PaperlayerView exact revision on arXiv

Source-linked AI summary

A Systematic Review of Aspect-based Sentiment Analysis: Domains, Methods, and Trends

Yan Cathy Hua, Paul Denny, Katerina Taskova, Jörg Wicker

arXiv:2311.10777v6cs.CLcs.LG

TL;DR

Existing ABSA reviews often emphasize methods and subtasks, leaving trends and relationships among research domains, datasets, and approaches less examined. This systematic review synthesizes studies across those dimensions and finds substantial concentration in product/service-review data and limited domain diversity, alongside trends in ABSA research.

  • Problem

    ABSA depends on domain and context, but prior reviews rarely quantify trends or relationships among research domains, datasets, and solution approaches.

    Source: §2.3 · p. 3, §2.4 · p. 4

  • Method

    The authors systematically reviewed ABSA literature, analyzing research domains, dataset domains, and solution approaches, with a targeted phase on generative-model in-context learning.

    Source: §3.4 · p. 6, §5 · p. 16

  • Results

    Most studies lacked a specific application domain (65.32%), while 70.95% of datasets were product/service reviews and public-sector research was scant.

    Source: §4.1 · p. 7

  • Takeaways & Limitations

    The review identifies potential systemic issues involving dataset and domain diversity, domain mismatch, and concentration in product/service-review resources.

    Source: §5 · p. 16, §6 · p. 21

    The sample covers English-language primary studies from four peer-reviewed databases and therefore does not generalize beyond that scope without assumptions.

    Source: §5.3 · p. 20

Abstract

from arXiv · show

Aspect-based sentiment analysis (ABSA) is a fine-grained type of sentiment analysis that identifies aspects and their associated opinions from a given text. With the surge of digital opinionated text data, ABSA gained increasing popularity for its ability to mine more detailed and targeted insights. Many review papers on ABSA subtasks and solution methodologies exist, however, few focus on trends over time or systemic issues relating to research application domains, datasets, and solution approaches. To fill the gap, this paper presents a systematic literature review (SLR) of ABSA studies with a focus on trends and high-level relationships among these fundamental components. This review is one of the largest SLRs on ABSA. To our knowledge, it is also the first to systematically examine the interrelations among ABSA research and data distribution across domains, as well as trends in solution paradigms and approaches. Our sample includes 727 primary studies screened from 8550 search results without time constraints via an innovative automatic filtering process. Our quantitative analysis not only identifies trends in nearly two decades of ABSA research development but also unveils a systemic lack of dataset and domain diversity as well as domain mismatch that may hinder the development of future ABSA research. We discuss these findings and their implications and propose suggestions for future research.

1 Introduction

ABSA extracts subjects and their associated opinions from opinionated text, enabling targeted analysis of sentiment. This review examines broad trends and relationships across ABSA research domains, datasets, and solution approaches.

1 Introduction

  • ABSA identifies subjects and associated opinions in text, addressing questions about what was mentioned, what people thought, and how opinions are distributed.The review situates this task amid the growing volume of digital opinionated text and the need for automated analysis.Source: §1 · pp. 1–2

1.1 Review Focus and Research Questions

  • The review asks how commercial-domain research and datasets are distributed, which subtask combinations are common, and how solution approaches have changed over time.Its questions also ask whether linguistic and traditional machine-learning approaches remain in use.Source: §1.1 · p. 2
  • The review combines 727 primary studies from 2008–2024, examines relationships among domains, data resources, and methodologies, and uses PDF-mining automatic filtering.The authors describe it as one of the largest scoped ABSA SLRs and, to their knowledge, the first to systematically examine these interrelations.Source: §1.1 · p. 2

1.2 Organisation of This Review

  • The paper is organized into background, methods, results, discussion, and conclusion sections, with fuller methods and supplementary figures in the appendices.Appendix A provides an introduction to ABSA and its subtasks; Appendix B gives methods details, and Appendix C provides additional results figures.Source: §1.2 · p. 2

2 Background

ABSA targets sentiment toward specific entities and aspects, with subtasks and solution designs shaped by context and domain. The review is motivated by gaps in evidence about dataset diversity, domain coverage, and changes in methods over time.

