PathReasoner: Modeling Reasoning Path with Equivalent Extension for Logical Question Answering

TL;DR

PathReasoner reframes logical reasoning as atom-level reasoning paths to address data sparsity and perceptual gaps in logic. It combines data-side expansion via Equivalent Path Extension with a transformer-based Reasoning Path Modeling module that uses in-atom and cross-atom attention plus high-order diffusion to capture complex logical structure. The approach yields state-of-the-art results on ReClor and competitive gains on LogiQA, with strong generalization to other reasoning tasks and efficient training convergence. This work provides a scalable, interpretable paradigm for logical reasoning that can extend to multi-modal settings beyond text.

Abstract

Logical reasoning task has attracted great interest since it was proposed. Faced with such a task, current competitive models, even large language models (e.g., ChatGPT and PaLM 2), still perform badly. Previous promising LMs struggle in logical consistency modeling and logical structure perception. To this end, we model the logical reasoning task by transforming each logical sample into reasoning paths and propose an architecture \textbf{PathReasoner}. It addresses the task from the views of both data and model. To expand the diversity of the logical samples, we propose an atom extension strategy supported by equivalent logical formulas, to form new reasoning paths. From the model perspective, we design a stack of transformer-style blocks. In particular, we propose a path-attention module to joint model in-atom and cross-atom relations with the high-order diffusion strategy. Experiments show that PathReasoner achieves competitive performances on two logical reasoning benchmarks and great generalization abilities.

Figures (10)

• Figure 1: Probing tests on representative LMs (e.g., RoBERTa). (a) is about model prediction consistency. (b) is related to the perception of logical connectives. Detailed pilot experiments are shown in the Appendix.
• Figure 2: The architecture of PathReasoner. Part (a) is Equivalent Path Extension, which aims to improve the diversity of samples. Part (b) is Reasoning Path Modeling, which is designed to model logical structures.
• Figure 3: Performances with different numbers of atoms.
• Figure 4: Performances with different numbers of new samples.
• Figure 5: Training Efficiency Analysis.