Morocco

This paper describes the submission of Team DevLake for the LoResMT 2026 Shared Task on Russian-Bashkir machine translation. We conducted a comprehensive comparative study of three distinct neural architectures: NLLB-200 (1.3B), M2M-100 (418M), and MarianMT (77M). To overcome hardware constraints, we employed parameter-efficient fine-tuning techniques (QLoRA) and extensive data filtering using a domain-specific BERT-based classifier. Our experiments demonstrate that the presence of the target language (Bashkir) in the model&#39;s pre-training data is the decisive factor for performance. Our best system, a fine-tuned NLLB-200-1.3B model augmented with exact match retrieval, achieved a CHRF++ score of 52.67. We also report on negative results with custom tokenization for smaller models, providing insights into the limitations of vocabulary adaptation without extensive pre-training.

EACL 2026

DevLake at LoResMT 2026: The Impact of Pre-training and Model Scale on Russian-Bashkir Low-Resource Translation

russian-bashkir

qlora

nllb

peft

large language models

fine-tuning

low-resource

machine translation

transformer

#### *Message from the General Chair, Aline Villavicencio*
I’m delighted and honoured to welcome you to the 19th Conference of the European Chapter of the Association for Computational Linguistics (EACL 2026), taking place in the beautiful city of Rabat, in Morocco, in March 24-29, 2026. EACL is the flagship European conference of the Association and EACL 2026 proudly continues our field’s tradition of excellence in scholarship, innovation, and inclusivity. I am deeply grateful to the many volunteers whose dedication, generosity, and tireless efforts have made this conference possible.
For the first time EACL is being hosted in the African continent. This is an important milestone for our community, and we are grateful to our Moroccan hosts for enabling this historic moment by bringing this edition of EACL to Rabat. We are also delighted that the Second Arabic NLP School is co-located with EACL. We hope attendees enjoy this wonderful opportunity to strengthen ties with the Computational Linguistics communities across the African continent. *[Read full message](https://drive.google.com/file/d/14NlmHvwM6fPJuMmOvVh7K0vtQbEyv3SZ/view?usp=sharing)*<br><br>

<html><button style="display: inline-flex; align-items: center; justify-content: center; white-space: nowrap; border-radius: 9999px; font-weight: bold; background: #7c3aed; color: white; font-family: 'Space Grotesk', sans-serif; height: 40px; font-size: 16px; padding: 0 20px; border: none; cursor: pointer" onclick="window.open('https://underline.io/events/522/reception','_blank')">Go to Workshops and Tutorials Program</button></html>
<br><br><br>

To access this event page, you need to log in with the **email address you registered with**. <br>Access credentials will be sent to your email from Underline -  subject line "Welcome to EACL 2026". Please be sure to check your spam email folder if you do not see an email confirmation right away.

Please log in

To access this event page, you are required to register.
Please complete your registration to continue.

We recommend reading [**the registration information**](https://2026.eacl.org/registration/) first.

**Online Registration Form**: https://acl.swoogo.com/eacl2026

Registration Required

Welcome to the 19th Conference of the European Chapter of the Association for Computational Linguistics (EACL). EACL 2026 will be held in Rabat, Morocco, from March 24–29, 2026. 

This paper describes the submission of Team DevLake for the LoResMT 2026 Shared Task on Russian-Bashkir machine translation. We conducted a comprehensive comparative study of three distinct neural architectures: NLLB-200 (1.3B), M2M-100 (418M), and MarianMT (77M). To overcome hardware constraints, we employed parameter-efficient fine-tuning techniques (QLoRA) and extensive data filtering using a domain-specific BERT-based classifier. Our experiments demonstrate that the presence of the target language (Bashkir) in the model's pre-training data is the decisive factor for performance. Our best system, a fine-tuned NLLB-200-1.3B model augmented with exact match retrieval, achieved a CHRF++ score of 52.67. We also report on negative results with custom tokenization for smaller models, providing insights into the limitations of vocabulary adaptation without extensive pre-training.

