AQE-RF: An Adaptive Quantifier Extension and Rule-Filtering Graph Network for Logical Reasoning of Text

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Computational Linguistics

Background:

• Neural models require strong contextual understanding for logical reasoning.
• Existing methods like neural architecture-based and LLM-driven strategies have limitations in fine-grained logic and explicit inference control.
• First-order logic approaches struggle with systematic quantifier handling and clear reasoning processes.

Purpose of the Study:

• To improve the logical reasoning capabilities of pretrained language models (PLMs).
• To address the limitations of current neural and LLM-driven approaches in handling complex logical structures.
• To develop a model that offers explicit inference control and interpretable reasoning paths.

Main Methods:

• Proposed AQE-RF model inspired by first-order logic and generalized quantifier (GQ) theory.
• Constructed a fine-grained text logical graph (FTLG) with GQ instantiation via option attention.
• Implemented rule-filtered deductive reasoning using conflict scores and dynamic programming (DP) for coherent inference path selection.

Main Results:

• AQE-RF demonstrates effectiveness in improving logical reasoning over existing methods.
• The model shows robustness across multiple benchmark datasets (LogiQA, ReClor, AR-LSAT).
• The approach successfully integrates explicit inference control and interpretable reasoning.

Conclusions:

• AQE-RF offers a significant advancement in enhancing logical reasoning for PLMs.
• The model's architecture effectively handles quantifiers and provides interpretable inference.
• This work contributes to more reliable and accurate natural language understanding systems.