From Large Language Model Predicates to Logic Tensor Networks: Neurosymbolic Offer Validation in Regulated Procurement
Summary: arXiv:2604.05539v1 Announce Type: new
Abstract: We present a neurosymbolic approach, i.e., combining symbolic and subsymbolic artificial intelligence, to validating offer documents in regulated public institutions. We employ a language model to extract information and then aggregate with an LTN (Logic Tensor Network) to make an auditable decision. In regulated public institutions, decisions must be made in a manner that is both factually correct and legally verifiable.
Introduction
The increasing complexity of procurement processes in regulated public institutions necessitates innovative solutions for validating offer documents. Traditional methods often fall short in terms of accuracy and efficiency. Our neurosymbolic approach bridges the gap between symbolic reasoning and subsymbolic learning, enabling a robust mechanism for offer validation.
The Neurosymbolic Approach
Our proposed model integrates two key components: large language models (LLMs) and Logic Tensor Networks (LTNs). This combination allows us to:
- Extract relevant information from unstructured text using LLMs.
- Perform logical reasoning and validation using LTNs.
- Generate auditable decisions that can be verified against established legal frameworks.
Benefits of the Proposed Model
One of the standout features of our neurosymbolic approach is its ability to link existing domain-specific knowledge with the semantic understanding of language models. This linkage provides several advantages:
- Interpretability: The decisions made by our model can be traced back to specific predicate values and rule truth values, enhancing transparency.
- Modular Predicate Extraction: The model can adapt to different domains by easily incorporating new rules and predicates.
- Support for Explainable AI (XAI): The system provides explanations for its decisions, fostering trust and compliance in regulated environments.
Experiments and Results
We conducted experiments using a real corpus of offer documents to evaluate the performance of our proposed pipeline. The results indicate that our approach achieves performance levels comparable to existing models while significantly enhancing interpretability and decision justification.
Key findings from our experiments include:
- The model successfully extracted and validated essential information from the offer documents.
- Decisions made by the model were auditable and aligned with legal requirements.
- User feedback highlighted the model’s effectiveness in providing clear explanations for its decisions.
Conclusion
In conclusion, our neurosymbolic approach to offer validation in regulated procurement offers a promising solution to existing challenges. By combining symbolic and subsymbolic AI, we provide a framework that not only meets the requirements of accuracy and legal compliance but also enhances the interpretability of AI-driven decisions. As public institutions continue to evolve, the integration of such advanced technologies will be crucial for maintaining transparency and trust in procurement processes.