PaperlayerView exact revision on arXiv

Source-linked AI summary

A Survey on Aspect-Based Sentiment Analysis: Tasks, Methods, and Challenges

Wenxuan Zhang, Xin Li, Yang Deng, Lidong Bing, Wai Lam

arXiv:2203.01054v2cs.CL

TL;DR

ABSA lacks a systematic modern review covering its diverse tasks, compound sentiment-element relations, and practical transfer settings. This survey develops an element-based taxonomy, reviews task methods and PLM-based models, and discusses cross-domain, cross-lingual, emerging, and open challenges. It synthesizes recent advances while identifying important performance and evaluation limitations.

  • Problem

    Existing surveys do not comprehensively cover recent compound ABSA tasks, PLM-based models, or cross-domain and cross-lingual transfer.

  • Method

    The survey organizes ABSA studies by concerned sentiment elements and reviews task solutions, PLM utilization, cross-domain and cross-lingual methods, emerging topics, and open challenges.

  • Results

    The survey provides a systematic modern review of ABSA advances and challenges, emphasizing compound tasks and the advances and limitations of PLM-based ABSA.

  • Takeaways & Limitations

    The taxonomy and synthesis outline current ABSA progress and potential future directions across tasks, transfer settings, and emerging challenges.

  • Takeaways & Limitations

    ASQP remains especially challenging, with current best-performing models achieving about 40% F1 on benchmark datasets.

Abstract

from arXiv · show

As an important fine-grained sentiment analysis problem, aspect-based sentiment analysis (ABSA), aiming to analyze and understand people's opinions at the aspect level, has been attracting considerable interest in the last decade. To handle ABSA in different scenarios, various tasks are introduced for analyzing different sentiment elements and their relations, including the aspect term, aspect category, opinion term, and sentiment polarity. Unlike early ABSA works focusing on a single sentiment element, many compound ABSA tasks involving multiple elements have been studied in recent years for capturing more complete aspect-level sentiment information. However, a systematic review of various ABSA tasks and their corresponding solutions is still lacking, which we aim to fill in this survey. More specifically, we provide a new taxonomy for ABSA which organizes existing studies from the axes of concerned sentiment elements, with an emphasis on recent advances of compound ABSA tasks. From the perspective of solutions, we summarize the utilization of pre-trained language models for ABSA, which improved the performance of ABSA to a new stage. Besides, techniques for building more practical ABSA systems in cross-domain/lingual scenarios are discussed. Finally, we review some emerging topics and discuss some open challenges to outlook potential future directions of ABSA.

1 INTRODUCTION

ABSA addresses the need for fine-grained opinion analysis by identifying sentiment elements and their relationships. This survey organizes recent tasks and methods, emphasizing compound tasks, PLMs, and transfer across domains and languages.

  • ABSA analyzes opinions at the aspect level because sentence- or document-level sentiment can assume a single topic and sentiment incorrectly.
  • Compound ABSA tasks jointly extract multiple sentiment elements and model their correspondence, addressing limitations of single-element tasks.
  • Pre-trained language models substantially improve ABSA performance, while the survey examines both their advances and limitations.
  • Cross-domain and cross-lingual ABSA are reviewed because distribution shifts may require retraining despite the expense of collecting aspect-level labels.
  • Earlier surveys did not comprehensively cover recent compound tasks, PLM-based ABSA, or cross-domain and cross-lingual transfer.
  • The survey introduces a taxonomy organized by concerned sentiment elements and reviews methods for single and compound ABSA tasks.

2.1 Four Sentiment Elements of ABSA

ABSA represents opinions through four sentiment elements: aspect terms, aspect categories, opinion terms, and sentiment polarities. These elements support distinct extraction or classification tasks and organize the survey’s task taxonomy.

  • Aspect categories define domain-specific aspects, while aspect terms explicitly name opinion targets and use “null” for implicit targets.
  • Opinion terms are expressions of sentiment toward a target, whereas sentiment polarity describes its positive, negative, or neutral orientation.
  • The taxonomy lists representative methods for these tasks, including supervised and semi-supervised approaches for aspect term extraction.
  • The survey distinguishes aspect terms from aspect categories and uses “target” or “aspect” as general expressions for opinion targets.

