PragMap-MT: A Bidirectional Elicited Dataset for Pragmatic Ambiguity in Low-Resource MT

Dataset Summary

This dataset accompanies the work:

An Elicitation-Matrix Approach to Pragmatic Context Modeling in Low-Resource Machine Translation: The Case of Akuapem Twi.

PragMap-MT is a pragmatics-focused, bidirectional Akuapem Twi-English translation dataset designed for context-sensitive translation research in low-resource settings.

It was created to address a core gap in MT evaluation and modeling: lexical correctness alone is often insufficient when multiple translations are pragmatically valid, and translation quality depends on social and communicative context.

Why Pragmatics Matters in Translation

Machine translation often performs well on lexical meaning, but pragmatic meaning is frequently under-specified, especially in low-resource languages. In Akuapem Twi-English translation, the same surface sentence can license multiple valid translations depending on social context, communicative intent, and participant relations.

Without explicit pragmatic modeling, systems can produce fluent but contextually inappropriate outputs.

This dataset addresses that gap by explicitly annotating context dimensions that influence translation choice.

The Two Sides of the Story

This release captures two complementary pragmatic phenomena:

1. Expansive mapping (1->M): Akuapem Twi -> English A single Akuapem Twi sentence can map to multiple valid English renderings depending on context.

2. Compressive mapping (M->1): English -> Akuapem Twi Multiple English formulations can converge into one Akuapem Twi rendering, while preserving communicative intent.

Together, these two directions model both ambiguity expansion and pragmatic compression, which are central to real-world translation behavior.

Included Files

• one_to_many_akan_eng_mappings_with_tags.json (1->M)
• many_to_one_akan_eng_mappings_with_tags.json (M->1)

Data Source and Partitioning

We construct a pragmatics-focused Akan-English dataset by partitioning a 25k-sentence English-Akuapem Twi corpus and designing an elicitation matrix to systematically elicit pragmatic variation.

From this corpus, we identify sentence pairs exhibiting expansive (1->M) or compressive (M->1) mappings, yielding 863 examples with clear pragmatic ambiguity, which are annotated to form the Annotated Pragmatic Set. The remaining ~23k pairs are retained as an unlabeled background set for evaluation.

We focus on this subset because it contains dense pragmatic variation; in low-resource pragmatics-focused MT, data quality is often more useful than simply scaling random examples.

Elicitation Matrix Methodology

The elicitation matrix organizes pragmatic context into seven dimensions:

• Audience Size
• Status
• Age
• Formality
• Gender
• Animacy
• Speech Act

For each sentence, values across these dimensions are used to identify translation behavior when multiple outputs are plausible.

Data Format

High-level JSON structure:

{
  "<source_sentence>": {
    "<candidate_translation_1>": {
      "AUDIENCE": "...",
      "STATUS": "...",
      "AGE": "...",
      "FORMALITY": "...",
      "GENDER": "...",
      "ANIMACY": "...",
      "SPEECH_ACT": "..."
    },
    "<candidate_translation_2>": { ... }
  }
}

Intended Uses

• Pragmatics-aware machine translation
• Context-aware reranking and candidate selection
• Evaluation of LLM pragmatic inference
• Low-resource sociolinguistic MT analysis

Limitations

• Pragmatic labels are structured abstractions and do not exhaust all discourse variables.
• Some examples remain genuinely ambiguous even after annotation.
• Distributional properties may shift as the dataset is expanded.

Growth and Versioning

This is a living dataset. We are continually increasing coverage, annotation density, and quality checks. The number of examples will increase in future releases.

Source Attribution (Required)

This dataset is derived from a subset of the following corpus. Please cite it when using this dataset:

@misc{azunre2021englishtwiparallelcorpusmachine,
      title={English-Twi Parallel Corpus for Machine Translation},
      author={Paul Azunre and Salomey Osei and Salomey Addo and Lawrence Asamoah Adu-Gyamfi and Stephen Moore and Bernard Adabankah and Bernard Opoku and Clara Asare-Nyarko and Samuel Nyarko and Cynthia Amoaba and Esther Dansoa Appiah and Felix Akwerh and Richard Nii Lante Lawson and Joel Budu and Emmanuel Debrah and Nana Boateng and Wisdom Ofori and Edwin Buabeng-Munkoh and Franklin Adjei and Isaac Kojo Essel Ampomah and Joseph Otoo and Reindorf Borkor and Standylove Birago Mensah and Lucien Mensah and Mark Amoako Marcel and Anokye Acheampong Amponsah and James Ben Hayfron-Acquah},
      year={2021},
      eprint={2103.15625},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2103.15625}
}

Citation

If you use this dataset, please cite both:

1. The source English-Twi corpus (Azunre et al., 2021)
2. This elicitation-matrix pragmatics dataset paper

@misc{pragmap_mt_2026,
  title={An Elicitation-Matrix Approach to Pragmatic Context Modeling in Low-Resource Machine Translation: The Case of Akuapem Twi},
  author={Kweku Yamoah and Godfred Agyapong and Kevin Scroggins and Neel Parekh and Detravious Brinkley and Chathuri Jayaweera and Shlok Gilda and Bonnie Dorr and Emmanuel Dorley},
  year={2026},
  note={Submitted to ANLP 2026},
  howpublished={ANLP 2026, FLAIRS},
  url={https://sites.google.com/view/flairs-anlp/anlp-2026}
}

Planned Venue

The associated paper is planned for ANLP 2026 (FLAIRS track/workshop):

• https://sites.google.com/view/flairs-anlp/anlp-2026

Licensing and Access

• This dataset is derived from a subset of the Azunre et al. English-Twi corpus.
• Users should comply with the original source dataset terms and citation requirements.
• Derivative releases should preserve attribution to both the source corpus and this annotation work.

Contact

For questions, corrections, or collaboration regarding new releases, please open an issue in the project repository or contact the dataset maintainers.