Addressing Hallucinations in Generative Language Models

解决生成语言模型中的幻觉

• 批准号: 10063716
• 金额: $ 5.66万
• 依托单位: ALGOMO LIMITED
• 依托单位国家: 英国
• 项目类别: Collaborative R&D
• 财政年份: 2023
• 资助国家: 英国
• 起止时间: 2023 至 无数据
• 项目状态: 已结题
• 来源: https://gtr.ukri.org/projects?ref=10063716
• 关键词: Addressing; Hallucinations; Generative; Language; Models

Advances in Deep Learning have made Generative Language Models (GLMs) a potential game-changer for customer service, as well as other commercial applications. These models can generate natural interactions without explicit instructions or manual data, unlike current conversational Artificial Intelligence (cAI) systems.However, this technology carries the risk of incorrect or fabricated answers, known as "hallucinations" which can undermine customer trust and harm a company's reputation.In this feasibility study, we will build a consortium that will research and assess which technologies could be utilized to properly address hallucinations generated by GLMs.This project fits Algomo perfectly, as it improves upon our core product, a cAI platform, and addresses the market need for trustworthy and reliable generative cAI.By developing a technology that effectively addresses hallucinations, we will increase trust and confidence in the outputs of GLMs, ultimately accelerating their commercial adoption not only for customer service but also for a variety of other use-cases.

深度学习的进步使生成语言模型（GLM）成为客户服务以及其他商业应用的潜在游戏规则改变者。这些模型可以在没有明确指令或手动数据的情况下生成自然的交互，不像当前的对话式人工智能（cAI）系统。然而，这项技术存在错误或捏造答案的风险，即所谓的“幻觉”，这可能会破坏客户的信任并损害公司的声誉。在这项可行性研究中，我们将建立一个联盟，将研究和评估哪些技术可以用来妥善解决GLM产生的幻觉。这个项目完全适合Algomo，因为它改进了我们的核心产品，cAI平台，通过开发一种有效解决幻觉的技术，我们将提高对GLM输出的信任和信心，最终加速其商业应用，不仅用于客户服务，还用于各种其他用例。