An Elicitation-Matrix Approach to Pragmatic Context Modeling in Low-Resource Machine Translation

The Case of Akuapem Twi

Abstract

Pragmatic ambiguity poses a major challenge for machine translation in low-resource languages like Akan, where a single English phrase may represent multiple pragmatic contexts and vice versa. To address this gap, we develop an elicitation matrix capturing key social and situational factors and use it to create a pragmatics‑focused Akan--English dataset of 863 annotated pairs. We then evaluate whether large language models (LLMs) can infer pragmatic context and whether explicit pragmatic tags improve translation selection choices. Across two models, three prompting strategies, and three experimental settings, human‑annotated pragmatic tags consistently yield the highest accuracy, with the largest gains on expansive (many‑to‑one) mappings. Chain‑of‑thought prompting further boosts performance. These findings indicate that pragmatic conditioning---rather than model size---is the primary driver of improvement, and they suggest that future models will benefit from incorporating pragmatic information during training and inference.