Few-shot stance detection using pre-trained word embeddings

Articles in Press

Document Type : Original Article

Authors

1 Iran, Tehran, Nasr Avenue, Tarbiat Modares University

2 Tarbiat Modares University

Abstract

Few-shot stance detection, which aims to determine whether a text expresses a supporting, opposing, or neutral position using only a small amount of labeled data, remains a major challenge in natural language processing. Traditional methods that rely on large annotated datasets often face limitations in real‑world applications. In this study, we propose SD-IWV[MR1.1], a lightweight approach that integrates pre‑trained word embeddings with part‑of‑speech (POS) tag embeddings to construct richer word representations, injecting both semantic and syntactic information into the model. In addition, the proposed method employs light fine‑tuning of pre‑trained vectors to improve generalization to unseen data, particularly in zero‑shot scenarios. Evaluations on the VAST dataset show that the proposed model achieves an F1 score of 0.732, demonstrating stable performance in handling topic ambiguity and cross‑domain adaptation, and offering notable improvements over traditional methods such as BiCond, SEKT, and TGA‑Net, as well as many BERT‑based models. Owing to its lightweight architecture, the model also requires significantly less training time compared to previous approaches

Keywords

Tabriz Journal of Electrical Engineering

Articles in Press, Corrected Proof
Available Online from 28 June 2026

Files

Share

How to cite

Statistics