PaperlayerView exact revision on arXiv

Source-linked AI summary

A Survey on Aspect-Based Sentiment Classification

Gianni Brauwers, Flavius Frasincar

arXiv:2203.14266v1cs.CL

TL;DR

The growing volume of online reviews creates demand for fine-grained sentiment analysis toward individual aspects. This survey synthesizes ABSC research through a three-part taxonomy, reviews representations and evaluation, and identifies attention-based deep learning and knowledge-enhanced approaches as central directions.

  • Problem

    ABSC research is rapidly evolving, while prior surveys do not comprehensively cover knowledge-based methods, transformer models, hybrid approaches, input representations, and evaluation.

  • Method

    The survey reviews ABSC inputs, evaluation techniques, and models using a taxonomy of knowledge-based, machine learning, and hybrid approaches.

  • Results

    The survey identifies attention-based deep learning as the leading direction and reports that combining symbolic knowledge with subsymbolic learning produces better results than either separately.

  • Takeaways & Limitations

    ABSC research can be understood through a unified taxonomy that connects model architectures with input representation, evaluation, and knowledge integration.

  • Takeaways & Limitations

    ABSC models are concentrated mainly on restaurant reviews, and datasets remain limited for languages other than English.

Abstract

from arXiv · show

With the constantly growing number of reviews and other sentiment-bearing texts on the Web, the demand for automatic sentiment analysis algorithms continues to expand. Aspect-based sentiment classification (ABSC) allows for the automatic extraction of highly fine-grained sentiment information from text documents or sentences. In this survey, the rapidly evolving state of the research on ABSC is reviewed. A novel taxonomy is proposed that categorizes the ABSC models into three major categories: knowledge-based, machine learning, and hybrid models. This taxonomy is accompanied with summarizing overviews of the reported model performances, and both technical and intuitive explanations of the various ABSC models. State-of-the-art ABSC models are discussed, such as models based on the transformer model, and hybrid deep learning models that incorporate knowledge bases. Additionally, various techniques for representing the model inputs and evaluating the model outputs are reviewed. Furthermore, trends in the research on ABSC are identified and a discussion is provided on the ways in which the field of ABSC can be advanced in the future.

1 INTRODUCTION

ABSC targets fine-grained sentiment toward individual aspects, extending sentiment analysis beyond general document or sentence polarity. The survey addresses gaps in prior reviews with a broader taxonomy and coverage of inputs, evaluation, and models.

  • Task scope: Sentiment analysis can use binary, neutral-inclusive, or intensity-based outputs and operate at different textual levels.The task’s granularity depends on sentiment type, analysis level, and target.
  • Task scope: ABSC identifies sentiment associated with particular aspects in documents or sentences, rather than only general text-level sentiment.ABSA methods may detect aspects, classify sentiment, and aggregate results at document or review level.
  • Survey contribution: The survey broadens prior ABSC reviews by covering knowledge bases, transformer models, and a taxonomy of knowledge-based, machine learning, and hybrid approaches.It also compares reported models and discusses current and future research directions.
  • Survey structure: The survey organizes discussion around input representation, performance evaluation, ABSC models, and additional research topics.These sections follow the main steps involved in designing ABSC methods.

2 INPUT REPRESENTATION

ABSC input representation converts aspect-specific text context into numeric vectors or matrices for classification. The survey distinguishes explicit and implicit aspects and reviews context selection, classical features, and embeddings.

  • Task inputs: ABSC models assume target aspects are identified and classify the sentiment y corresponding to each aspect a.The survey formalizes ABSA with sentiment, aspect, holder, and time, but focuses on the sentiment–aspect pair.
  • Context: Context is represented separately for each aspect, with methods differing according to whether the aspect is explicit or implicit.Multiple aspects require aspect-specific representations rather than a single undifferentiated text representation.
  • Context: Explicit-aspect representations may use nearby words, but physical proximity can miss sentiment expressions located farther from the target.More robust methods avoid relying solely on word distance.
  • Feature representations: Single-vector models use features such as bag-of-words and TF-IDF, while sequence models can process sets of word vectors or matrices.Bag-of-words ignores word order and can produce high-dimensional vectors; TF-IDF rescales word frequency by corpus-wide frequency.
  • Feature limitations: Bag-of-words features can make classification difficult because they disregard structure and create large feature spaces.Stop-word filtering and n-grams address different issues but do not eliminate the dimensionality challenge.
  • Feature representations: Word embeddings represent words as limited-size vectors, while contextual embeddings such as ELMo and BERT vary with surrounding context.Contextual embeddings address the fact that a word can have multiple meanings in different contexts.