2.1 ABSA: a fine-grained sentiment analysis

  • ABSA extends sentiment analysis beyond document- or sentence-level polarity to more targeted sentiment about entities or their aspects.Sentiment analysis may classify polarity or intensity; ABSA aims to identify the subjects of opinions and their sentiment.Source: §2.1 · p. 2

2.2 ABSA and its subtasks

  • Core ABSA subtasks include aspect and opinion extraction and aspect-sentiment classification, while composite tasks combine subtasks in end-to-end solutions.Examples include AOPE, APCE, ASTE, and ASQE/ASQP; multi-task learning can use shared learning to capture inter-task relations.Source: §2.2 · p. 3

2.3 The context- and domain-dependency challenges

  • ABSA is domain- and context-dependent: the same expression can carry different meanings or sentiment across contexts, and some tasks remain difficult for language models.The cited challenging cases include implicit and multi-word aspect extraction, extraction and category detection on mixed-domain corpora, and context-dependent sentiment classification.Source: §2.3 · p. 3
  • Domain-specific datasets and resources can support ABSA quality, but annotated data is costly and limited, making domain adaptation an ongoing challenge.The review raises the possibility that dataset-reliant methods and skewed dataset-domain distributions could hinder performance and generalisability.Source: §2.3 · p. 3, §2.4 · p. 3

2.4 Review Rationale

  • Earlier ABSA work and benchmarks concentrated on product and service reviews, while other domains—including public-sector areas—also contain opinionated text relevant to ABSA.A cited review of 62 public datasets found 53 were reviews of restaurants, hotels, or digital products; only one covered the public sector.Source: §2.4 · pp. 3–4
  • Prior surveys largely detail subtasks or methods, while few quantify combinations or trends; the authors use a large-scale SLR to address these evidence gaps.The review asks about domain and dataset concentration, subtask formulations, and changes in solution approaches over time.Source: §2.4 · p. 4

3 Methods

The authors conduct a systematic review of ABSA studies, manually classifying research and dataset domains and quantitatively analyzing the extracted data. A targeted second phase adds a review of in-context learning studies through July 2024.

3.1 Main Procedures

  • The main review screened 4,191 database results without publication-year constraints and included 519 peer-reviewed studies after automatic and manual screening.The team then recorded study data using a planned scheme, cleaned the extracted data, and conducted quantitative analysis against its research questions.Source: §3.1 · pp. 4–5

3.2 Main SLR Sample Summary

  • Included ABSA publications began in 2008 and grew noticeably after 2014, alongside increased extracted and in-scope study counts.Pre-2008 results were excluded because they were not ABSA-focused; the authors suggest the later growth likely reflects the emergence of deep-learning approaches, especially sequence models.Source: §3.2 · p. 5

3.3 Note on “domain” mapping

  • The review distinguishes a study’s research application domain from its dataset domain and manually maps both into 19 categories.Research domain is non-specific unless a study names an application or use case; dataset domain is based on the dataset’s described source or selection context.Source: §3.3 · p. 5

3.4 Phase-2 targeted review on in-context learning

  • The Phase-2 review focuses on in-context-learning implementations of generative models for ABSA, excluding fine-tuned approaches and extending the search through July 2024.The combined pool contained 790 in-scope studies; keyword filtering selected 208 for manual review, including 91 new-search studies and 117 from the earlier SLR.Source: §3.4 · p. 6
  • The Phase-2 search omitted SpringerLink because persistent interface issues prevented navigation of its search results during collection.Source: §3.4 · p. 6

4 Results

The review finds concentrated ABSA research and dataset use, alongside shifts in subtasks and solution approaches. Its results span domain distributions, research paradigms, methods, and emerging generative approaches.

4.1 Results for RQ1

  • 65.32% of reviewed studies had no specific application domain, while 70.95% of study–dataset-domain pairs used product/service review datasets; the review also found domain mismatch and low diversity.Only 12 datasets appeared among combinations used by at least 10 studies, and 78.20% of those combinations involved four SemEval datasets.Source: §4.1 · p. 7, Figure 2 · p. 8

4.1.1 Detailed Results for RQ1

  • 65.32% of studies had a non-specific research domain, while product/service review accounted for 24.28%; healthcare and policy research each had single-digit study counts.Product/service review remained a major category over time, but non-specific studies consistently overtook it from 2015.Source: §4.1.1 · p. 8
  • Among 757 frequent domain–dataset combinations, 95.77% involved non-specific research domains, and four SemEval datasets accounted for 78.20%.The 757 combinations involved only 12 distinct datasets; two SemEval datasets alone accounted for 50.33% of the entries.Source: §4.1.1 · pp. 8–9, Table 3 · p. 10
  • Of 20 cross-domain or domain-agnostic studies, 17 used only product/service review datasets and 14 used at least one SemEval dataset.Only three included other dataset domains: movie reviews, singer reviews, or generic tweets.Source: §4.1.1 · p. 9

4.2 Results for RQ2

  • 32.37% of studies combined aspect extraction and sentiment classification, while 30.83% focused on sentiment classification alone; composite ASTE studies rose to 10 in 2022.The authors suggest the ASTE rise could signal a shift toward multi-task learning methods for end-to-end ABSA.Source: §4.2 · p. 9

