Authors: Kai Sun, Richong Zhang, Samuel Mensah, Yongyi Mao, Xudong Liu. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP).
Introduction
Aspect-Based Sentiment Analysis (ABSA) aims to identify the sentiment expressed toward specific entities or attributes within a sentence. While traditional models excel at capturing local context, they often struggle with complex linguistic structures where opinion words are distant from their target aspects. This paper introduces the CDT (Convolution over Dependency Tree) model, which leverages Graph Convolutional Networks (GCNs) to bridge these gaps by propagating sentiment information along syntactic paths.
The Challenge of Long-Range Dependencies
In sentiment analysis, identifying the sentiment toward a specific aspect is more granular than general sentence-level classification. For instance, in the sentence "The food was delicious, but the service was slow," the sentiment is positive for "food" but negative for "service."
A significant challenge in ABSA is long-range dependency. In complex sentences, an opinion word might be separated from its target aspect by many irrelevant words. Standard sequential models like LSTMs process text linearly and often dilute the connection between these distant but related terms.
As shown in the figure above, the linear distance between "branzino" and "difficult" is 11 words. However, in a dependency tree, these words are connected via a much shorter syntactic path, allowing the model to "shortcut" the linear distance.
Methodology: The CDT Model
The proposed CDT model integrates sequential context with syntactic structure through three primary stages:
- Contextual Embedding (Bi-LSTM): The sentence is first processed by a Bi-directional Long Short-Term Memory network to capture local context and word order.
- Syntactic Enhancement (GCN): The Bi-LSTM hidden states serve as initial node features for a GCN. This network operates directly on the sentence's dependency tree, allowing information to flow along grammatical edges. This enables aspect words to aggregate sentiment from their relevant modifiers regardless of their position in the sentence.
- Aspect-Centric Pooling: After several GCN layers, the model performs average pooling on the vectors representing the aspect words. Because the GCN has propagated relevant information into these specific nodes, the aspect vectors alone provide a robust representation for sentiment classification.
Experimental Results
The CDT model was evaluated on four benchmark datasets: Rest14, Laptop14, Rest16, and a Twitter collection. Using pre-trained GloVe vectors and the Stanford Parser for dependency trees, the model achieved state-of-the-art (SOTA) performance:
- Accuracy: CDT reached 82.30% on Rest14 and 77.19% on Laptop14, outperforming previous leaders like MGAN and TNet.
- F1-Score: The model showed consistent improvements, particularly on the Rest16 dataset (69.93).
- Ablation Studies: Comparisons with "ASP-GCN" (GCN only) and "ASP-BiLSTM" (LSTM only) demonstrated that combining sequential context with syntactic structure is essential for high-accuracy sentiment analysis.
Conclusion
The CDT model demonstrates that syntax acts as a powerful shortcut for sentiment propagation. By stacking GCNs on top of LSTMs, the model captures both the "flow" of natural language and the underlying "logic" of grammar. While the model's performance is dependent on the quality of the underlying dependency parser, it represents a significant step forward in handling the structural complexity of human sentiment.
Summary updated