2.2 ABSA Definition

ABSA identifies one or more sentiment elements in a text item, including dependency relations when multiple elements are involved. The survey therefore separates single-element and compound tasks.

  • ABSA identifies sentiment elements of interest in a text item, either singly or as multiple elements linked by dependency relations.
  • Single ABSA tasks predict one sentiment element, whereas compound tasks predict multiple coupled elements.
  • Although benchmark datasets commonly treat a sentence as the text item, the described methods naturally handle texts of any length.
  • Aspect term extraction extracts all aspect terms, while aspect-opinion pair extraction extracts all (a, o) pairs.
  • The taxonomy organizes ABSA studies by concerned sentiment elements and presents representative methods for each task.

2.3 Modeling Paradigms

The survey organizes ABSA solutions around four unified end-to-end paradigms and a pipeline paradigm for complex tasks. These paradigms map different input-output formats to shared computational frameworks.

  • Overview: ABSA commonly uses SeqClass, TokenClass, MRC, and Seq2Seq paradigms, while complex tasks may use pipelines combining multiple models.The paradigms differ in their input-output formats and modeling procedures.
  • Sequence-level Classification (SeqClass): Sequence-level classification encodes input text and applies a classifier to predict one-hot or multi-hot labels.The input may include multiple parts, such as a sentence and a specific aspect.
  • Token-level Classification (TokenClass): Token-level classification encodes an input sequence and decodes a label for each token.Decoders may use multilayer perceptrons with softmax or conditional random fields, together with schemes such as BIOES.
  • Machine Reading Comprehension (MRC): MRC extracts text spans from input text conditioned on a task-specific query by predicting span start and end positions.For ATE, the query can ask which aspect terms occur in the text.
  • Sequence-to-Sequence (Seq2Seq): Seq2Seq encodes an input sequence and generates an output sequence, such as desired sentiment elements or labels.The decoder generates tokens step by step using the encoded input and previous outputs.
  • Pipeline Method (Pipeline): Pipeline methods sequentially pass predictions between models, making them easier to implement but vulnerable to error propagation.Early errors can make final compound-task predictions incorrect, and pipelines often perform poorly on complex compound ABSA tasks.

2.4 Datasets & Evaluations

ABSA research relies on annotated benchmark datasets covering different sentiment elements, domains, and languages, with exact-match evaluation commonly summarized through classification metrics.

  • Datasets: Annotated datasets support ABSA method development and are documented by language, domain, annotated sentiment elements, and URL.The survey presents an overview of commonly used datasets and corresponding evaluation metrics.
  • Datasets: SemEval-2014, SemEval-2015, and SemEval-2016 datasets are widely used laptop and restaurant benchmarks for several ABSA tasks.They include annotations for aspect categories, aspect terms, and sentiment polarities, although coverage is incomplete across datasets.
  • Datasets: Opinion-term annotations were later added through another dataset and combined with earlier resources to create ASTE-Data-V2.ASTE-Data-V2 contains aspect term, opinion term, and sentiment information.
  • Datasets: Specialized resources include MAMS, which contains multiple aspects with different polarities, ASC-QA for Chinese QA-style reviews, and ARTS for robustness testing.These datasets target more challenging or specialized ABSA settings.
  • Evaluations: Exact-match evaluation counts a prediction as correct only when all predicted elements match the human annotation, after which accuracy, precision, recall, and F1 can be calculated.These metrics are used to compare methods across tasks and datasets.

3 SINGLE ABSA TASKS

Single ABSA tasks predict one sentiment element at a time: aspect terms, aspect categories, opinion terms, or sentiment polarity. The survey describes task-specific formulations and methods, including supervised, semi-supervised, unsupervised, neural, and syntactic approaches.

