PaperlayerView exact revision on arXiv

Source-linked AI summary

chatClimate: Grounding Conversational AI in Climate Science

Saeid Ashraf Vaghefi, Qian Wang, Veruska Muccione, Jingwei Ni, Mathias Kraus, Julia Bingler, Tobias Schimanski, Chiara Colesanti-Senni, Nicolas Webersinke, Christrian Huggel, Markus Leippold

arXiv:2304.05510v2cs.CL

TL;DR

LLMs can hallucinate or provide outdated answers, a concern in climate science where information should be accurate, current, and reliable. This paper builds chatClimate around IPCC AR6 and compares three chatbot scenarios; hybrid chatClimate outperformed GPT-4 and standalone chatClimate in answer accuracy.

  • Problem

    LLMs can hallucinate or provide outdated information, while climate questions require accurate, reliable, and timely answers.

    Source: §1 · p. 1, Large Language Models · p. 3

  • Method

    The authors integrate IPCC AR6 into searchable external memory and evaluate GPT-4, standalone chatClimate, and hybrid chatClimate on climate questions.

    Source: §1 · p. 2, chatClimate Pipeline · p. 3, Input Prompts and chatBots · p. 4

  • Results

    Hybrid chatClimate outperformed GPT-4 and standalone chatClimate in answer accuracy when given access to IPCC AR6.

    Source: §6 · p. 6

  • Takeaways & Limitations

    The approach may provide decision-makers and the public with trustworthy climate information and can scale to chatbots in specialized domains.

    Source: §1 · p. 2

    The authors did not fully explore chain-of-thought prompting by testing different prompts.

    Source: Limitations · p. 6

Abstract

from arXiv · show

Large Language Models (LLMs) have made significant progress in recent years, achieving remarkable results in question-answering tasks (QA). However, they still face two major challenges: hallucination and outdated information after the training phase. These challenges take center stage in critical domains like climate change, where obtaining accurate and up-to-date information from reliable sources in a limited time is essential and difficult. To overcome these barriers, one potential solution is to provide LLMs with access to external, scientifically accurate, and robust sources (long-term memory) to continuously update their knowledge and prevent the propagation of inaccurate, incorrect, or outdated information. In this study, we enhanced GPT-4 by integrating the information from the Sixth Assessment Report of the Intergovernmental (IPCC AR6), the most comprehensive, up-to-date, and reliable source in this domain. We present our conversational AI prototype, available at www.chatclimate.ai and demonstrate its ability to answer challenging questions accurately in three different QA scenarios: asking from 1) GPT-4, 2) chatClimate, and 3) hybrid chatClimate. The answers and their sources were evaluated by our team of IPCC authors, who used their expert knowledge to score the accuracy of the answers from 1 (very-low) to 5 (very-high). The evaluation showed that the hybrid chatClimate provided more accurate answers, highlighting the effectiveness of our solution. This approach can be easily scaled for chatbots in specific domains, enabling the delivery of reliable and accurate information.

1 Introduction

LLMs can hallucinate or provide outdated information, a particular concern for climate questions requiring accurate, reliable, timely answers. The paper presents chatClimate, which uses IPCC AR6 information, and evaluates its answers against GPT-4 and a hybrid setup.

1 Introduction

  • Hallucination and outdated information are major LLM issues, especially for climate questions requiring accurate, reliable, and timely information.The introduction also emphasizes authoritative references and citations for understanding climate risks and supporting mitigation strategies.Source: §1 · p. 1
  • Accurate sources and responses can support informed decision-making; governments may seek feedback on report statements or literature supporting claims.The passage links this information exchange to dialogue between government entities, scientists, national delegations, and policymakers.Source: §1 · p. 2
  • External sources, described as long-term memory, can update LLM knowledge without retraining and help prevent incorrect or outdated information from propagating.The passage notes that model training is computationally expensive and has other downsides.Source: §1 · p. 2
  • The authors introduce chatClimate, a climate-focused conversational AI using IPCC AR6, and evaluate 13 challenging questions across GPT-4, standalone chatClimate, and hybrid chatClimate.The reports cover the climate system, impacts, and solutions related to adaptation, mitigation, and climate-resilient development.Source: §1 · p. 2
  • The prototype is not IPCC-endorsed and may hallucinate, generate harmful instructions or biased content, or provide incorrect information.The paper states these risks in a disclaimer.Source: §1 · p. 2
  • Hybrid chatClimate outperformed GPT-4 and standalone chatClimate in answer accuracy when given access to IPCC AR6.The authors attribute the higher performance to integrating up-to-date, domain-specific data.Source: §1 · p. 2

2 Background

The background reviews LLM capabilities and their hallucination and staleness risks, climate-related NLP applications, conversational QA, and retrieval-based long-term memory approaches.

