Schema-Guided Hierarchical Information Extraction and Semantic Evaluation Using Generative AI presents a framework for extracting complex, nested data from unstructured documents and automatically evaluating the accuracy of that data against a gold standard. The authors developed this system to address the challenges of processing Health Technology Assessment (HTA) reports, which contain specialized, hierarchical information—such as drug indications and reimbursement recommendations—that vary in structure and length across different organizations.
The Schema-Based Framework
The core of the approach is a JSON schema that acts as a unified information model. This schema encodes domain knowledge by defining the attributes to be extracted, their hierarchical relationships, and constraints such as data types, patterns, and allowed values (enums).
By providing this schema in the prompt, the researchers guide a generative AI model to perform the extraction in a single, zero-shot API call. This method replaces manual extraction by domain experts, who previously had to interpret and record these complex attributes by hand. The schema also provides the necessary context for the evaluation phase, ensuring that the extracted output remains consistent with the required structure.
Automated Semantic Evaluation
Because the extracted data is hierarchical and of variable length, comparing it to a gold standard is not a simple task. The authors introduced a path-based matching algorithm that traverses the nested structure of the JSON objects to align corresponding elements.
When an exact string match is not found, the framework uses a second generative AI model to perform a semantic comparison. This comparison classifies results into four categories:
Exact match: Minor formatting differences (e.g., different dash types).
Semantic match: Lexically different but identical in meaning (e.g., "fever in kids" vs. "pyrexia in children").
Useful match: Information that is not identical but remains valuable for HTA research (e.g., "fever" vs. "fever in children under 12").
Non-match: No similarity found.
Performance and Efficiency
In tests using documents from the National Institute for Health and Care Excellence (NICE), the researchers extracted 12 out of 14 attributes with an F1 score exceeding 90%. The process proved significantly faster than manual extraction, requiring approximately 30 times less time than a human domain expert.
The authors report that the framework is generalizable across different generative AI models and transferable to HTA documents from different organizations and in different languages.
Considerations for Implementation
The researchers note that the effectiveness of this framework relies on the quality of the schema and the ability of the chosen generative AI model to follow instructions. While the study achieved high accuracy, the process requires a human-generated gold standard for the initial evaluation phase. Additionally, the authors emphasize that the "useful match" category is specifically tailored to the needs of HTA research, where the goal is to support reimbursement decisions rather than clinical practice, suggesting that the definition of a "successful" extraction may vary depending on the specific domain requirements.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!