3 PERFORMANCE EVALUATION

ABSC evaluation compares predicted sentiment labels with true labels using standard classification measures and alternatives for ordinal or ranking-sensitive errors. Different measures emphasize different aspects of model quality.

  • Classification metrics: Accuracy, precision, recall, and F1-measure are the most commonly used metrics for comparing ABSC predictions with true sentiment labels.These measures assess classification quality from complementary perspectives.
  • Error measures: Mean squared error and ranking loss provide alternatives for evaluating prediction errors when sentiment labels are ordinal.Both measures assess errors in classification predictions, with lower values indicating better performance.
  • Error measures: MSE penalizes large errors more than small errors because the prediction error is squared.Ranking loss treats small and large errors more equally than MSE.
  • Ranking metrics: AUC measures how well a model separates classes using the area under the ROC curve.The ROC curve relates recall to the true positive rate.

4 SENTIMENT CLASSIFICATION

The survey classifies ABSC models into knowledge-based, machine learning, and hybrid categories. It summarizes representative models, data types, and reported performance measures within this taxonomy.

  • Taxonomy: ABSC models are organized into three major categories: knowledge-based approaches, machine learning models, and hybrid models.The taxonomy also contains subcategories for more detailed discussion.
  • Model comparison: The model overview tables report model types, data used, and performance measures across datasets and domains.Rows may contain multiple entries when a model is evaluated on different datasets or reimplemented elsewhere.

4.1 Knowledge-Based

Knowledge-based ABSC methods use knowledge bases to support sentiment classification, emphasizing transparent reasoning without model training. The survey covers dictionary-, ontology-, and discourse-based approaches, including methods that use contextual relations to determine sentiment.

  • Overview: Knowledge-based methods use stored rules, relations, and assumptions to define features for ABSC.They are also known as symbolic AI approaches.
  • Overview: Their main advantages are interpretability and transparency, while constructing the required knowledge bases can take considerable time.These methods require no training time, but knowledge-base construction remains costly in effort.
  • Dictionary-Based: Dictionary-based methods build a feature vector whose elements encode sentiment scores for context words surrounding an aspect.WordNet and SentiWordNet are examples of dictionaries used to represent linguistic relations and sentiment information.
  • Ontology-Based: Ontology-based methods use relations between real-world entities to identify relevant context and determine an aspect’s sentiment label.The sentiment decision can then use a classifier, such as a dictionary-based sentiment classifier.
  • Discourse-Based: Discourse-based methods use rhetorical or cross-document structures to identify important words and aggregate sentiment across related reviews.RST organizes text into elementary discourse units, while the SMACk system analyzes relations among arguments in responding reviews.

4.2 Machine Learning

Machine-learning ABSC models span classical classifiers, tree ensembles, recurrent and convolutional networks, and attention-based architectures. Their suitability varies with data characteristics, computational resources, feature-engineering costs, and task performance.

  • Model families: Machine-learning ABSC models include support vector machines, tree-based models, deep learning models, and attention-based deep learning models.The models learn patterns from feature vectors and corresponding sentiment labels.
  • Classical models: SVMs separate aspect feature vectors into sentiment classes and can use kernels, one-versus-all classification, and robust high-dimensional learning.Kernel transformations improve separability but make learned coefficients harder to interpret; SVMs are also described as robust to noisy data.
  • Classical models: Decision trees can outperform SVMs on some datasets, whereas SVMs perform better on other problems.The survey reports both outcomes, indicating that model performance depends on the problem setting.
  • Classical models: Random forests reduce decision-tree overfitting by restricting features and training each tree on a bootstrapped data sample.Their reported SemEval-2014 ABSC results are mixed, while gradient-boosted and extra trees may provide slight improvements over random forests.
  • Deep learning: RNNs process word sequences, while RecNNs combine word representations through tree structures that can follow relations beyond sequential word order.RecNNs require a tree structure, commonly produced by a parser, and can process inputs according to the relations defining that tree.
  • Attention-based deep learning: Attention-based models summarize aspect-relevant context, capture long-term dependencies, and include transformer architectures that use self-attention without a separate RNN or CNN base.The survey notes that attention weights can offer interpretability, although their usefulness as explanations remains controversial; transformer variants achieve significant improvements over previous models.

