Deep learning-based automatic action extraction from structured chemical synthesis procedures.

Area of Science:

• Chemistry
• Natural Language Processing
• Machine Learning

Background:

• Chemical synthesis procedures are often described in unstructured text within patents.
• Extracting actionable information from these procedures is challenging but crucial for advancing chemical research.

Purpose of the Study:

• To develop and validate a machine learning methodology for extracting structured actions from chemical synthesis procedures in patents.
• To bridge the gap between chemistry and natural language processing (NLP) by transforming experimental procedures into a usable format.

Main Methods:

• A pipeline combining machine learning algorithms and scripts was developed to process USPTO and EPO patents.
• Key tasks included classifying patent paragraphs for chemical procedures and converting sentences into a structured format.
• Artificial neural networks, including Long Short-Term Memory (LSTM), bidirectional LSTMs, Transformers, and fine-tuned T5, were employed.

Main Results:

• The bidirectional LSTM achieved a high accuracy of 0.939 in classifying chemical procedure paragraphs.
• The Transformer model demonstrated superior performance in the second task, achieving a BLEU score of 0.951 for structured output.
• The pipeline successfully transforms unstructured synthesis procedures into a structured, actionable dataset.

Conclusions:

• The developed methodology effectively extracts structured actions from chemical synthesis procedures, creating a valuable dataset.
• This facilitates AI-driven approaches for streamlining synthetic pathways, predicting outcomes, and optimizing reaction conditions.
• The structured dataset enhances accessibility and utility of information within chemical synthesis procedures for researchers.