Arjun Biswas

Using a large language model to create a chatbot companion for a major climate report

22-Apr-24

Large language models (LLMs) are text models trained on a large corpus of information and capable of generating natural language responses to user queries . One application of LLMs is to create virtual assistants or chatbots that can provide information about a particular domain, often limiting the LLM to work with a particular corpus or database in a 'retrieval augmented generation' approach. Here we describe the creation and evaluation of a chatbot designed to act as a companion or guide to a new report about tipping points in the climate system. The Global Tipping Points Report is a recent and comprehensive review of scientific understanding around abrupt changes (so-called tipping points") in the climate system, including both physical climate and societal shifts. The report was presented at COP28 and incorporated contributions from more than 200 researchers over 90 organizations. We created a chatbot as an additional channel for readers to engage with the large and complex report. The chatbot was intended to be capable of answering questions about report content without going beyond the scope of the report, misrepresenting any content or introducing falsifications. To increase trust, the chatbot was also designed to clearly reference the sections of the report used to generate each answer. Since this was a novel communications mechanism for the report-writing team, an additional goal was to evaluate the chatbot's effectiveness as a way of improving the delivery of scientific information to a general audience."