4.3 Hybrid

Hybrid ABSC models combine machine learning with knowledge bases to address limited data, with dictionary-, ontology-, and other structure-enhanced approaches reporting benefits in several settings.

  • Hybrid models: Hybrid ABSC models combine machine learning with knowledge bases, addressing limited data availability that can hinder model training, especially in niche domains and languages.Knowledge bases can provide features or other information to compensate for insufficient training data.
  • Dictionary-enhanced machine learning: Dictionary-enhanced methods incorporate sentiment scores or lexicon features into machine learning models such as SVMs, LSTMs, CNNs, and Bi-GRUs.The survey describes SentiWordNet-based features and lexicon-enhanced neural models across multiple applications.
  • Other hybrid approaches: Hybrid systems can also combine symbolic and subsymbolic AI, as SenticNet and Sentic LSTM integrate linguistic knowledge, logic, and deep learning for sentiment analysis.The survey reports that Sentic LSTM performs better than symbolic and subsymbolic AI used separately.
  • Ontology-enhanced machine learning: Ontology-enhanced models use concepts and sentiment information as features or combine ontology-based classification with a machine learning backup classifier.These approaches can define concept-based representations or resolve missing or conflicting ontology classifications with a learned model.
  • Ontology-enhanced machine learning: Ontology features yield significantly higher F1-score in one SemEval-2016 evaluation and can preserve aspect-extraction performance with only 20% of the original training data.The reported performance drop is less than 10% for the ontology-enhanced SVM, while base methods decline significantly.

5 RELATED TOPICS

The survey relates ABSC to interconnected tasks and linguistic phenomena, emphasizing that aggregation cannot be fully separated from classification and that sarcasm and thwarting remain difficult, underexplored challenges.

  • ABSA task relationships: ABSA comprises detection, classification, and aggregation, but these steps are interdependent and may need to be designed jointly.Information extracted during classification can be useful during aggregation.
  • Thwarting and sarcasm: Thwarting reverses a document’s overall sentiment relative to the sentiment expressed through most of its text, making word-level aggregation unreliable.The challenge arises because the broader context can contradict the dominant local sentiment.
  • Thwarting and sarcasm: Sarcasm uses language contrary to its intended meaning and is especially common in social media, where emoticons and hashtags may also matter.The survey notes that sarcasm recognition is particularly difficult and often involves additional social-media cues.
  • Thwarting and sarcasm: Thwarting and sarcasm make sentiment analysis non-trivial, yet they have received little attention specifically in ABSC research.The survey describes ontology, attention-based, and cognitive-feature approaches from related sentiment-analysis literature.
  • Emotion analysis: Emotion analysis differs from sentiment analysis by considering a broad range of emotions, such as joy, sadness, and anger, rather than primarily polarity labels or scores.The survey presents emotion analysis as a related topic in conjunction with ABSC.

6 CONCLUSION

The survey synthesizes ABSC models, datasets, evaluation practices, and future research directions. It highlights attention-based deep learning, knowledge-enhanced methods, data limitations, and applications beyond current sentence-level settings.

  • The survey organizes ABSC around input representation, aspect sentiment classification, and output evaluation, supported by a taxonomy and performance tables.It combines technical explanations, intuitive descriptions, and reported model performances.
  • Performance comparisons are difficult when studies use independently collected datasets, motivating more consistent dataset development.Early ABSC studies often scraped and compiled their own Web datasets.
  • ABSC research remains constrained by limited datasets outside English, making knowledge bases and simpler models important alternatives or complements.The survey reports that knowledge bases can help compensate for limited data, although current integrations remain relatively basic.
  • Weakly supervised methods can address situations where labeled ABSC data is expensive to obtain.The survey gives joint aspect extraction and sentiment classification as an example of weak supervision using an attention-based sentiment dictionary.
  • Attention-based deep learning methods rapidly surpassed earlier approaches, while transformer models achieved state-of-the-art ABSC performance.The survey identifies attention-based architectures as the field’s prevailing direction.
  • Future ABSC applications should address implicit and explicit aspects together, review-level sentiment, and sentiment changes over time.The survey also identifies specialized opinion-based search engines and new temporal datasets as directions for further work.
Loading 2203.14266v1…