will take place in Miami, Florida from Nov 12th to Nov 16th, 2024, at the Hyatt Regency Miami Hote and on Underline for remote participants.

This paper presents SETU-ADAPT's submissions to the WMT 2024 Biomedical Shared Task, where we participated for the language pairs English-to-French and English-to-German. Our approach focused on fine-tuning Large Language Models (LLMs), using in-domain and synthetic data, employing different data retrieval strategies. We introduce a novel MT framework, involving three autonomous agents: a Translator Agent, an Evaluator Agent and a Reviewer Agent. We present our findings and report the quality of the outputs.

The presentation explores the WMT 2024 English-to-Low-Resource Multimodal Translation Task, where seven teams submitted translations for both automatic and human evaluation. It delves into the various approaches adopted by each team, along with an analysis of their results.

Findings of WMT2024 English-to-Low Resource Multimodal Translation Task

An overview of non-repetitive translation task

Findings of the WMT 2024 Shared Task on Non-Repetitive Translation

Translating for low-resource languages is challenging due to limited quality data. We improved translation performance by integrating target sentence length as an additional feature to source sentence during training. We developed transformer models and evaluated across 8 language directions (English <=> Assamese, Manipuri, Khasi, and Mizo), using four different length encoding methods. Comparing these models to the baseline, we submitted two systems per language direction and present our findings here.​

A3-108 Controlling Token Generation in Low Resource Machine Translation Systems

This paper describes the submission systems of DLUT-NLP team for the WMT24 low-resource Indic language translation shared task. We participated in the translation task of four language pairs, including en-as, en-mz, en-kha, en-mni.

DLUT-NLP Machine Translation Systems for WMT24 Low-Resource Indic Language Translation

Machine Translation for low-resource languages poses significant challenges, primarily due to the limited availability of data.The WMT24 Low-Resource Indic Neural Machine Translation task challenges us to employ innovative techniques to improve machine translation for low-resource Indian languages. Our proposed solution leverages advancements in neural machine translation, focusing on methodologies such as back-translation and fine-tuning. By fine-tuning pretrained models like mBART, we achieved significant progress in translating languages such as Manipuri and Khasi. The best score was achieved for the English-to-Khasi (en-kh) primary model, with the highest BLEU score of 0.0492, chrF score of 0.3316, and METEOR score of 0.2589 (on scale of 0 to 1) and comparable scores for other language pairs.

MTNLP-IIITH: Machine Translation for Low-Resource Indic Languages

Our submission focuses on Assamese-to-English and English-to- Assamese translation, using LLMs as a baseline further fine-tuned with in-domain and synthetic data through various data retrieval strategies. To address low-resource challenges, we incorporated synthetic data generation and few-shot prompting, significantly improving translation accuracy for both language pairs. We present our findings and report on the quality of outputs achieved through these methods.

The SETU-ADAPT Submissions to the WMT24 Low-Resource Indic Language Translation Task

In this paper, we describe our system for the WMT 24 shared task of Low-Resource Indic Language Translation. We consider eng ↔ {as, kha, lus, mni} as participating language pairs. In this shared task, we explore finetun- ing of a pre-trained machine translation model, where the pretraining objective includes align- ment of embeddings of tokens from the 22 scheduled Indian languages by a carefully con- structed alignment augmentation strategy (Lin et al., 2020). Our primary system1 is based on language-specific finetuning on this pre-trained model. We achieve chrF2 scores of 50.6, 42.3, 54.9, and 66.3 on the official public test sets for eng→as, eng→kha, eng→lus, eng→mni re- spectively. We also explore multilingual train- ing with/without language grouping and freez- ing of encoder and/or embedding layers.

NLIP_Lab-IITH Low-Resource MT System for WMT24 Indic MT Shared Task

This paper presents our machine translation system that was developed for the WAT2024 MultiIndic MT shared task. We built our system for the Sindhi-English language pair. We developed two MT systems. The first system was our baseline system where Sindhi was translated into English. In the second system, we used Hindi as a pivot for the translation of text. In both the cases, we had identified the name entities and translated them into English as a preprocessing step. Once this was done, the standard NMT process was followed to train and generate MT outputs for the task. The systems were tested on the hidden dataset of the shared task

System Description of BV-SLP for Sindhi-English Machine Translation in MultiIndic22MT 2024 Shared Task

This paper describes NLIP Lab’s multilingual machine translation system for the WAT24 shared task on multilingual Indic MT task for 22 scheduled languages belonging to 4 lan- guage families. We explore pre-training for In- dic languages using alignment agreement objec- tives. We utilize bi-lingual dictionaries to sub- stitute words from source sentences. Further- more, we fine-tuned language direction-specific multilingual translation models using small and high-quality seed data. Our primary submis- sion is a 243M parameters multilingual trans- lation model covering 22 Indic languages. In the IN22-Gen benchmark, we achieved an av- erage chrF++ score of 46.80 and 18.19 BLEU score for the En-Indic direction. In the Indic- En direction, we achieved an average chrF++ score of 56.34 and 30.82 BLEU score. In the In22-Conv benchmark, we achieved an aver- age chrF++ score of 43.43 and BLEU score of 16.58 in the En-Indic direction, and in the Indic- En direction, we achieved an average of 52.44 and 29.77 for chrF++ and BLEU respectively. Our model1 is competitive with IndicTransv1 (474M parameter model).

NLIP_Lab-IITH Multilingual MT System for WAT24 MT Shared Task

Multimodal machine translation leverages multiple data modalities to enhance translation quality, particularly for low-resourced languages. This paper uses a multimodal model that integrates visual information with textual data to improve translation accuracy from English to Hindi, Malayalam, Bengali, and Hausa. This approach employs a gated fusion mechanism to effectively combine the outputs of textual and visual encoders, enabling more nuanced translations that consider both language and contextual visual cues. The model's performance was evaluated against the text-only machine translation model based on BLEU, ChrF2 and TER. Experimental results demonstrate that the multimodal approach consistently outperforms the text-only baseline, highlighting the potential of integrating visual information in low-resourced language translation tasks.

Downloads

Next from EMNLP 2024

Findings of WMT2024 English-to-Low Resource Multimodal Translation Task

Similar lecture

The SETU-ADAPT Submissions to WMT 2024 Chat Translation Tasks