LQM: Linguistically Motivated Multidimensional Quality Metrics for Machine Translation

The real gap LQM addresses is that existing quality metrics treat translation errors as language-agnostic, but Arabic diglossia means a technically accurate rendering can still be socially or pragmatically wrong in ways no surface-level metric catches. The six-level hierarchy is an attempt to make those failures legible to automated pipelines, not just human annotators.

This connects directly to the cluster of evaluation-reliability work we covered in mid-April. The 'Fabricator or dynamic translator?' piece (arXiv, April 16) examined how LLMs produce spurious or misleading output during translation without any framework for categorizing what kind of failure is occurring. LQM is essentially proposing the taxonomy that work lacked. Separately, 'Context Over Content: Exposing Evaluation Faking in Automated Judges' raised the concern that automated evaluators are already unreliable on standard benchmarks, which makes the case for richer error taxonomies more urgent, not less.

The meaningful test is whether LQM's taxonomy gets adopted by any of the commercial MT evaluation pipelines cited in the 'Fabricator or dynamic translator?' study. If it remains a corpus annotation tool used only in academic settings within the next 12 months, the framework's practical reach is limited regardless of its linguistic rigor.