4.3.1 Paradigms

  • 66.28% of studies used somewhat-supervised paradigms, including 60.89% fully supervised; 19.65% did not require labelled data, and 14.07% were hybrid.The review classifies non-machine-learning approaches under the unsupervised paradigm.Source: §4.3.1 · p. 10, Table 4 · p. 11

4.4 Approaches

  • Among 519 reviewed studies, 60.31% (N=313) employed deep-learning approaches, which took dominance from 2017; over half, 54.53% (N=283), also included traditional machine learning.Pure linguistic or statistical approaches appeared in 67 studies (12.91% of the 519) and persisted over time.Source: §4.4 · pp. 11, 13, 14, 15

4.5 ICL and generative approach in ABSA - Phase-2 targeted review results

  • Of five 2024 studies testing in-context learning on ABSA, four found zero- and 5-shot foundation models below fine-tuned or fully trained deep-learning models; 18 studies fine-tuned generative LLMs in 2022–2024.Most fine-tuned generative LLM studies framed composite ABSA tasks as sequence-to-sequence generation.Source: §4.5 · pp. 15–16

5 Discussion

The discussion links skewed ABSA domain and dataset resources, dataset-reliant methods, and solution trends to systemic research challenges. It also identifies future directions and limits of the review.

5.1.1 The Out-of-sync Research and Dataset Domains

  • 70.95% (N=447) of 519 studies used product/service-review datasets, while 90.48% (N=656) of studies in the non-specific research domain did so.The review suggests scarce non-commercial datasets may have led generic studies to rely on popular benchmarks, potentially hindering development and evaluation across domains.Source: §5.1.1 · p. 17
  • About half of the education, healthcare, and public-policy datasets were author-created, and the most-researched of these domains had only 12 studies (2.31% out of 519).The review connects the scarcity of public datasets to the cost and challenge of research in these domains.Source: §5.1.1 · p. 18

5.1.2 The Dominance and Limitations of the SemEval Datasets

  • 78.20% (N=592) of 757 study/domain/dataset combinations used four SemEval datasets, which often omit complex multi-aspect, multi-sentiment cases.All 35 models tested on more complex same-source datasets showed performance drops, reaching 69.73% in [89].Source: §5.1.2 · p. 19

5.1.3 The Reliance on Labelled-Datasets

  • 60.89% (N=316) of studies were fully supervised, and 17 of 20 cross-domain or domain-agnostic studies relied only on product/service-review datasets.The review notes that domain knowledge shapes rules, features, and learned representations, and that remote or narrower training domains can cap performance on context-heavy tasks [48].Source: §5.1.3 · p. 19
  • 60.31% (N=313) of studies used deep learning, and RNN-based solutions made up 55.91% (N=175) of those approaches, with sequentiality limiting long-distance relations [74].Traditional attention can struggle with multi-word aspects or multi-aspect sentences.Source: §5.1.3 · p. 19

5.1.4 The Potential of Generative LLMs and Foundation Models

  • Only five of 208 recent studies tested ICL, mostly with subpar results against other methods and simple zero-shot instructions.Fine-tuned generative LLMs using Seq2Seq text generation have seen increasing adoption and promising task performance, but retain costs, labelled-data needs, and domain-transfer challenges [73]; larger, more diverse datasets are also needed for effective fine-tuning or comprehensive testing.Source: §5.1.4 · p. 20

5.2 Ideas for Future Research

  • Future work should broaden ethically shared datasets, develop less data-dependent methods and robustness evaluation, and continue examining end-to-end, hybrid, and generative approaches.The review also calls for attention to domain adaptation, alternative training schemes, and effective feature and knowledge injection.Source: §5.2 · p. 20

5.3 Limitations

  • The review covers English-language studies from four peer-reviewed databases and does not compare solution performance across studies.The authors also note that search, screening, and categorisation may have missed relevant studies, datasets, or approaches.Source: §5.3 · pp. 20–21

6 Conclusion

The conclusion synthesizes ABSA research across application domains, dataset domains, and methodologies, identifying potential systemic issues and areas for future attention.

6 Conclusion

  • The review identifies product/service-review dataset predominance among studies without a specific research domain, alongside dataset-reliant methods such as supervised machine learning.It discusses possible mutual reinforcement between resources and methodologies and proposes areas for further research.Source: §6 · p. 21

A Aspect-Based Sentiment Analysis (ABSA)

The appendix introduces ABSA and its subtasks, then points readers to reviews covering particular methods and task areas.