Large Language Models

  • LLMs can generate text and answer questions, but hallucinations may be semantically incorrect or unsupported, and information may become outdated after training.The passage notes that most models were trained only through the end of 2021.Source: Large Language Models · pp. 2–3

NLP and Climate Change

  • Climate-related NLP applications include financial disclosure analysis, climate-and-health topic modeling, and fact-checking climate claims..Source: NLP and Climate Change · p. 3

Question Answering and chatBots

  • Climate QA systems can make scientific information more accessible and support communication among experts, policymakers, and stakeholders.The passage also describes potential support for informed decision-making and climate mitigation and adaptation strategies.Source: Question Answering and chatBots · p. 3

Long-term Memory and Agents for LLMs

  • Retrieval agents can access long-term memory, insert retrieved information into prompts, and support domain-specific answers using in-house databases.The passage describes applications in healthcare and finance as well as domain-specific chatbots.Source: Long-term Memory and Agents for LLMs · p. 3

3 Method

The method converts IPCC AR6 reports into searchable external memory, retrieves relevant passages for user questions, and supplies them to GPT-4 through prompts tailored to three chatbot scenarios.

chatClimate Pipeline

  • IPCC AR6 reports are converted from PDFs to JSON, chunked, embedded, indexed, and searched semantically to retrieve information relevant to each query.For a question, the system embeds it and uses dot-product similarity to identify the five nearest vectors, then decodes retrieved vectors into text and uses them to refine prompts.Source: chatClimate Pipeline · p. 3

Tools and External APIs

  • The toolchain converts reports to JSON, chunks text for embedding, stores vectors, and uses GPT-4's chatcompletion endpoint to answer with indexed embeddings.The paper names PDF parsr, LangChain, and OpenAI's text-embedding-ada-002 as tools in this pipeline.Source: Tools and External APIs · pp. 3–4

Input Prompts and chatBots

  • The three prompts distinguish hybrid chatClimate, IPCC-only chatClimate, and GPT-4 without added information or answer instructions.Hybrid chatClimate receives five retrieved passages and may use both IPCC information and in-house knowledge; chatClimate is instructed to answer only from IPCC AR6.Source: Input Prompts and chatBots · p. 4

4 Experiments

The experiments compare three chatbot scenarios on 13 questions, with IPCC AR6 authors assessing answer accuracy. An example contrasts their answers to a question about limiting warming to 1.5°C.

chatBots and Questions

  • IPCC AR6 authors assessed answers from hybrid chatClimate, chatClimate, and GPT-4 across 13 questions.The study notes that source attribution can facilitate the trickle-back process required by governments and other stakeholders working with IPCC reports.Source: chatBots and Questions · p. 4
  • For the 1.5°C question, hybrid chatClimate and chatClimate state emissions reductions and timing, while GPT-4 gives a more general answer.The passage identifies this as a mitigation question and points to Table 2 for returned answers.Source: chatBots and Questions · p. 4

Evaluation of Answers (accuracy score)

  • The evaluation examines chatbot accuracy, not other interaction factors such as clarity, tone, speed, consistency, or user satisfaction.The paper explicitly limits this work's evaluation to accuracy.Source: Evaluation of Answers (accuracy score) · p. 4

5 Discussion

The discussion reports that hybrid chatClimate answers were more accurate than standalone chatClimate and GPT-4, and examines retrieval settings, risks, limitations, and future work.

Expert Cross-Check of the Answers

  • Hybrid chatClimate answers were more accurate than standalone chatClimate and GPT-4; the systems gave consistent 1.5°C timing estimates of 2030–2052 for Question 2.For Question 1, hybrid chatClimate and chatClimate discussed CO2 reductions, while GPT-4 gave a more general response; Figures 2 and 3 cross-check references for Questions 1 and 2.Source: Expert Cross-Check of the Answers · pp. 4–5, Figure 2 · p. 5, Figure 3 · p. 5

Prompts Engineering

  • Retrieving 10 or 15 nearest vectors added external information, and the K-15 scenario provided more information about the regions most affected by climate change.The authors use this comparison to highlight the importance of prompt engineering.Source: Prompts Engineering · p. 5

Personalized-GPTs or GPT-n, Risk Management

  • The authors caution that external data sources may contain inaccuracies or biases and call for regulation of database integration and prompt engineering in chatbots.They also argue that regulation matters because users may lack the knowledge to verify answers or references.Source: Personalized-GPTs or GPT-n, Risk… · pp. 5–6

Limitations

  • The study did not fully explore chain-of-thought prompting by testing different prompts.The authors expect it may improve accuracy and list it as future work.Source: Limitations · p. 6

Future works

  • The authors invite community feedback through the chatClimate website to improve its question-answering capabilities.They state a goal of providing accurate and reliable climate-change information.Source: Future works · p. 6

6 Conclusion

The conclusion reports that hybrid chatClimate outperformed GPT-4 and standalone chatClimate in answer accuracy when given access to IPCC AR6. It highlights external knowledge, retrieval, and prompt design as central findings.

6 Conclusion

  • Hybrid chatClimate outperformed GPT-4 and standalone chatClimate in answer accuracy when given access to IPCC AR6.The authors attribute the higher performance to integrating up-to-date, domain-specific data, which they say addresses hallucination and outdated information issues.Source: §6 · p. 6
  • The authors conclude that external knowledge and instructions on using it could refine hallucination and outdated-information issues, while retrieval and prompt engineering can help LLMs provide sources.They report testing K-5, K-10, and K-15 nearest knowledge pieces and call retrieval hyperparameter tuning important.Source: §6 · p. 6
  • The conclusion calls for regulation of LLM training, fine-tuning, and chatbot applications, including external database integration and prompt engineering.It notes that users may not be able to verify answers or references, making biased data or engineered prompts potentially harmful.Source: §6 · p. 6

Expert cross-check of the QA scenarios Q3-13

The section continues the QA analysis for questions 3–13, with supplementary tables comparing chatbot answers and a separate cross-check of Question 9 references across retrieval settings.

Loading 2304.05510v2…