LQM: Linguistically Motivated Multidimensional Quality Metrics for Machine Translation

LQM overview connecting six linguistic levels to translation error examples: sociolinguistics and code or register selection in Moroccan Arabic; pragmatics and illocutionary force in German; semantics and polysemy in French; morphosyntax and verbal features in Mandarin; orthography and typos in Korean; and graphetics and character encoding in Spanish. — **LQM at a glance.** Six linguistic levels, illustrated with translation errors across languages. SRC: source; TRG: reference translation; MT: machine translation.

LQM Error Explorer

About the project

LQM is a language-agnostic framework for diagnosing machine translation errors across six linguistic levels: sociolinguistics, pragmatics, semantics, morphosyntax, orthography, and graphetics. It captures dialect, cultural, and contextual issues that broad evaluation schemes can miss.

The project evaluates six large language models on a bidirectional corpus of 3,850 sentences spanning seven Arabic dialects, using expert span-level annotations and severity-weighted quality scores. This companion explorer lets you filter the annotations using the LQM or MQM taxonomy and inspect highlighted error spans.

In this explorer, we present LQM framework applied on the following languages/dialects: Egyptian Arabic Dialect (EGY), English (ENG), Jordanian Arabic Dialect (JOR), Mauritanian Arabic Dialect (MAU), Moroccan Arabic Dialect (MOR), Palestinian Arabic Dialect (PAL), Emerati Arabic Dialect (UAE), and Yemeni Arabic Dialect (YEM).

Read the paper: Magdy et al. (2026) ↗Cite this work

—

matching examples

—

matching errors

—

models represented

—

avg. errors / example

Filtered overview

Select a bar label to filter.

Examples by model

Errors by category

Errors by severity

Examples

Per page

Citation

@inproceedings{magdy-etal-2026-lqm,
  title = "{LQM}: Linguistically Motivated Multidimensional Quality Metrics for Machine Translation",
  author = "Magdy, Samar M. and
    Alwajih, Fakhraddin and
    El Mekki, Abdellah and
    El Sayed, Wesam and
    Abdul-Mageed, Muhammad",
  booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {ACL} 2026",
  month = jul,
  year = "2026",
  address = "San Diego, California, United States",
  publisher = "Association for Computational Linguistics",
  pages = "40470--40493",
  doi = "10.18653/v1/2026.findings-acl.2012",
  url = "https://aclanthology.org/2026.findings-acl.2012/"
}

DOI: 10.18653/v1/2026.findings-acl.2012

Disclaimer: This data excludes records that were used for the test set of AlexandriaX-2026, Subtask 3: Dialectal Arabic MT Error Detection and Classification.