[Paper Review] DAN: Dual-View Representation Learning for Adapting Stance Classifiers to New Domains

Chang Xu, Cécile Paris+4 more|arXiv (Cornell University)|Mar 13, 2020

Topic Modeling33 references6 citations

TL;DR

This paper proposes DAN, a dual-view representation learning framework that improves stance classifier adaptation to new domains by separately modeling subjective (opinion-based) and objective (fact-based) stance expressions. By learning domain-invariant features in two distinct views and fusing them optimally, DAN achieves better alignment and stance discrimination than single-view methods, consistently outperforming state-of-the-art baselines across 20 adaptation tasks with lower proxy A-distance values.

ABSTRACT

We address the issue of having a limited number of annotations for stance classification in a new domain, by adapting out-of-domain classifiers with domain adaptation. Existing approaches often align different domains in a single, global feature space (or view), which may fail to fully capture the richness of the languages used for expressing stances, leading to reduced adaptability on stance data. In this paper, we identify two major types of stance expressions that are linguistically distinct, and we propose a tailored dual-view adaptation network (DAN) to adapt these expressions across domains. The proposed model first learns a separate view for domain transfer in each expression channel and then selects the best adapted parts of both views for optimal transfer. We find that the learned view features can be more easily aligned and more stance-discriminative in either or both views, leading to more transferable overall features after combining the views. Results from extensive experiments show that our method can enhance the state-of-the-art single-view methods in matching stance data across different domains, and that it consistently improves those methods on various adaptation tasks.

Motivation & Objective

• To address the challenge of limited annotated stance data in new domains by improving domain adaptation for stance classification.
• To identify and model two distinct linguistic types of stance expressions: subjective (opinion-laden) and objective (fact-based) utterances.
• To design a dual-view adaptation network (DAN) that learns domain-invariant, stance-discriminative features in separate views for improved transferability.
• To validate that joint learning and optimal fusion of dual-view features lead to better alignment and performance than single-view approaches.
• To demonstrate the effectiveness of DAN across diverse adaptation tasks using quantitative and qualitative analysis.

Proposed method

• DAN decomposes each input utterance into two views: one for subjective expressions (opinions) and one for objective expressions (facts), enabling separate feature learning.
• The model employs view-specific domain examiners (discriminators) to enforce domain invariance in each view via adversarial training.
• Each view learns stance-discriminative features through shared encoder networks tailored to the linguistic characteristics of its respective expression type.
• The method performs view fusion by combining the best-performing features from both views using a learnable fusion mechanism to produce a unified, transferable representation.
• All components are trained jointly using back-propagation with a multi-task loss combining stance classification and domain adversarial objectives.
• The framework is compatible with existing single-view domain adaptation methods (e.g., DANN, WDGRL), which are enhanced by integrating the dual-view learning mechanism.

Experimental results

Research questions

• RQ1 Can modeling subjective and objective stance expressions as separate views improve domain adaptation performance in stance classification?
• RQ2 Does learning domain-invariant features in two distinct linguistic views lead to better alignment between source and target domains than a single global feature space?
• RQ3 How does the fusion of dual-view features affect overall transferability and stance discrimination compared to single-view methods?
• RQ4 To what extent does the proposed dual-view adaptation reduce proxy A-distance between domains, indicating reduced domain discrepancy?
• RQ5 Can the DAN framework be effectively integrated with and improve upon existing state-of-the-art single-view domain adaptation methods for stance classification?

Key findings

• D-DANN and D-WDGRL, the DAN-enhanced versions of DANN and WDGRL, achieved consistently lower proxy A-distance values than their original counterparts across all 20 adaptation tasks, indicating improved domain alignment.
• The dual-view features learned by DAN were more domain-invariant and stance-discriminative than those from single-view methods, as visually confirmed by t-SNE projections.
• The qualitative analysis showed that the subjective and objective views captured distinct linguistic patterns, with each view contributing unique discriminative capacity to the final representation.
• The optimal fusion of dual-view features led to more transferable representations than using either view alone, demonstrating the necessity of view-specific learning and fusion.
• The proposed method improved the performance of baseline single-view adaptation methods across diverse adaptation tasks, validating its generalizability and effectiveness.
• The results confirm that fine-grained adaptation through dual views reduces domain discrepancy and enhances stance classification generalization in low-resource settings.