Scholarlink

Large Language Models (LLMs) have profoundly influenced the research in Natural Language Processing (NLP). One of their most popular applications is in generative AI by the development of Generative Pre-trained Transformer (GPT) models. The open-source availability of state-of-the-art generative AI models has improved learning experiences through the development of assistive tools for code completion, paraphrasing, and content enhancements. However, most of these tools have been designed for the English language and do not cater to users whose first language is a low-resourced Indian language. LLMs have also highlighted a remarkable potential in neural machine translation and transliteration with the development of open-source models such as IndicTrans and IndicXlit. In this paper, we have built upon the existing works for the English language by developing a working prototype of an assistive writing tool, Bahubhaashee, for three low-resource Indian languages specifically Hindi, Marathi, and Tamil. This tool also computes the coauthorship metrics for the entire writing session to quantify the usage of AI in the written content. We evaluated this tool by conducting a user study with 7 students who had proficiency in at most two languages apart from English. It was deduced from the coauthorship metrics that the users write most of the text independently (RB = 0.625) and did not directly accept any text generated by GPT-2 as it is (RA = 0).

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

A Novel Multilingual Human-AI Collaborative Writing Design for Indian Languages

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Gaurav Pendharkar,
  • Shibani Antonette,
  • Ratnavel Rajalakshmi

摘要

Large Language Models (LLMs) have profoundly influenced the research in Natural Language Processing (NLP). One of their most popular applications is in generative AI by the development of Generative Pre-trained Transformer (GPT) models. The open-source availability of state-of-the-art generative AI models has improved learning experiences through the development of assistive tools for code completion, paraphrasing, and content enhancements. However, most of these tools have been designed for the English language and do not cater to users whose first language is a low-resourced Indian language. LLMs have also highlighted a remarkable potential in neural machine translation and transliteration with the development of open-source models such as IndicTrans and IndicXlit. In this paper, we have built upon the existing works for the English language by developing a working prototype of an assistive writing tool, Bahubhaashee, for three low-resource Indian languages specifically Hindi, Marathi, and Tamil. This tool also computes the coauthorship metrics for the entire writing session to quantify the usage of AI in the written content. We evaluated this tool by conducting a user study with 7 students who had proficiency in at most two languages apart from English. It was deduced from the coauthorship metrics that the users write most of the text independently (RB = 0.625) and did not directly accept any text generated by GPT-2 as it is (RA = 0).