Construction Data Knowledge Graphs for Hallucination Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI systems in construction data processing are prone to generating non-factual outputs or 'hallucinations', leading to misinformed decisions and potential catastrophic outcomes in precision-oriented and safety-related contexts.
Innovation Solution
A computer system and method that utilizes a first generator model to map unstructured construction data to structured entity relations and a second generator model to map structured entity relations to unstructured data, ensuring cycle consistency and preventing hallucinations through training with positive and negative reference data, and employing first-order logic and neural network reliability checks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a single generator model is used to process construction data, then the processing speed is faster, but the reliability of output decreases due to hallucinations
Solution Approach 1:
The system divides the data processing task into two distinct generator models: the first generator model processes unstructured construction data into structured format, and the second generator model processes structured data into unstructured output. This segmentation allows each model to specialize in specific data transformations, improving overall reliability while maintaining manageable complexity through clear functional separation.
Solution Approach 2:
Structured data serves as an intermediary representation between the two generator models. The first model transforms unstructured construction data into structured intermediate form, which then feeds into the second model that generates the final unstructured output. This intermediary structured representation acts as a verification layer that prevents hallucinations from propagating through the system.
2Reliability
If cycle consistency training is implemented with two generator models, then hallucination prevention improves, but training time and computational resources increase
Solution Approach 1:
The system implements cycle consistency training where the output of the second generator model is fed back to the first generator model, and vice versa. This creates a feedback loop that enforces consistency: if the second model processes the output of the first model and returns to the original input, the cycle should be consistent. This feedback mechanism prevents hallucinations by ensuring that generated content can be traced back to the original structured representation, though it requires additional training iterations and computational resources.
3Measurement precision
If first-order logic verification is applied to check generator outputs, then output accuracy improves, but processing speed decreases
Solution Approach 1:
The system performs first-order logic verification as a preliminary check on the outputs generated by the generator models. By verifying logical consistency before finalizing the output, the system ensures accuracy in construction data processing. This preliminary verification step, while adding processing time, prevents erroneous outputs from propagating further, ultimately improving overall system reliability and reducing rework.
Data Source
AI summary
A construction knowledge base comprises structured construction data, arranged in a knowledge graph (G) with entities and relations between the entities. A user query (X) is mapped to a set of entity relations ({circumflex over (R)}) using a first generator model (Md), trained to map unstructured construction data to structured entity relations. A subgraph (Z) is retrieved from the knowledge graph (G), using the set of entity relations for the query ({circumflex over (R)}). The subgraph (Z) is mapped to unstructured data (Y) as output for the query, using a second generator model (Mg), inverse to the first generator model (Md) and trained to map structured entity relations to unstructured construction data. The generator models (Mg, (Md)) are trained for cycle consistency, whereby structured entity relations output by the first generator model (Md) are input to the second generator model (Mg), and unstructured data output by the second generator model (Mg) is input to the first generator model (Md).