workshop paper

This paper describes a submission to the LoResMT 2026 Shared Task for the Russian-Kazakh, Russian-Bashkir, and English-Chuvash tracks. The primary approach involves parameter-efficient fine-tuning (LoRA) of the Tencent HY-MT1.5-7B multilingual model. For the Russian-Kazakh and Russian-Bashkir pairs, LoRA adaptation was employed to correct the model's default Arabic script output to Cyrillic. For the extremely low-resource English-Chuvash pair, two strategies were compared: mixed training on authentic English-Chuvash and Russian-Chuvash data versus training exclusively on a synthetic English-Chuvash corpus created via pivoting through Russian. Baseline systems included NLLB 1.3B (distilled) for Russian-Kazakh and Russian-Bashkir, and Gemma 2 3B for English-Chuvash. Results demonstrate that adapting a strong multilingual backbone with LoRA yields significant improvements over baselines while successfully addressing script mismatch challenges. Code for training and inference is released at: https://github.com/defdet/low-resource-langs-mt-adapt

Script Correction and Synthetic Pivoting: Adapting Tencent HY-MT for Low-Resource Turkic Translation

We describe our submission to the Turkic languages translation challenge at LoResMT 2026, which focuses on translation from Russian into Kyrgyz. Our approach leverages parallel data, synthetic translations, a comprehensive filtering pipeline and a four-stage curriculum learning strategy. We compare our system with contemporary baselines and present the model that achieves a chrF++ score of 49.1 and takes first place in the competition.

Data-Centric Approach at the LoResMT 2026 Turkic Translation Challenge: Russian-Kyrgyz

The challenges of building speech-to-text translation (ST) systems (e.g., a relative lack of parallel speech–text data and robustness to noise in audio) are exacerbated for low-resource language pairs. In this work, we seek to improve low-resource ST by building on previous studies that regularize ST training with the connectionist temporal classification (CTC) loss. By systematically evaluating a diverse range of linguistic annotations as CTC labels across multiple auxiliary loss configurations, we improve speech translation systems for both low- and high-resource settings. These improvements over both a standard end-to-end ST system and a speech LLM indicate a need for continued research on regularizing speech representations in ST.

CTC Regularization for Low-Resource Speech-to-Text Translation

Noun compounds are generally considered an open challenge for NLP systems, given to the difficulty of interpreting the implicit semantic relation between modifier and head, although the advent of Large Language Models (LLMs) recently led to remarkable performance leaps. However, most evaluations have been carried out on English benchmarks.In our work, we test LLMs on compound semantics understanding in Chinese, adopting two different evaluation scenarios: an extrinsic evaluation in a Natural Language Inference task, and an intrinsic evaluation in which models are directly asked to predict the semantic relation linking the two constituents.Our results show that the bigger and more recent LLMs are able to surpass supervised baselines in the inference task, especially when tested under the few-shot setting. In the more challenging task of selecting the correct interpretation of the compounds out of a fine-grained typology of semantic relations between head and modifier, the best Chinese LLM (Qwen-plus) manages to select the correct option in about one third of the cases.

Large Language Models Put to the Test on Chinese Noun Compounds: Experi-
ments on Natural Language Inference and Compound Semantics

Figures of Speech (FOS) consist of multi-word phrases that are deeply intertwined with culture. While Neural Machine Translation (NMT) performs relatively well with the figurative expressions of high-resource languages, it often faces challenges when dealing with low-resource languages like Sinhala due to limited available data. To address this limitation, we introduce a corpus of 2,344 Sinhala figures of speech with cultural and cross-lingual annotations. We examine this dataset to classify the cultural origins of the figures of speech and to identify their cross-lingual equivalents. Additionally, we have developed a binary classifier to differentiate between two types of FOS in the dataset, achieving an accuracy rate of approximately 92%. We also evaluate the performance of existing LLMs on this dataset. Our findings reveal significant shortcomings in the current capabilities of LLMs, as these models often struggle to accurately convey idiomatic meanings. By making this dataset publicly available, we offer a crucial benchmark for future research in low-resource NLP and culturally aware machine translation.

