Leveraging Text-to-Text Pretrained Language Models for Question Answering in Chemistry.

Area of Science:

• Chemistry
• Computer Science
• Artificial Intelligence

Background:

• Existing chemistry question answering (QA) systems often struggle with complex queries and dynamic knowledge bases.
• Previous iterations relied on knowledge graph embeddings, limiting flexibility and requiring retraining for data evolution.

Purpose of the Study:

• To develop an advanced chemistry QA system, Marie, utilizing a text-to-text pretrained language model.
• To improve data retrieval accuracy and efficiency from a dynamic knowledge graph, "The World Avatar" (TWA).
• To overcome limitations of previous QA approaches, particularly those dependent on knowledge graph embeddings.

Main Methods:

• Leveraging a fine-tuned Flan-T5 model to translate natural language questions directly into SPARQL queries.
• Integrating the QA system with "The World Avatar" (TWA), a time-evolving knowledge graph containing chemical information.
• Eliminating the need for separate entity and relation linking components.

Main Results:

• The Marie QA system demonstrates high accuracy in retrieving data for complex queries with multiple relation hops.
• The system effectively balances query correctness with processing speed for practical applications.
• Achieved superior accuracy and question-answering capabilities compared to the predecessor system relying on knowledge graph embeddings.

Conclusions:

• The fine-tuned Flan-T5 approach offers a significant advancement in chemistry QA systems.
• Marie provides a flexible and accurate solution for querying dynamic chemical knowledge graphs without constant retraining.
• This methodology enhances the ability to handle complex chemical information retrieval tasks efficiently.