Abstract
This study explores the application of task-independent memory biases—specifically recency and frequency of access—to improve the efficiency of intelligent agents navigating ambiguous information. By framing Word Sense Disambiguation (WSD) as a memory retrieval problem within the Soar cognitive architecture, the researchers demonstrate that a computationally efficient approximation of human-like memory decay enables agents to scale to large knowledge bases like WordNet while maintaining sub-millisecond query times.
Introduction
For artificial agents to achieve general intelligence, they must manage vast long-term declarative memories. However, as memory size increases, the "needle in a haystack" problem intensifies. When faced with an ambiguous cue, such as the word "bank," an agent must efficiently retrieve the most relevant sense from a multitude of possibilities. While cognitive architectures like ACT-R have utilized "base-level activation" to prioritize frequent and recent information, these models have historically struggled to scale to the hundreds of thousands of entries found in real-world datasets. This research bridges the gap between cognitively inspired memory models and the computational requirements of large-scale, real-time applications.
Methodology
The authors utilize Word Sense Disambiguation (WSD) as a proxy for general memory retrieval. The task involves identifying the correct sense of a word within a given context using the WordNet 3.0 lexicon, which contains over 212,000 senses.
Memory Biases and Models
The study implemented several retrieval strategies within the Soar cognitive architecture:
- Individual Biases: Testing "Recency" (last used sense) and "Frequency" (most used sense) in isolation.
- Base-level Activation: A standard cognitive model where memory strength decays exponentially over time.
- Threshold Model: A hybrid heuristic that prioritizes recency if the last access was within a specific window, otherwise defaulting to global frequency.
Computational Optimization
To address the scalability of base-level activation—where every memory's score typically requires updating at every time step—the authors developed a locally efficient approximation. This "lazy" update mechanism only calculates a memory's activation upon access, utilizing a history of the last 10 accesses to estimate decay.
Experimental Results
The approach was evaluated across three major corpora: SemCor, Senseval-2, and Senseval-3.
Performance Gains
The results indicate that historical access patterns are highly predictive of correct word senses. Memory-based models significantly outperformed random selection (38.73% on SemCor) and traditional NLP baselines like the Lesk algorithm.
As shown in the performance data, the base-level activation model achieved approximately 78.47% accuracy on the SemCor dataset by the 10th run. This demonstrates that the agent "learns" the regularities of its environment through repeated exposure, even without complex linguistic context analysis.
Scalability and Efficiency
A critical finding was the efficiency of the optimized retrieval system. The "lazy" update approximation maintained high fidelity to the perfect decay model while ensuring real-time responsiveness.
The system handled the entire WordNet database with a maximum query time of 1.34ms, well within the 50ms threshold required for reactive robotic control. Furthermore, the model fidelity remained above 90% across all datasets, confirming that the approximation does not significantly degrade decision quality.
Conclusion
This research demonstrates that task-independent memory biases provide a powerful, computationally efficient mechanism for handling ambiguity in large-scale knowledge bases. By leveraging the history of memory access, agents can achieve high accuracy in tasks like WSD without relying on domain-specific linguistic rules. Future work will focus on integrating environmental context through "spreading activation" and testing these biases in non-linguistic domains to further validate their general utility for intelligent agents.
Summary updated