The project focuses on Amharic and Afaan Oromo and examines AI translation in public and institutional communication, where errors can have practical consequences.

Rather than evaluating translation only for fluency or sentence-level similarity, the study asks whether a translation preserves the meaning people need in order to understand instructions, interpret rights and obligations, follow deadlines, recognize warnings, and take the correct action.

Repository status

This is a development repository. A validated public dataset has not yet been released.

The current corpus is undergoing provenance review, independent human evaluation, adjudication, and inter-rater reliability analysis. Materials in active review should not be treated as a gold-standard benchmark.

The first planned public release is:

v1.0 — Human-Validated Evaluation Set

Research question

How does AI translation infrastructure shape communicative access for speakers of under-resourced languages?

The project evaluates translation across four dimensions:

Semantic fidelity

Does the translation preserve the core meaning of the source?

Pragmatic fidelity

Does it preserve communicative force, such as warnings, obligations, permissions, uncertainty, emphasis, and tone where those features matter?

Institutional fidelity

Does it preserve institutional meaning, including eligibility conditions, deadlines, rights, responsibilities, procedures, and administrative terminology?

Actionability

Could a reasonable user take the correct action based on the translated information?

These dimensions reflect the project's central claim:

Translation availability is not the same as communicative access.

A translation can sound fluent and still fail if it changes the practical meaning of a deadline, eligibility rule, warning, right, institutional term, or required action.

Languages

The initial study uses:

English as the principal source language
Amharic
Afaan Oromo

Amharic and Afaan Oromo are treated as analytically important cases. They are not intended to represent all under-resourced languages.

Translation systems in the initial study

Google Translate

Outputs are collected from the public Google Translate interface using a documented collection protocol.

Meta SeamlessM4T v2 Large

Model:

facebook/seamless-m4t-v2-large

This model is used as a research-model condition. Results from this study should not be interpreted as evidence about the translation system currently used in Facebook, Instagram, or other Meta production environments unless separate evidence establishes that connection.

Current development corpus

The current development corpus contains 40 aligned public-institutional communication units drawn from:

education
civic and election communication
health and public information

Official institutional translations are treated as institutional reference translations, not as unquestioned gold standards.

That distinction matters because official translations may themselves contain errors, inconsistencies, or different translation choices.

Human evaluation

Two reviewers independently evaluate each target-language translation using 1–5 scales for:

semantic fidelity
pragmatic fidelity
institutional fidelity
actionability

The review process is:

Reviewer 1 → Reviewer 2 → inter-rater reliability → discussion → adjudication

Original reviewer ratings are preserved. Adjudicated ratings are stored separately so that agreement can be measured before reviewers discuss differences.

Error taxonomy

The project also uses a multi-label error taxonomy:

omission
addition
negation_flip
modality_shift
eligibility_shift
deadline_or_number_error
institutional_term_error
register_or_tone_shift
ambiguity
cultural_pragmatic_loss
actionability_failure

These error labels complement the four 1–5 evaluation dimensions rather than replacing them.

Planned public release

The validated public release will include, where rights and provenance permit:

stable item IDs
source provenance
source and target languages
institutional reference translations
translation-system outputs
system-condition metadata
independent reviewer scores
adjudicated scores
error annotations
rights and provenance notes
dataset version information
limitations and intended-use documentation

Intended use

This dataset is intended to support research on:

AI translation evaluation
multilingual communication
language access
institutional communication
translation error analysis
human-in-the-loop quality assurance

It may also support the development of methods for identifying translations that require human review before they are used in high-consequence communication.

Not intended for

This dataset is not intended to replace qualified human translation in legal, medical, immigration, public-benefit, eligibility, or other high-consequence settings.

It should not be used to make decisions about individuals or communities.

Limitations

The initial dataset is intentionally small and language-specific.

Amharic and Afaan Oromo do not represent all under-resourced languages, and the findings should not be generalized to all multilingual settings without further evidence.

Institutional reference translations may themselves contain limitations.

Results from a research model should not be generalized to undocumented production systems.

The first release will therefore be presented as a carefully validated research dataset, not as a universal benchmark for translation quality.

Reproducibility and methods

Methods, source-provenance documentation, reviewer materials, code, and reproducible workflows are maintained in the associated GitHub repository:

https://github.com/Endalk-Chala/translation-infrastructure-language-access

Researcher

Endalkachew H. Chala
Center for an Informed Public, University of Washington

Website: https://endalk-chala.github.io/
GitHub: https://github.com/Endalk-Chala
Hugging Face: https://huggingface.co/endalkchala
ORCID: https://orcid.org/0000-0001-6210-6706

Citation

A formal citation will be added with the first validated public release.

Downloads last month: 28

Total file size:

18 kB