Research

Low-resource MT, multilingual NLP, and model interpretability

With a focus on the languages and communities that current language technology serves worst. Full list also available on Google Scholar.

Publications

  1. 01

    AmanaMT: Translation Direction Predicts Automatic Metric Reliability

    Sole author

    Benchmarked 10 MT evaluation metrics across 730,748 segments and 33 languages. Translation direction is the strongest predictor of metric–human agreement (η² = 0.31), far ahead of morphological type (η² = 0.07). Also identifies a Simpson's Paradox in multi-year evaluation.

    PaperCode
    WMT 2026 @ EMNLP
  2. 02

    Bringing Mapudungun into the Modern MT Ecosystem

    First author

    Fine-tuned NLLB-200 on Mapudungun–Spanish across eight tokenization strategies. A novel Morfessor-VC segmentation matches a 5× larger model on arn→es (43.2 vs. 42.9 chrF++).

    PaperCode
    AmericasNLP 2026 @ ACL

Working Papers

  1. 01

    Drawing Legs on a Snake: Handling Out-of-Vocabulary Characters in Pretrained Chinese MT

    First author

    Found that pretrained MT models can't represent many Chinese characters, and showed that a new embedding per missing character fixes names and ordinary words while radical-based decomposition never helps; proposed a dictionary-plus-pronunciation initialization for characters with no training data.

    Code
    In preparation
  2. 02

    Span F1 Misreads Cross-Lingual Retrieval

    First author

    Multilingual needle-in-a-haystack tests score correct English answers as wrong. This artifact explains most of the non-Latin-script gap. It also manufactures cross-script transfer from fine-tuning and reshuffles MLQA rankings across 33 models in lm-evaluation-harness. We propose script-aware scoring as a drop-in fix.

    Code
    In preparation
  3. 03

    Lost in Translation or Lost in the Model?

    First author

    Do multilingual LLMs flatten culturally untranslatable words, like Portuguese saudade, toward their English glosses? A matched-pair study of 105 untranslatable concepts across 28 languages and 7 models finds they don't. Untranslatables stay consistently farther from English than exact translations through the models' middle layers, and this holds after removing language-identity effects.

    Code
    In preparation
  4. 04

    Donor or Seed? Separating Donor Effects from Training Noise in Low-Resource Translation

    Co-author

    Shows that comparing donor languages for low-resource NLLB fine-tuning on single runs mostly measures seed noise. A donor × seed variance decomposition across six targets finds donor effects are real but small, and similarity proxies fail to predict them.

    Code
    In preparation
  5. 05

    Speech Registers for Multimodal MT

    Co-author

    Applies source-token registration (Qu et al. 2025) to improve zero-shot performance across all four {S,T}×{S,T} MT directions; benchmarks NLLB-200 (3.3B), fine-tuned NLLB (600M), and Whisper across 420 directions and 8 language pairs.

    Code
    In preparation