Abstract
Braid is a neuro-symbolic reasoning engine designed to bridge the gap between the rigid precision of formal logic and the flexible intuition of deep learning. By integrating neural networks for "fuzzy" matching and on-the-fly rule generation, Braid solves complex story-understanding tasks while providing human-verifiable, step-by-step logical explanations. It achieves state-of-the-art results on narrative benchmarks while overcoming the traditional "brittleness" of symbolic AI.
Introduction
Traditional symbolic AI provides transparent reasoning but is notoriously brittle, failing when linguistic variations occur (e.g., "put" vs. "place") or when encountering gaps in pre-programmed common-sense rules. Conversely, modern neural networks handle linguistic nuances effectively but operate as "black boxes," lacking verifiable reasoning paths. Braid addresses this "knowledge acquisition bottleneck" and the lack of interpretability by weaving neural components into a symbolic reasoning framework.
Methods
Braid's architecture centers on a symbolic reasoner enhanced by neural modules at key decision points:
- Fuzzy Unification: Neural functions (BERT-based) calculate similarity between terms in context, allowing the system to recognize logically equivalent but linguistically different terms.
- Dynamic Rule Generation (DRG): A fine-tuned T5 model generates "if-then" common-sense rules at runtime when the reasoner encounters gaps in its knowledge base, utilizing datasets like GLUCOSE.
- Distributed Proof Graph Builder: To manage the expanded search space created by fuzzy logic, a "Master" node coordinates parallel "Worker" nodes to explore reasoning paths simultaneously.
- Integer Linear Programming (ILP): Once a graph of potential explanations is built, ILP is used to select the optimal proof based on neural confidence scores and logical consistency.
Key Results and Explainability
Braid was evaluated using the ROC Story Cloze Test, a benchmark requiring systems to choose the most plausible ending for a four-sentence story.
- Performance: Braid achieved 89.88% accuracy on the ROC Story Cloze Test, outperforming contemporary models like GPT-2 (86.5%) and HintNet (79.2%).
- Explainability: Unlike pure neural models, Braid generates a "Proof Tree" that can be converted into natural language. This allows the system to justify its choice (e.g., explaining that Zoey moved a plant because plants need light and windows provide light).
- Zero-Shot Capability: On complex children's stories, Braid reached 60% accuracy without task-specific training, nearly doubling the performance of other state-of-the-art systems at the time.
Discussion and Future Work
Braid demonstrates that knowledge can be injected at runtime rather than being pre-compiled, significantly reducing the brittleness of expert systems. However, the system remains dependent on initial semantic parsing accuracy and the manual definition of narrative frames. Future work aims to automate frame discovery and reduce the computational overhead associated with large-scale ILP solving.
Summary updated