[Paper Review] Position Bias Mitigation: A Knowledge-Aware Graph Model for Emotion Cause Extraction

Hanqi Yan, Lin Gui+2 more|arXiv (Cornell University)|Jun 7, 2021

Topic Modeling26 references4 citations

TL;DR

This paper proposes a knowledge-aware graph model (KAG) that mitigates position bias in Emotion Cause Extraction by modeling semantic dependencies through commonsense knowledge paths from ConceptNet. By constructing a clause graph with sequence and knowledge edges, and using an extended Relation-GCN for node representation learning, KAG achieves state-of-the-art performance on the original dataset and significantly improved robustness against adversarial examples that disrupt positional cues.

ABSTRACT

The Emotion Cause Extraction (ECE)} task aims to identify clauses which contain emotion-evoking information for a particular emotion expressed in text. We observe that a widely-used ECE dataset exhibits a bias that the majority of annotated cause clauses are either directly before their associated emotion clauses or are the emotion clauses themselves. Existing models for ECE tend to explore such relative position information and suffer from the dataset bias. To investigate the degree of reliance of existing ECE models on clause relative positions, we propose a novel strategy to generate adversarial examples in which the relative position information is no longer the indicative feature of cause clauses. We test the performance of existing models on such adversarial examples and observe a significant performance drop. To address the dataset bias, we propose a novel graph-based method to explicitly model the emotion triggering paths by leveraging the commonsense knowledge to enhance the semantic dependencies between a candidate clause and an emotion clause. Experimental results show that our proposed approach performs on par with the existing state-of-the-art methods on the original ECE dataset, and is more robust against adversarial attacks compared to existing models.

Motivation & Objective

• To investigate the extent to which existing ECE models rely on clause position relative to emotion clauses due to dataset bias.
• To generate adversarial examples that disrupt positional cues, thereby testing model robustness.
• To develop a graph-based model that leverages commonsense knowledge to enhance semantic dependency modeling between candidate clauses and emotion clauses.
• To improve generalization by reducing dependence on positional features, especially in cases where causes are not near emotion clauses.

Proposed method

• Construct a clause graph where nodes represent clauses and edges include Sequence-Edges (S-Edges) linking consecutive clauses to capture local context.
• Introduce Knowledge-Edges (K-Edges) that connect a candidate clause to the emotion clause if a knowledge path exists in ConceptNet between their key entities.
• Initialize node features using a Bi-LSTM and Transformer encoder to obtain clause representations, with document-level attention to integrate emotion context.
• Extract knowledge paths from ConceptNet between clause entities and encode them via Bi-LSTM to produce path representations.
• Extend Relation-GCN to update node representations by aggregating messages from both S-Edges and K-Edges, enabling joint modeling of sequential and knowledge-based dependencies.
• Perform clause classification via graph neural network inference to predict cause clauses, with attention mechanisms to weigh relevant paths and clauses.

Experimental results

Research questions

• RQ1 To what extent do existing ECE models depend on the relative position of clauses for cause prediction?
• RQ2 How effective is the proposed adversarial example generation strategy in disrupting position-based inductive biases?
• RQ3 Can a knowledge-aware graph model reduce reliance on positional cues while maintaining performance on standard benchmarks?
• RQ4 How does the integration of commonsense knowledge from ConceptNet improve model robustness against position-agnostic adversarial examples?

Key findings

• The proposed KAG model achieves competitive F1 scores of 48.52 (micro) and 62.39 (macro) on the original ECE dataset, matching state-of-the-art performance.
• On adversarial examples where positional cues are removed, KAG shows the smallest average F1 drop of 16.97%, outperforming other models in robustness.
• Models like PAE-DGL and RTHN suffer significant performance drops—up to 33%—on adversarial samples, indicating strong reliance on positional features.
• Case studies confirm that KAG effectively attends to multiple valid knowledge paths, such as 'arguments → fight → angry' and 'molest → irritate → exasperate → angry', to identify correct cause clauses.
• The model demonstrates attention to relevant knowledge paths, but also reveals limitations in handling spurious paths, such as 'unemployment → situation → trouble → worried', which are not grounded in event-level reasoning.
• The study confirms that dataset bias toward proximal causes leads to models overfitting to position, and that knowledge-enhanced graph modeling can mitigate this issue.