Building a Trustworthy Nanoscience ChatBot

January 23, 2024

ChatBot enlarge

The domain-specific chatbot can "read" relevant documents when composing a reply. Relevance is determined using a machine-computed measure of semantic similarity.

Scientific Achievement

The Center for Functional Nanomaterials has developed a nanoscience chatbot that grounds its answers in a set of trusted documents, eliminating the tendency to fabricate facts, a behavior that renders existing chatbots unusable for science.

Significance and Impact

  • Adapting AI language systems to science domains would accelerate research.
  • Grounding responses using document retrieval makes domain-specific chatbots provide trustworthy answers.

Research Details

The Center for Functional Nanomaterials has developed a nanoscience chatbot that grounds its answers in a corpus of trusted documents. This was done by combining machine learning (ML) methods for document reading, similarity retrieval, and text generation. The overall system can be provided with a corpus of trusted scientific publications, after which it will reply to user questions by sourcing answers from the documents. This makes its replies domain-specific, grounded in trusted research, and sourced. These methods pave the way towards sophisticated AI systems for scientists.

  • Demonstrated how ML methods can be combined to enable science chatbots.
  • Demonstrated ML exploration of publication figures and data.
  • Explored the limits of domain-specific AI systems.

Publication Reference

Kevin G. Yager “Domain-specific chatbots for science using embeddings” Digital Discovery 2, 1850 (2023).

DOI: https://doi.org/10.1039/D3DD00112A
OSTI: https://www.osti.gov/biblio/2202517

Brookhaven Press Release: “Brainstorming with a Bot: CFN’s Kevin Yager develops a chatbot with an expertise in nanomaterials” 

Acknowledgment of Support

This research was carried out by the Center for Functional Nanomaterials, which is a U.S. DOE Office of Science Facility, at Brookhaven National Laboratory under Contract No. DE-SC0012704.
 

2024-21820  |  INT/EXT  |  Newsroom