Scholarlink

This paper analyses distance measurement methods in multidimensional vector space to quantify the similarity of responses between LLMs and citizens of individual countries. Four LLMs created in different countries and cultural environments (ChatGPT, QWEN, YaLM and JAIS) were analyzed, and it was assumed that the LLMs’ attitudes towards science and technology would be similar to the attitudes of the population of the countries in which the LLMs were created. This transfer of attitudes was expected because textual materials in the languages of the peoples of the countries they were created are used for training LLMs. The obtained results indicate that this transfer did occur, and given the specific topic, the influence of the attitudes of Western countries, which manage to influence the attitudes of LLMs through the use of English-language content for training, was not observed.

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

Methods for Analyzing Similarity of Attitudes and Their Application to Science and Technology Across Residents of Countries and Large Language Models

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Kristian Dokic,
  • Bojan Radisic,
  • Barbara Pisker,
  • Mirjana Radman-Funaric

摘要

This paper analyses distance measurement methods in multidimensional vector space to quantify the similarity of responses between LLMs and citizens of individual countries. Four LLMs created in different countries and cultural environments (ChatGPT, QWEN, YaLM and JAIS) were analyzed, and it was assumed that the LLMs’ attitudes towards science and technology would be similar to the attitudes of the population of the countries in which the LLMs were created. This transfer of attitudes was expected because textual materials in the languages of the peoples of the countries they were created are used for training LLMs. The obtained results indicate that this transfer did occur, and given the specific topic, the influence of the attitudes of Western countries, which manage to influence the attitudes of LLMs through the use of English-language content for training, was not observed.