Mind Reasoning Manners: Enhancing Type Perception for Generalized Zero-Shot Logical Reasoning Over Text.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Machine Learning

Background:

• Logical reasoning tasks, often framed as multiple-choice question answering (MCQA), typically assume similar reasoning type distributions between training and testing data.
• This assumption is often violated in real-world applications, necessitating research into zero-shot capabilities and improved reasoning type perception.

Purpose of the Study:

• To address limitations in current logical reasoning benchmarks by proposing a new framework for generalized zero-shot logical reasoning.
• To develop a model that can effectively perceive and generalize across different types of logical reasoning without explicit type supervision during training.

Main Methods:

• Introduced ZsLR (Zero-shot Logical Reasoning), a benchmark with six splits based on three type sampling strategies for evaluating generalized zero-shot performance.
• Proposed TaCo (Type-aware model), which employs heuristic input reconstruction and text graph construction with a global node.
• Integrated graph reasoning and contrastive learning within TaCo to enhance the perception of reasoning types in global representations.

Main Results:

• TaCo demonstrated superiority over state-of-the-art (SOTA) methods in both zero-shot and full-data settings.
• The model's generalization capability was verified on other logical reasoning datasets.
• Experiments confirmed TaCo's effectiveness in improving type perception for logical reasoning tasks.

Conclusions:

• The proposed ZsLR benchmark and TaCo model offer significant advancements for generalized zero-shot logical reasoning.
• TaCo effectively enhances model understanding and generalization across diverse reasoning types, addressing key limitations of existing approaches.
• The findings suggest TaCo's potential for broader applications in complex text-based reasoning tasks.