  • Task Overview: The four single ABSA tasks are aspect term extraction, aspect category detection, aspect sentiment classification, and opinion term extraction.Each task targets one of the four principal sentiment elements.
  • Aspect Term Extraction (ATE): ATE extracts explicit aspect expressions and includes supervised, semi-supervised, and unsupervised approaches.Supervised methods commonly formulate ATE as token-level classification using sequence-labeling models.
  • Aspect Term Extraction (ATE): Around 80% F1 scores are reported for supervised ATE on benchmark datasets, but these methods require substantial labeled data for sophisticated neural models.This data requirement motivates semi-supervised approaches using unlabeled text and pseudo-labeled examples.
  • Aspect Category Detection (ACD): ACD identifies domain-specific aspect categories, often as a multi-label classification problem, and can aggregate or recover targets beyond individual aspect terms.Unsupervised ACD first extracts candidate aspect terms, then maps or clusters them into predefined categories.
  • Opinion Term Extraction (OTE): OTE includes AOCE, which predicts aspect and opinion terms together, and TOWE, which extracts opinions corresponding to a given aspect term.TOWE models aspect-specific representations to identify the associated opinion expression.
  • Aspect Sentiment Classification (ASC): ASC predicts sentiment polarity for a specific aspect, either an aspect term or an aspect category, using contextual and structural information.Neural models fuse aspect and sentence context, while GNN-based methods exploit dependency-tree syntax.

4 COMPOUND ABSA TASKS

Compound ABSA tasks jointly extract multiple sentiment elements and their relations, extending single-element analysis to pairs, triplets, and quadruplets. The survey reviews pipeline, joint, unified, classification, MRC, and Seq2Seq solutions while highlighting unresolved comparisons and ASQP difficulty.

  • Compound ABSA Tasks: Compound ABSA tasks extract multiple sentiment elements while coupling them as pairs, triplets, or quadruplets.Their purpose is to capture more complete aspect-level sentiment structure than separate element extraction.
  • Aspect-Opinion Pair Extraction (AOPE): AOPE pairs extracted aspect and opinion terms, addressing the missing pairwise relation in aspect-opinion co-extraction.Pipeline solutions first extract elements and then classify candidate pairs, while unified methods jointly model terms and relations.
  • End-to-End ABSA (E2E-ABSA): E2E-ABSA extracts aspect terms and sentiment polarities simultaneously through pipeline, joint, or unified methods.Unified tagging can combine aspect-boundary labels with polarity labels, while opinion extraction is often used as an auxiliary task.
  • End-to-End ABSA (E2E-ABSA): Around 70% F1 is reported for recent pipeline, unified, and joint E2E-ABSA methods, leaving the most suitable paradigm unclear.Earlier studies favor different paradigms, so performance comparisons remain unresolved.
  • Aspect Category Sentiment Analysis (ACSA): ACSA jointly predicts aspect categories and their sentiment polarities, with unified methods organized into four types.The types are Cartesian product, add-one-dimension, hierarchy classification, and Seq2Seq modeling; Seq2Seq results outperform earlier classification models and show advantages in few-shot and zero-shot settings.
  • Aspect Sentiment Quad Prediction (ASQP): ASQP predicts aspect, aspect term, polarity, and opinion term together, providing the most complete but most challenging ABSA output.Implicit elements can be represented with “null”; current best-performing models achieve about 40% F1 on benchmark datasets.

5 ABSA WITH PRE-TRAINED LANGUAGE MODELS

Pre-trained language models have substantially advanced ABSA by providing contextualized representations that improve performance across tasks, while raising challenges around task-specific structure, efficient representations, and robustness.

  • PLMs address limitations of context-independent word embeddings by supplying contextualized representations for ABSA models.Conventional architectures coupled task-specific neural models with Word2Vec or GloVe, whose gains eventually reached a bottleneck.
  • Simple PLM integration can outperform carefully designed neural architectures on End-to-End ABSA and establish state-of-the-art ASC results.Reported approaches replace word embeddings or feed sentence–aspect pairs into PLMs, sometimes with additional domain- and task-specific post-training.
  • Complex ABSA tasks require designs that model dependency relations among sequence- or token-level labels beyond using PLMs as embedding layers.The survey distinguishes task complexity from the challenge of adequately inducing knowledge learned during PLM pre-training.
  • PLMs also support generative data augmentation and improved dependency trees, extending their utility beyond serving as model backbones.Generative augmentation produces semantic-preserved training data, while better dependency trees can benefit relation-aware ABSA methods.
  • PLM-based ABSA still faces redundancy and robustness issues, including over 25% performance drops on adversarial examples for ASC.The survey calls for meaningful sparse structures, more efficient ABSA representations, and stronger robustness.

