Recognizing Conflict Opinions in Aspect-level Sentiment Classification with Dual Attention Networks

Xingwei Tan, Yi Cai, Changxi ZhuEMNLP 2019 · DOI

In the world of sentiment analysis, opinions are often treated as binary: positive or negative. However, human thought is rarely that simple. We frequently encounter "conflict opinions," such as "the atmosphere was attractive, but a little uncomfortable." In Natural Language Processing (NLP), standard models often struggle with these dialectic thoughts, where contradictory views are held simultaneously.

A research team from the South China University of Technology presented a solution at EMNLP-IJCNLP 2019, arguing that ignoring these nuances makes real-world sentiment systems unreliable. They proposed a novel architecture designed specifically to capture the complexity of mixed feelings.

The Challenge of Conflict Opinions

Most Aspect-level Sentiment Classification (ALSC) tasks categorize opinions into four buckets: positive, negative, neutral, or conflict. Existing models consistently fail in the "conflict" category for two primary reasons:

  1. Data Sparsity: Conflict opinions are rare in standard datasets, meaning models lack sufficient examples to learn their patterns.
  2. Complex Expressions: Conflict opinions combine positive and negative cues. Standard "single-attention" models usually fixate on one dominant part of a sentence (e.g., "attractive") while ignoring the other (e.g., "uncomfortable"), leading to misclassification.

The D-AT-GRU Architecture

To address these issues, the researchers introduced the Dual Attention-based GRU (D-AT-GRU). This model shifts the paradigm from multi-class classification to a multi-label approach. Instead of choosing one of four categories, the model asks two binary questions: "Is there positive sentiment?" and "Is there negative sentiment?" If both are present, the label is automatically identified as "conflict."

The D-AT-GRU architecture utilizes dual attention heads (αp and αn) to separately identify positive and negative sentiment features from the hidden states of a GRU backbone.

Key Components:

  • Dual Attention Mechanism: The model employs two separate attention heads. One is dedicated to hunting for positive expressions, while the other searches for negative ones.
  • Orthogonal Regularization: To prevent the two heads from focusing on the same words, a mathematical constraint forces the attention maps to be distinct. This ensures the "positive" head isn't distracted by "negative" words and vice versa.
  • GRU Backbone: Gated Recurrent Units (GRU) are used to extract contextual features, providing a robust foundation for the attention layers.

Experimental Results

The D-AT-GRU model was evaluated on the SemEval 2014 dataset. While overall accuracy improved, the most significant gains were seen in the difficult conflict category.

Performance comparison showing that D-AT-GRU significantly outperforms baselines in the conflict category, achieving 40.38% accuracy compared to the 25.00% of the previous state-of-the-art (GCAE).

As shown in the results, D-AT-GRU reached an overall accuracy of 78.50%. More impressively, it hit 40.38% in the conflict category—a massive leap from the 11.54% achieved by standard AT-LSTM and the 25.00% of the previous state-of-the-art, GCAE. Ablation studies confirmed that the orthogonal regularization was vital; without it, the dual attention heads tended to overlap, and conflict detection accuracy dropped significantly.

Conclusion

By breaking "conflict" into its constituent parts (positive + negative), the D-AT-GRU model leverages abundant pure-sentiment data to solve the problem of data scarcity in mixed-sentiment cases. While a 40% accuracy rate indicates that conflict opinions remain a significant challenge, this dual-attention approach represents a major step toward AI that can truly understand human nuance and dialectic thinking.

Summary updated

BibTeX
@inproceedings{tan2019recognizing,
  author = {Xingwei Tan and Yi Cai and Changxi Zhu},
  title = {{Recognizing Conflict Opinions in Aspect-level Sentiment Classification with Dual Attention Networks}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2019},
  doi = {10.18653/v1/d19-1342},
}
Figure