SinFoS: A Parallel Dataset for Translating Sinhala Figures of Speech

We present the annotation of Swedish multiword expressions under the PARSEME annotation scheme, including a new release and a historical overview of previous releases. We provide an overview of the evolution of the Swedish datasets and of inter-annotator agreement. We discuss general guidelines and the development of Swedish-specific guidelines for particle verbs and multiword tokens, as well as additional challenges for the Swedish annotation. We also conduct an initial comparison of Swedish and other Germanic languages, identifying aspects where the PARSEME guidelines require revision to ensure better consistency across languages.

Swedish Multiword Expression Corpora in PARSEME

This paper presents the development of a corpus of annotated multiword expressions (MWEs) for Ukrainian. The resource covers four major categories of MWEs: verbal, nominal, adjectival/adverbial, and functional. We describe the methodology used for data selection, the annotation scheme, and the procedures employed during annotation. In addition, the paper discusses some specific types of MWE constructions, illustrating their usage with numerous examples and addressing complex and borderline cases. The resulting corpus is an important resource for linguistic studies and NLP tasks involving MWEs, and is publicly accessible https://gitlab.com/parseme/sharedtask-data/-/tree/master/2.0?ref_type=heads.

Ukrainian Multiword Expressions Corpus: Creation, Annotation, and Linguistic
Analysis

This study investigates whether eye-tracking measures predict if a word is the final token of a multi-word expression (MWE), focusing on two understudied MWE types: fixed expressions (e.g., \textit{due to}) and phrasal verbs (e.g., \textit{turn out}). Using mixed-effects logistic regression, we compared tokens in MWE contexts with the same tokens in non-MWE contexts. Results reveal a clear difference in processing. For fixed expressions, reading-time measures significantly predict MWEhood. In contrast, phrasal verbs show no consistent predictive effects. Additionally, we compared the reading-time models to models that included GPT-2 surprisal as a predictor. While surprisal does predict MWEhood, it fails to capture the distinction between types. These findings highlight the need to consider MWE typology in models of formulaic language processing.

Cognitive Signatures of Multi-Word Expressions: Reading-Time and Surprisal

In recent years, language models, both encoder-only and generative, have been applied to a variety of downstream NLP tasks, includingsequence labeling tasks like automatic multi-word expression identification (MWEI). Multiple studies show that, in general, fine-tunedencoder-only models like BERT tend to outperform pretrained generative LLMs on downstream tasks (Arzideh et al., 2025; Ochoa et al.,2025; Bucher and Martini, 2024; Sebok et al., 2025). However, such comparisons are sparse for MWEI, in particular for French, in partdue to the lack of comprehensive gold-standard datasets. In this study, we address this research gap by comparing CamemBERT with gpt-oss and Qwen3 for MWEI, using the French subcorpus of the newly released PARSEME dataset. CamemBERT outperforms both LLMs by large margins in precision, recall, and F1. We complement this numerical evaluation with a qualitative analysis of prediction errors.

Cheese it up: CamemBERT Outperforms Large Language Models for Identifica-
tion of French Multi-word Expressions

IT-systems generate log messages containing important information about the system's health. To gather information about system entities, we extract technical terms and proper nouns as multi-word expressions (MWEs) from a wide range of log messages from 16 different real systems. We apply Gries' information-theoretic approach which iteratively calculates the best MWE candidates using an eight-dimensional ranking method. These candidates are evaluated in an annotation study, achieving a precision of 66 %. This value is significantly higher than evaluations on general-purpose texts, demonstrating the higher occurrence of compound technical terms and proper nouns in log messages. The MWEs found can be used to reduce the number of nodes in a system behavior graph while increasing the information density of the nodes.

Downloads

Next from EACL 2026

Script Correction and Synthetic Pivoting: Adapting Tencent HY-MT for Low-Resource Turkic Translation

Similar lecture

Script Correction and Synthetic Pivoting: Adapting Tencent HY-MT for Low-Resource Turkic Translation