Abstract
The ability to conduct logical reasoning is a fundamental aspect of intelligent human behavior, and thus an important problem along the way to human-level artificial intelligence. Traditionally, logic-based symbolic methods from the field of knowledge representation and reasoning have been used to equip agents with capabilities that resemble human logical reasoning qualities. More recently, however, there has been an increasing interest in using machine learning rather than logic-based symbolic formalisms to tackle these tasks. In this paper, we employ state-of-the-art methods for training deep neural networks to devise a novel model that is able to learn how to effectively perform logical reasoning in the form of basic ontology reasoning.
Introduction
In the field of Artificial Intelligence, a long-standing divide exists between Knowledge Representation and Reasoning (KRR) and Machine Learning (ML). Traditional KRR relies on symbolic logic and mathematical proof theory. While precise, these methods are often "brittle," struggling with incomplete, noisy, or conflicting data, and frequently failing to scale. Conversely, Deep Learning is highly scalable and robust to noise but typically treats reasoning as a simple prediction task lacking a formal logical foundation.
This research introduces the Recursive Reasoning Network (RRN), a novel deep learning architecture designed to bridge this gap. The RRN performs ontology reasoning—automatically inferring new information from a set of rules and facts—combining the theoretical desirability of logic-based methods with the practical flexibility of deep learning.
The Recursive Reasoning Network (RRN)
Unlike standard neural networks that process input in a single pass, the RRN is designed to "sift" through data multiple times to perform multi-hop reasoning. The process occurs in two distinct stages:
- Embedding Generation (Reasoning Phase): The model initializes random vector representations (embeddings) for every individual in the dataset. It then iteratively updates these embeddings by processing all known triples (facts). These updates are designed to encode both the facts and potential inferences. Through repetition, information "transpires" between embeddings, allowing the model to connect dots across multiple logical steps.
- Query Answering (Prediction Phase): Once the embeddings are saturated with knowledge, simple Multi-Layer Perceptrons (MLPs) answer specific queries. For instance, to determine if one individual is the "great aunt" of another, their embeddings are fed into a trained MLP for a final prediction.
Experimental Evaluation
The RRN was evaluated on five diverse datasets, ranging from synthetic benchmarks to massive real-world knowledge bases:
- Family Trees: A dataset requiring complex multi-hop inferences (e.g., identifying a "first cousin once removed").
- Countries: Focused on adjacency and hierarchical locations (regions and subregions).
- DBpedia: A large-scale dataset extracted from Wikipedia, utilizing the 101 most frequent classes and 518 relation types.
- Claros: A formal catalog of archaeological artifacts.
- UMLS Reasoning: A biomedical ontology describing medical domain concepts.
Key Results and Insights
The RRN demonstrated exceptional performance, achieving near-perfect accuracy across most benchmarks. A core insight of this work is that logical inference can be treated as a state-update problem; by recursively updating embeddings, the logic of the ontology is baked directly into the vector space.
Key findings include:
- High Precision: On datasets like Family Trees and Claros, the model achieved over 99% accuracy for both facts and inferred relations.
- Scalability: Even on the complex DBpedia dataset, the model maintained an accuracy of 98.9% for relations.
- Task Variance: The model generally performed better at predicting class memberships than relations, as class inferences typically depend on fewer logical steps.
- Robustness: Unlike symbolic logic, the RRN can resolve data conflicts statistically, providing sensible predictions even when faced with messy or incomplete data.
Limitations and Future Work
The primary limitation of the current RRN is explainability. While highly accurate, the model acts as a "black box," providing conclusions without human-readable justifications or proofs. Future work aims to extend the architecture to allow the model to induce rules and explanations, evolving the RRN from a predictor into a fully interpretable neural reasoner.
Summary updated