Deep Learning for Aspect-Level Sentiment Classification: Survey, Vision, and Challenges

Jie Zhou, Jimmy Xiangji Huang, Qin Chen, Qinmin Vivian Hu, Tingting Wang, Liang He

Introduction: The Need for Granular Sentiment

Standard sentiment analysis often operates at the document or sentence level, assuming a single overarching opinion. However, real-world text is rarely that simple. A sentence-level classifier might label a review like "The pizza was delicious, but the service was incredibly slow" as "neutral" or "mixed," missing the vital detail that the user is happy with the food but unhappy with the service.

Aspect-Level Sentiment Classification (ASC) addresses this by assigning sentiment polarities (positive, negative, or neutral) to specific "aspects" or terms mentioned in the text. This nuance is critical for applications like product feedback analysis and market research, where understanding specific pain points is more valuable than a general score.

An illustration of Aspect-Level Sentiment Classification (ASC) identifying distinct sentiments for different aspects within a single sentence.

A Taxonomy of Deep Learning Approaches

While deep learning has revolutionized ASC, the research landscape has historically been fragmented, with researchers using varied datasets and inconsistent evaluation metrics. This paper introduces a systematic taxonomy to categorize models into five distinct families based on their underlying architecture, providing a clear roadmap for the field.

A systematic taxonomy of deep learning approaches for ASC, categorized by architectural families.
  1. Recursive Neural Networks (RecNN): These models leverage the structural syntax of a sentence, such as a parse tree, to build representations from the bottom up.
  2. Recurrent Neural Networks (RNN): Utilizing sequence models like LSTMs and GRUs, these approaches often use "Target-Dependent" LSTMs (TD-LSTM) to model the specific context preceding and following an aspect.
  3. Attention-based RNNs: Currently the most prevalent family, these models use attention mechanisms to "focus" on the most relevant words. They are divided into Basic (one-way attention) and Interactive (where the aspect and context influence each other's representations).
  4. Convolutional Neural Networks (CNN): These use filters to capture local patterns and n-grams, often employing gating mechanisms to incorporate aspect-specific information.
  5. Memory Networks: These utilize an external memory component to store context, allowing for multiple "hops" or passes over the data to capture long-distance dependencies.

Standardizing Datasets and Evaluation

A significant contribution of this work is the unification of major benchmarks into a standardized XML/JSON format. This includes the SemEval (2014, 2015, 2016) datasets for Laptops and Restaurants, a large-scale Twitter dataset, and specialized sets like SentiHood and MPQA. By standardizing these "rules of the game," the authors enable the first truly rigorous head-to-head comparison of state-of-the-art models.

Key Insights and Benchmarking Results

The authors' re-implementation and benchmarking of over a dozen models yielded several critical insights for future research:

  • Aspect Awareness is Mandatory: Models that explicitly incorporate aspect information (e.g., via aspect embeddings or targeted attention) consistently outperform generic sentiment classifiers.
  • The Power of Interaction: "Interactive" attention models (like IAN) generally perform better than basic ones. Modeling the bidirectional relationship—how the context defines the aspect and how the aspect highlights specific words in the context—is crucial for accuracy.
  • RNNs > CNNs: For the specific task of ASC, RNN-based architectures tend to outperform CNNs because they are better at capturing the long-range sequential dependencies inherent in complex sentences.
  • The "Neutral" Struggle: A major finding is that almost all models struggle significantly with the "neutral" class. F1 scores for neutral sentiment are often low, likely due to class imbalance in the training data.
  • Performance Baselines: On the popular Restaurants14 dataset, top-tier models like AOA-LSTM (Attention-over-Attention) and RAM (Recurrent Attention on Memory) achieved accuracies around 80-81%. However, simple baselines like ContextAvg (averaging word embeddings) perform surprisingly well, sometimes beating older, more complex neural architectures.

Challenges and Future Directions

The paper identifies several frontiers for the field:

  • The BERT Era: Integrating large-scale pre-trained models like BERT is the next logical step for a massive performance boost beyond traditional LSTMs.
  • Explainability: While attention weights provide some insight, improving the "interpretability" of these models is vital for industry adoption.
  • Commonsense Knowledge: Current models struggle with implicit sentiment. Knowing that a "long wait" is bad requires commonsense knowledge that isn't always present in the text itself.
  • Joint Tasks: Future work should focus on "End-to-End" ASC—identifying the aspects and their sentiments simultaneously, rather than assuming aspects are pre-identified.

Summary updated

Abstract

This survey focuses on deep learning-based aspect-level sentiment classification (ASC), which aims to decide the sentiment polarity for an aspect mentioned within the document. Along with the success of applying deep learning in many applications, deep learning-based ASC has attracted a lot of interest from both academia and industry in recent years. However, there still lack a systematic taxonomy of existing approaches and comparison of their performance, which are the gaps that our survey aims to fill. Furthermore, to quantitatively evaluate the performance of various approaches, the standardization of the evaluation methodology and shared datasets is necessary. In this paper, an in-depth overview of the current state-of-the-art deep learning-based methods is given, showing the tremendous progress that has already been made in ASC. In particular, first, a comprehensive review of recent research efforts on deep learning-based ASC is provided. More concretely, we design a taxonomy of deep learning-based ASC and provide a comprehensive summary of the state-of-the-art methods. Then, we collect all benchmark ASC datasets for researchers to study and conduct extensive experiments over five public standard datasets with various commonly used evaluation measures. Finally, we discuss some of the most challenging open problems and point out promising future research directions in this field.
BibTeX
@article{zhou2019deep,
  author = {Jie Zhou and Jimmy Xiangji Huang and Qin Chen and Qinmin Vivian Hu and Tingting Wang and Liang He},
  title = {{Deep Learning for Aspect-Level Sentiment Classification: Survey, Vision, and Challenges}},
  journal = {IEEE Access},
  year = {2019},
  doi = {10.1109/access.2019.2920075},
}
Figure