A.1 Definition and examples

  • ABSA identifies sentiment toward specific entities or attributes called aspects, which may be explicit or implicit and can carry different sentiments within one text.Examples span restaurant, laptop, and course reviews, illustrating explicit and implicit aspects and their associated opinions.Source: §A.1 · p. 23

A.2 ABSA Subtasks

  • ABSA subtasks include extracting aspects and opinions and classifying aspect sentiment; end-to-end models perform multiple subtasks together to capture inter-task relations.Composite formulations include aspect-opinion pairs, aspect-polarity pairs, sentiment triplets, and quadruplets.Source: §A.2 · pp. 23–24

A.3 Other ABSA Reviews

  • Existing reviews cover ABSA subtasks and approaches including explicit and implicit extraction, deep learning, end-to-end tasks, pretrained LLMs, and multimodal ABSA.The review directs readers seeking detailed solution methodologies to those prior surveys.Source: §A.3 · p. 24

B Full SLR Methodology

This section provides a detailed account of the systematic literature review methodology and procedures.

B Full SLR Methodology

B.1 Research identification

The review searched four digital databases in October 2022, using a broad-search, later-filter strategy to identify ABSA studies.

B.1 Research identification

  • The authors exported 4,191 research-paper files from four databases on 24–25 October 2022.The exports included PDF and BibTeX or equivalent files.Source: §B.1 · p. 25
  • Because database search parameters were limited, the authors used a “search broad and filter later” strategy.Table B.1 details the search strings, criteria, and exported files for each database.Source: §B.1 · p. 25
  • The search strategy combined terms for ABSA, extraction, and opinion mining while avoiding false positives from the broader sentiment-analysis field.Search strings were selected based on pilot trials, with keyword-only references and coarser-grained studies filtered later.Source: §B.1 · p. 25

B.2 Selection of primary studies

Automatic PDF screening followed by manual review narrowed the search results to 519 in-scope studies, using criteria refined after a pilot examination.

B.2 Selection of primary studies

  • A pilot examination of 100 files informed refined inclusion and exclusion criteria, including cases where ABSA keywords appeared only in references, appendices, or introductions.Some papers discussed ABSA in reviews or introductions while studying coarser-grained sentiment analysis or opinion mining.Source: §B.2 · p. 26
  • The review excluded popularity criteria such as citation counts to reduce mainstream-method over-dominance from citation chaining.Source: §B.2 · p. 26
  • The screening pipeline used PDF mining for automatic exclusion, followed by manual screening against the inclusion and exclusion criteria.The automatic workflow used Pandas and PyMuPDF to collect file metadata and PDF keyword-occurrence information.Source: §B.2 · p. 26
  • 519 papers remained in scope after manual screening of the 914 files that passed automatic exclusion.The automatic process excluded files using predefined criteria, with several exclusion steps manually checked and others spot-checked.Source: §B.2 · p. 26

B.3 Data extraction and synthesis

The authors manually extracted study and dataset information, standardized and grouped domain descriptions, and quantitatively synthesized the resulting data.

B.3 Data extraction and synthesis

  • For each of the 519 publications, the authors recorded research focus, domains, subtasks, datasets, models, architectures, paradigms, and specific approaches.Recorded approaches included attention mechanisms and Naïve Bayes classifiers.Source: §B.3 · p. 26
  • After cleaning recording errors and naming variations, the authors analyzed the synthesized data quantitatively with Pandas.The mappings aimed to maintain consistency between research-domain and dataset-domain categories.Source: §B.3 · pp. 26–27
  • Research domains defaulted to “non-specific” unless a study stated an application domain or use case as its motivation.Research and dataset domains were mapped into 19 common categories; a study using Yelp and Amazon reviews could therefore have a non-specific research domain but a product/service-review dataset domain.Source: §B.3 · pp. 26–27
  • Dataset domains were assessed per dataset, with names standardized and descriptions checked before categorization.Datasets without a specific domain filter, such as general Twitter tweets, were classified as “non-specific.”Source: §B.3 · pp. 26–27

C Additional Results

The supplementary results catalogue studies by subtasks, methods, datasets, publication type, and dataset language, while the supplied results show a concentration in a few datasets.

C Additional Results

  • Conference papers outnumbered journal articles until 2022, with the gap closing since 2016.The authors note that the late-October 2022 search, journal publication lag, and possible changes in publication channels could contribute to this pattern.Source: Figure C.1 · p. 28
  • Figure C.2 tracks included-study counts by publication year for the top five dataset languages.Source: Figure C.2 · p. 29
  • Supplementary tables report study counts by subtask combinations, individual subtasks, deep-learning approaches, traditional machine-learning approaches, and datasets.The dataset table covers 1,179 dataset uses across 519 studies and 218 datasets.Source: Table C.1 · p. 30, Table C.2 · p. 30, Table C.3 · p. 31
  • 211 studies (17.90%) used SemEval 2014 Restaurant, and 189 (16.03%) used SemEval 2014 Laptop, the two most frequent listed datasets.Source: C Additional Results · p. 32, Table C.5 · p. 32
Loading 2311.10777v6…