Scholarlink

The adoption of financial technology (fintech) among Ugandan women is significantly hampered by language barrier and limited financial literacy, as many fintech applications are available only in English. To address this gap, this paper demonstrates the applicability of explainable machine translation (MT) models tailored for translating financial content from English to Luganda. We leveraged explainable AI techniques like BertViz, SHAP and attention matrix heatmaps for understanding MT decisions, thereby fostering user trust. We curated a high-quality dataset of financial content extracted from social media posts, fintech websites and blogs to train and fine-tune the models. We fine-tuned four transformer models (Helsinki-NLP/opus-mt-en-lg, facebook/nllb-200-distilled-600M, flores101_mm100_175M, and facebook/m2m100_418M) then evaluated them using the Bilingual Evaluation Score (BLEU). After hyperparameter optimization, the NLLB-200 model outperformed the other models with a 24.5% BLEU score. Fintech service providers can leverage our work to develop multilingual interfaces, making financial services more accessible to Luganda-speaking women and men. By enabling users to better understand and interact with financial tools, our work supports informed financial decision-making and promotes greater financial inclusion. This work also provides a benchmark for developing responsible software systems and trustworthy digital fintech spaces.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Explainable English-Luganda Machine Translation Models for Building Inclusive Fintech Applications

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Belinda Marion Kobusingye,
  • Margaret Nagwovuma,
  • Barbara Nansamba,
  • Ggaliwango Marvin

摘要

The adoption of financial technology (fintech) among Ugandan women is significantly hampered by language barrier and limited financial literacy, as many fintech applications are available only in English. To address this gap, this paper demonstrates the applicability of explainable machine translation (MT) models tailored for translating financial content from English to Luganda. We leveraged explainable AI techniques like BertViz, SHAP and attention matrix heatmaps for understanding MT decisions, thereby fostering user trust. We curated a high-quality dataset of financial content extracted from social media posts, fintech websites and blogs to train and fine-tune the models. We fine-tuned four transformer models (Helsinki-NLP/opus-mt-en-lg, facebook/nllb-200-distilled-600M, flores101_mm100_175M, and facebook/m2m100_418M) then evaluated them using the Bilingual Evaluation Score (BLEU). After hyperparameter optimization, the NLLB-200 model outperformed the other models with a 24.5% BLEU score. Fintech service providers can leverage our work to develop multilingual interfaces, making financial services more accessible to Luganda-speaking women and men. By enabling users to better understand and interact with financial tools, our work supports informed financial decision-making and promotes greater financial inclusion. This work also provides a benchmark for developing responsible software systems and trustworthy digital fintech spaces.