6 TRANSFERABLE ABSA

Transferable ABSA addresses distribution shifts across domains and languages, where labeled data is costly and same-distribution assumptions often fail. The survey reviews feature, data, translation, embedding, and multilingual-PLM strategies while identifying compound cross-lingual transfer as an open challenge.

  • 6.1 Cross-domain ABSA: Cross-domain ABSA is needed because models trained and tested within one distribution may fail when domains change, while new fine-grained labels are expensive to collect.Domain adaptation provides lower-cost alternatives to repeatedly gathering large labeled datasets.
  • 6.1 Cross-domain ABSA: Feature-based transfer learns domain-independent representations using syntactic features or auxiliary dependency prediction tasks.These approaches aim to reduce domain shift at the word or representation level.
  • 6.1 Cross-domain ABSA: Data-based transfer adjusts training-data distributions through high-precision syntactic patterns, domain-independent opinion terms, and pseudo-labeled target data.Pseudo-labeled target examples can be added to source-domain training data for cross-domain modeling.
  • 6.1 Cross-domain ABSA: Combining feature- and data-based transfer, including instance re-weighting and domain-specific PLM pre-training, improves cross-domain E2E-ABSA generalization.The survey presents this consolidation as a promising direction, with PLMs also providing domain-independent representations.
  • 6.2 Cross-lingual ABSA: Cross-lingual ABSA commonly uses translation and label projection, cross-lingual embeddings, or multilingual PLMs fine-tuned on source-language data.Translated target-language data and alignment-free label projection can produce pseudo-labeled data that supports strong XABSA baselines.
  • 6.2 Cross-lingual ABSA: XABSA remains underexplored, especially for difficult compound tasks, and multilingual PLMs require better adaptation to target-language knowledge.Existing studies primarily address relatively easier tasks such as cross-lingual aspect term extraction.

7 CHALLENGES AND FUTURE DIRECTIONS

The survey identifies future directions for making ABSA more reliable and practical, including larger and more diverse datasets, multimodal analysis, unified models, and broader lifelong-learning settings.

  • 7.1 Quest for Larger and More Challenging Datasets: Small datasets derived largely from SemEval make model comparisons difficult, particularly for parameter-heavy PLM-based systems.The survey recommends larger datasets and more challenging collections spanning domains, languages, and opinion-sharing platforms.
  • Multimodal ABSA should exploit related textual and visual information because users express opinions through images as well as text.Existing multimodal work mainly targets simple tasks, while multimodal E2E-ABSA introduces joint learning with auxiliary cross-modal relation detection.
  • Unified ABSA models could handle multiple related tasks simultaneously instead of requiring a separate task-specific model for each task.The survey motivates this direction by noting that ideas and model designs can often transfer across closely related ABSA tasks.
  • Lifelong ABSA must move beyond domain-incremental ASC toward learning changing classes and tasks, including varying aspect-category sets.More advanced systems also need to adapt when applications introduce different categories and cross-domain settings.

8 CONCLUSIONS

The survey comprehensively reviews ABSA tasks, methods, challenges, and future directions, emphasizing compound tasks, PLMs, and transferable systems.

  • The survey organizes ABSA research by sentiment elements and modeling paradigms, emphasizing recent compound-task advances.It covers task solutions, PLM benefits and limitations, cross-domain and cross-lingual progress, and emerging challenges.
  • The paper frames cross-domain and cross-lingual ABSA as routes toward more practical systems while identifying open challenges for future research.
Loading 2203.01054v2…