Collaborative Knowledge Infusion for Low-resource Stance Detection

TL;DR

This work tackles low-resource stance detection by jointly infusing target-related background knowledge from multiple sources and learning efficiently with a parameter-efficient adaptor. A knowledge verifier aligns semantic knowledge from Wikipedia and the Internet, mitigating wrong or irrelevant infusions, while a collaborative adaptor preserves backbone knowledge and enables task-specific learning with few trainable parameters. A staged optimization strategy with label smoothing and weighted loss stabilizes training under data imbalance. Across VAST, PStance, and COVID-19-Stance, the approach achieves state-of-the-art results, demonstrating strong data efficiency and cross-target generalization, with practical implications for robust stance analysis in low-resource domains.

Abstract

Stance detection is the view towards a specific target by a given context (\textit{e.g.} tweets, commercial reviews). Target-related knowledge is often needed to assist stance detection models in understanding the target well and making detection correctly. However, prevailing works for knowledge-infused stance detection predominantly incorporate target knowledge from a singular source that lacks knowledge verification in limited domain knowledge. The low-resource training data further increases the challenge for the data-driven large models in this task. To address those challenges, we propose a collaborative knowledge infusion approach for low-resource stance detection tasks, employing a combination of aligned knowledge enhancement and efficient parameter learning techniques. Specifically, our stance detection approach leverages target background knowledge collaboratively from different knowledge sources with the help of knowledge alignment. Additionally, we also introduce the parameter-efficient collaborative adaptor with a staged optimization algorithm, which collaboratively addresses the challenges associated with low-resource stance detection tasks from both network structure and learning perspectives. To assess the effectiveness of our method, we conduct extensive experiments on three public stance detection datasets, including low-resource and cross-target settings. The results demonstrate significant performance improvements compared to the existing stance detection approaches.

Figures (3)

• Figure 1: Overview of our stance detection architecture: knowledge alignment, parameter-efficient learning, and staged optimization. Knowledge alignment collaboratively selects semantic similar knowledge of the target from different knowledge sources. Parameter-efficient learning introduces the collaborative adaptor and knowledge augmentation into the stance detection model to perform low-resource learning. Staged optimization algorithm optimizes the classifier with label smoothing, then pushes the classifier edge aligning to data distribution with the help of weighted loss.
• Figure 2: Knowledge alignment. The target's collaborative knowledge is the knowledge with a higher semantic similarity score from Wikipedia or the Internet. The target's Wikipedia knowledge is obtained by Wikipedia's API. The target's knowledge from the Internet is obtained by Google retrieval.
• Figure 3: Overview of collaborative adaptor in efficient-parameter learning. $Q, K, V$ are the query, key, and value of the Transformer module respectively. LoRA denotes the low-rank adaptor, and the gate is a controller for LoRA.