Learning to activate logic rules for textual reasoning.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Cognitive Science

Background:

• Current textual reasoning models often lack human-like interpretability and logical accuracy.
• Existing models struggle to learn explicit reasoning processes.

Purpose of the Study:

• To develop a novel reasoning model that learns to activate logic rules explicitly.
• To improve interpretability and logical accuracy in textual reasoning.
• To mimic human cognitive activities using a specialized memory structure.

Main Methods:

• Utilized deep reinforcement learning to train a novel reasoning model.
• Employed Memory Networks with a specialized memory storing relational tuples (mimicking Image Schemas).
• Redefined textual reasoning as a sequential decision-making process involving memory modification and retrieval, with logic rules as state-transition functions.

Main Results:

• Achieved an average error rate of 0.7% on the bAbI-20 benchmark.
• Required less than 1,000 training samples.
• Demonstrated the ability to jointly address variable binding and relation activating problems in logic rule activation.

Conclusions:

• The proposed model offers a significant step towards human-like reasoning in AI.
• Explicitly activating logic rules via deep reinforcement learning enhances model interpretability and accuracy.
• The model shows high efficiency, achieving strong performance with limited training data and no supporting facts.