Abstract
In the sensitive domain of human rights, automated classification systems must be both accurate and transparent. This paper presents a framework based on Relational Tsetlin Machines (RTMs) designed to classify human rights violations from unstructured text. Unlike traditional "black-box" neural networks, the RTM utilizes logic-based learning to distill complex natural language into human-readable Horn Clauses. By employing Herbrand semantics and entity generalization, the model achieves a mean accuracy of 94.71%, outperforming standard LSTM and CNN architectures by nearly 20%. This approach provides intrinsic explainability, allowing legal professionals to verify the underlying logic of the model's decisions.
Introduction
Automated monitoring of human rights violations is critical for NGOs and government agencies. However, traditional deep learning models, while powerful, lack the transparency required for legal and advocacy work. This paper addresses the need for explainable AI (XAI) in the human rights domain by proposing a relational reasoning framework that moves beyond simple classification to logical rule extraction.
The Problem of Opaque Models
Deep learning models like CNNs and LSTMs are notoriously difficult to interpret. In human rights contexts, identifying a "serious violation" is insufficient without an explanation that a human lawyer can verify. Furthermore, natural language reports are often semi-structured and require models that can generalize across different entities (e.g., recognizing that "Security Forces" and "Police" may play similar roles in a violation).
Methodology
The core of the proposed framework is the Relational Tsetlin Machine (RTM). The RTM replaces standard backpropagation with a team of automata that learn conjunctive clauses.
- Relational Mapping: Using Herbrand semantics, the system maps natural language relations (such as Support or Suppress) into a propositional space.
- Entity Generalization: To improve generalization, specific entities are replaced with variables. This allows the model to learn the structural pattern of a violation rather than memorizing specific names.
- Pipeline: The process involves input sentence processing, relation and entity extraction, generalization, and finally, feature vector generation for the RTM.
Results and Performance
The RTM was evaluated against data from the PULSAR system, which parses reports from organizations like Amnesty International. The RTM achieved a mean accuracy of 94.71%, significantly higher than LSTMs (73.9%) and CNNs (74.77%). Notably, a standard Tsetlin Machine without relational logic only reached 68.22%, highlighting the importance of the relational component in handling natural language structures.
Explainability through Horn Clauses
The final output of the RTM is a set of Horn Clauses. These are logical rules in the form "If X and Y are true, then Z is a violation." This provides intrinsic explainability, as the model's decision-making logic is directly accessible to the user, leaving no room for the textual ambiguity common in post-hoc explanation methods.
Conclusion
The RTM framework offers a robust solution for explainable relation classification in the human rights domain. By combining high accuracy with logical transparency, it provides a tool that can be trusted by legal experts and human rights advocates. Future work will focus on handling more complex, multi-clause sentences and exploring recursive logical structures to scale the system to larger, more diverse datasets.
Summary updated