[Paper Review] Aspect Based Sentiment Analysis with Gated Convolutional Networks

Wei Xue, Tao Li|arXiv (Cornell University)|May 18, 2018

Sentiment Analysis and Opinion Mining17 references67 citations

TL;DR

The paper introduces GCAE, a CNN-based model with Gated Tanh-ReLU Units that selectively extract aspect-specific sentiment features for ABSA tasks (ACSA and ATSA), offering faster training and competitive accuracy without recurrent layers or traditional attention mechanisms.

ABSTRACT

Aspect based sentiment analysis (ABSA) can provide more detailed information than general sentiment analysis, because it aims to predict the sentiment polarities of the given aspects or entities in text. We summarize previous approaches into two subtasks: aspect-category sentiment analysis (ACSA) and aspect-term sentiment analysis (ATSA). Most previous approaches employ long short-term memory and attention mechanisms to predict the sentiment polarity of the concerned targets, which are often complicated and need more training time. We propose a model based on convolutional neural networks and gating mechanisms, which is more accurate and efficient. First, the novel Gated Tanh-ReLU Units can selectively output the sentiment features according to the given aspect or entity. The architecture is much simpler than attention layer used in the existing models. Second, the computations of our model could be easily parallelized during training, because convolutional layers do not have time dependency as in LSTM layers, and gating units also work independently. The experiments on SemEval datasets demonstrate the efficiency and effectiveness of our models.

Motivation & Objective

• Motivate ABSA to provide fine-grained sentiment toward predefined aspects or targets in text (ACSA and ATSA).
• Propose a CNN-based architecture using gating mechanisms to selectively propagate sentiment features conditioned on the given aspect.
• Eliminate dependence on LSTMs/attention to improve training efficiency while maintaining or improving accuracy.
• Extend the model to handle both aspect categories (ACSA) and aspect terms (ATSA) with shared CNN components.
• Demonstrate effectiveness on SemEval restaurant and laptop datasets through thorough experiments.

Proposed method

• Propose Gated Convolutional Network with Aspect Embedding (GCAE) leveraging two parallel convolutional streams for sentiment and aspect features.
• Introduce Gated Tanh-ReLU Units (GTRU) that fuse sentiment and aspect cues via element-wise gating at each position.
• Compute sentiment features a_i via ReLU-activated convolution with aspect embedding, compute aspect features s_i via tanh-activated convolution, and output c_i = s_i * a_i.
• Apply max-over-time pooling on the gated features to form a fixed-size sentence representation for classification.
• For ATSA, add a small CNN over aspect terms to generate guiding features for the gates; for ACSA, use a single aspect embedding.
• Train end-to-end with cross-entropy loss over predicted sentiment polarities.

Experimental results

Research questions

• RQ1 Can a CNN-based architecture with per-position gated units match or surpass LSTM/attention-based ABSA models on ACSA and ATSA tasks?
• RQ2 Do gating mechanisms conditioned on aspect information improve the extraction of aspect-specific sentiment features compared to standard CNNs?
• RQ3 Is the proposed GCAE training-time efficient due to parallelizable convolutions and gating, relative to LSTM/attention-based approaches?
• RQ4 How does GCAE perform on hard ABSA test sets featuring sentences with multiple aspects and mixed sentiments?
• RQ5 Can the ATSA extension with an auxiliary CNN over aspect terms effectively guide sentiment composition for multi-word targets?

Key findings

• GCAE achieves higher accuracy than several neural baselines on restaurant ABSA tasks and shows notable gains on hard test sets with multiple sentiments per sentence.
• GCAE demonstrates faster training convergence than LSTM/attention-based models (training time substantially reduced).
• GTRU gating conditioned on aspect information improves selective propagation of sentiment features, outperforming GTU and GLU variants in ABSA tasks.
• ATSA extension with aspect-term CNNs provides competitive results against LSTM-based baselines, with GCAE often outperforming in hard-test scenarios.
• Convolutional architecture enables parallelizable computation, achieving efficiency without sacrificing accuracy.