Source-linked AI summary
Cross-Target Stance Classification with Self-Attention Networks
Chang Xu, Cecile Paris, Surya Nepal, Ross Sparks
TL;DR
Stance classification typically requires a classifier trained for each target, creating a need to generalize across related targets. The paper proposes CrossNet, a self-attention model that extracts shared domain-specific aspects from a source target, and reports improved cross-target performance in certain domains. Its source-target choice and transfer setting remain limited.
Problem
Traditional stance classification trains a target-specific classifier, while related targets may share domain aspects that could support cross-target generalization.
Method
CrossNet uses self-attention to identify domain-specific aspects in stance-bearing sentences and transfer target-independent information from a source target to a destination target.
Results
Cross-target F1-score improves 6.6% over BiCond, compared with 3.0% in-target, while CrossNet achieves superior results over baselines in certain domains.
Takeaways & Limitations
CrossNet can benefit stance inference across related targets by capturing domain-specific information shared between them.
Takeaways & Limitations
The model uses one source target selected by common sense, and cross-target performance remains 26% lower than in-target performance on average.
Abstract
from arXiv · showhide
In stance classification, the target on which the stance is made defines the boundary of the task, and a classifier is usually trained for prediction on the same target. In this work, we explore the potential for generalizing classifiers between different targets, and propose a neural model that can apply what has been learned from a source target to a destination target. We show that our model can find useful information shared between relevant targets which improves generalization in certain scenarios.
1 Introduction
The paper studies whether stance classifiers can generalize from an annotated source target to a related destination target by using shared domain aspects. CrossNet applies self-attention to extract such information, improving classification in certain domains while source-target selection remains limited.
- Motivation: Traditional stance classification trains a target-specific classifier, requiring new annotated ground-truth data for each unseen target.Cross-target adaptation is presented as an alternative that can reuse knowledge from existing related targets.
- Motivation: Cross-target classification is harder because language models may be incompatible across targets, despite shared aspects such as environmental damage.The paper illustrates this overlap with sentences about reef destruction across climate-change and mining-project targets.
- Cross-target stance classification: CrossNet learns domain-specific aspects from a source target and applies them to prediction on a related destination target.The model is designed to transfer target-independent information across related targets.
- Limitations: The source target is selected using common sense, while more sophisticated source-target selection is left for future work.This limits how systematically the method chooses which related target should provide transferable knowledge.
2 Model
CrossNet encodes a stance-bearing sentence conditionally on its target, extracts domain-specific aspects with self-attention, and predicts the stance from that aspect representation.
- CrossNet takes a stance-bearing sentence and descriptive target as inputs and yields a predicted stance label through four layers.The architecture comprises embedding, context encoding, aspect attention, and prediction layers.
- 2.2 Context Encoding Layer: A target BiLSTM first encodes target context, then initializes a separate sentence BiLSTM to produce target-conditioned sentence representations.The forward and backward states are aligned when initializing the sentence encoder.
- 2.3 Aspect Attention Layer: The aspect attention layer learns recurring domain aspects across sentences and uses self-attention to identify sentence parts compatible with the sentence’s overall semantics.Shared layer parameters are stimulated by repeated domain-aspect mentions, while attention weights select core sentence information.
- 2.4 Prediction Layer: Attention-weighted hidden states form a domain-aspect encoding AP that the prediction layer feeds into an MLP and softmax to obtain class probabilities.The output contains predicted probabilities for each of the C stance classes.
- 2.5 Model Training: CrossNet is trained with multiclass cross-entropy loss plus L2 regularization over the trainable model parameters.The loss uses ground-truth class indicators and predicted class probabilities.
3 Experiments
Experiments evaluate CrossNet against two baselines on cross-target stance classification using SemEval and Australian mining-project tweets. Quantitative and qualitative results show stronger cross-target gains, while attention visualizations identify domain-relevant aspects and expose destination-specific vocabulary limitations.
- 3.4 Classification Performance: BiCond outperforms BiLSTM across all target configurations, indicating that conditional target encoding better captures sentence–target dependency than concatenation.
- 3.4 Classification Performance: CrossNet improves over both baselines in almost all cases, with average F1-score gains of 6.6% cross-target versus 3.0% in-target over BiCond.It is slightly worse than BiCond on LA in-target, but the difference is not statistically significant.
- 3.4 Classification Performance: Transfer ratios show a 26% average performance drop from in-target to cross-target settings, highlighting the remaining need for destination-target information.The study does not explore destination-specific information in model building.
- 3.5 Visualization of Attention: Attention heatmaps highlight domain-relevant words, such as “feminist,” “rights,” and “equality” for women’s rights and “president” and “dreams” for politics.The visualized examples are tweets correctly classified by CrossNet.
- 3.5 Visualization of Attention: Source-trained attention may miss destination-specific words such as “abortion” and “trumps” because they are rare in the source-target corpus.
- 3.5 Visualization of Attention: CrossNet transfers domain aspects from climate-change data to mining-project tweets by attending to words such as “reef,” “destroy,” “environmental,” and “disaster.”Highly attended words are also reported to include sentiment-bearing terms that contribute to stance prediction.
4 Conclusion and Future Work
CrossNet extracts target-independent, high-level domain information for cross-target stance classification and outperforms several baselines in certain domains. Future work targets source-target selection, multiple knowledge sources, and multilingual evaluation.
- CrossNet extracts target-independent information and achieves superior results over several baselines in certain domains.The model is reported to perceive high-level domain-specific information in stance-bearing sentences.
- Effective source-target selection is crucial for achieving satisfactory generalization on destination targets.The authors suggest learning correlations between target closeness and generalization performance to guide selection.
- Future extensions include transferring knowledge from multiple related sources and evaluating multilingual generalization.The current mining-project model uses one source target, while related community and economic targets may provide additional knowledge.