Technical Summary: Induce, Align, Predict: Zero-Shot Stance Detection via Cognitive Inductive Reasoning

Problem Definition

Zero-shot stance detection (ZSSD) aims to determine the stance of a text toward targets that were not seen during training. This task is critical for analyzing dynamic, polarized online discourse where new topics emerge rapidly and labeled data is scarce. While Large Language Models (LLMs) offer zero-shot capabilities, prompting-based approaches often fail to handle complex reasoning and lack robust generalization to novel targets. Conversely, LLM-enhanced methods (LEM) that combine LLM knowledge with supervised neural models still rely heavily on substantial labeled data and struggle to generalize beyond instance-level lexical patterns, limiting their interpretability and adaptability.

Methodology: The Cognitive Inductive Reasoning Framework (CIRF)

The authors propose CIRF, a schema-driven framework that bridges linguistic inputs and abstract reasoning through automatic schema induction and schema-guided inference. The framework operates via two primary components:

1. Unsupervised Schema Induction (USI)

This module leverages LLMs to abstract structured reasoning patterns from raw text without requiring labeled data. The process involves four stages:

• FOL Reasoning Generation: An LLM generates a reasoning chain for each sentence-target pair in the form of First-Order Logic (FOL).
• FOL Interpretation and Abstraction: The LLM analyzes the internal structure of the FOL chains, generating logically equivalent but structurally diverse expressions. It then summarizes these variants into generalized FOL templates that capture essential logic independent of specific targets or vocabulary (e.g., abstracting "is robot" and "helps humans" into generic "is target" and "meets condition").
• Schema Clustering and Hierarchical Abstraction: The LLM clusters these abstracted templates into logical categories (schemas) based on semantic and inference pattern similarity. For large clusters, a hierarchical strategy is used to synthesize intermediate logic chains before forming the final schema-level template.
• Schema Graph Construction: Induced schemas are represented as nodes in a multi-relational graph, with edges indicating logical relations such as causality, contrast, or implication.

2. Schema-Enhanced Graph Kernel Model (SEGKM)

This component performs the actual stance prediction by aligning input instances with the induced schemas:

• FOL Graph Construction: Input arguments are converted into FOL graphs where nodes represent predicates and edges encode logical relations (implication, conjunction, negation).
• Schema Knowledge as Graph Filters: The induced schema graphs are decomposed into subgraph filters centered on schema nodes. These filters serve as the basis for the graph kernel.
• Kernel-Based Node Representation: A learnable graph kernel aligns the local subgraph of each input node with the schema filters. This involves a semantic-aware deep kernel function (inspired by random walk kernels) that computes similarity between the input subgraph and schema filters, capturing higher-order structural alignment.
• Hierarchical Representation: Multiple layers of this kernel-based feature extraction are stacked to capture both fine-grained local logic and higher-order composite reasoning motifs. The final graph representation aggregates features from all layers for stance prediction.

Key Contributions

1. CIRF Framework: The proposal of the first schema-driven approach for ZSSD that bridges linguistic input and abstract reasoning via automatic induction and application of cognitive reasoning schemas.
2. Unified Architecture: The introduction of a unified framework combining unsupervised schema induction (abstracting FOL patterns into multi-relational schema graphs) and the SEGKM for effective schema-guided inference and generalization to unseen targets.
3. Performance and Efficiency: Extensive experiments demonstrating that CIRF establishes new state-of-the-art (SOTA) results on SemEval-2016, VAST, and COVID-19-Stance benchmarks. Crucially, it achieves competitive performance using only 30% of the labeled data required by LLM-enhanced methods, highlighting its efficiency in low-resource settings.

Experimental Results

The authors evaluated CIRF on three benchmarks:

• SemEval-2016 (SEM16): CIRF achieved an average F1 score of 76.2, outperforming the previous SOTA (FOLAR) by 1.9 points.
• VAST: CIRF achieved an average F1 of 80.9, surpassing FOLAR by 1.9 points and LCDA by 0.6 points. It showed particular robustness on this dataset due to its diverse topics.
• COVID-19-Stance: CIRF outperformed the most competitive method (LogiMDF) by 3.7 points.

Low-Resource Performance:
In experiments with reduced labeled data (e.g., 10% or 30% of the training set), CIRF consistently outperformed strong baselines (FOLAR, LCDA, LogiMDF). For instance, on the COVID-19 dataset with only 10% supervision, CIRF outperformed LogiMDF by 2.8 points.

Ablation and Sensitivity:

• Removing the cognitive schema or the SEGKM module resulted in the most significant performance drops, confirming the necessity of explicit schema-guided reasoning.
• The model demonstrated robustness to the number of induced schemas and the number of selected schema graphs, with performance remaining stable across a wide range of hyperparameters.
• Scaling CIRF with more powerful LLM backbones (GPT-4o, DeepSeek-v3) further improved performance, suggesting the framework effectively capitalizes on enhanced reasoning capabilities.

Significance and Claims

The paper claims that CIRF represents a new paradigm for schema-driven zero-shot stance detection. By treating entire FOL reasoning chains as units of schema induction, the framework successfully abstracts and transfers not just individual semantic items but the relational patterns binding them. This allows for robust generalization to novel and diverse targets by combining the interpretability of symbolic schemas with the adaptability of neural inference.

The authors assert that their work addresses the critical bottleneck of data dependency in current ZSSD approaches. By leveraging unsupervised abstraction of first-order logic patterns, CIRF enables robust generalization and interpretability across diverse targets without relying on large amounts of labeled data. The paper concludes that explicit schema-guided abstraction is essential for robust zero-shot generalization, a capability that conventional neural architectures and standard prompting strategies struggle to achieve. Future work is directed toward scaling schema induction for larger corpora and integrating richer world knowledge or multimodal signals.