PaperlayerView exact revision on arXiv

Source-linked AI summary

Stance Detection on Social Media: State of the Art and Trends

Abeer AlDayel, Walid Magdy

arXiv:2006.03644v5cs.SIcs.CL

TL;DR

Social-media stance detection addresses the challenge of identifying viewpoints when sentiment analysis is insufficient and when stance is shaped by language and social interaction. This paper surveys research across several domains, covering tasks, targets, features, methods, applications, and benchmarks. It concludes that network features often outperform textual features and that SVM-based supervised methods are effective, while existing approaches remain dependent on human-annotated data.

  • Problem

    Social-media stance is often confused with sentiment, although the survey states that sentiment and stance are orthogonal and that sentiment has only a minuscule effect on accurate stance detection.

  • Method

    The paper surveys social-media stance detection across NLP, computational social science, and Web science, covering tasks, targets, features, machine-learning approaches, applications, and trends.

  • Results

    Network features were superior to content features in most surveyed studies, while supervised SVM methods were found most effective across different settings.

  • Takeaways & Limitations

    The survey provides an overview of current approaches and identifies future research directions for stance detection on social media.

  • Takeaways & Limitations

    Existing stance-detection techniques have a huge dependency on human-annotated data, motivating efforts to enrich available data.

Abstract

from arXiv · show

Stance detection on social media is an emerging opinion mining paradigm for various social and political applications in which sentiment analysis may be sub-optimal. There has been a growing research interest for developing effective methods for stance detection methods varying among multiple communities including natural language processing, web science, and social computing. This paper surveys the work on stance detection within those communities and situates its usage within current opinion mining techniques in social media. It presents an exhaustive review of stance detection techniques on social media, including the task definition, different types of targets in stance detection, features set used, and various machine learning approaches applied. The survey reports state-of-the-art results on the existing benchmark datasets on stance detection, and discusses the most effective approaches. In addition, this study explores the emerging trends and different applications of stance detection on social media. The study concludes by discussing the gaps in the current existing research and highlights the possible future directions for stance detection on social media.

1 INTRODUCTION

Stance detection on social media has grown as a research area for analyzing views on controversial social and political topics. This survey maps its development, task variants, methods, applications, and future directions across multiple research communities.

  • Social media stance detection examines perspectives toward controversial social issues and political topics, including abortion, climate change, feminism, referendums, and elections.
  • Stance work distinguishes detecting expressed views in posts from predicting unexpressed views about undisclosed topics or future events.
  • Target settings include one topic, multiple related targets such as Clinton and Trump, or claims whose support is analyzed for applications including fake-news detection.
  • The survey spans NLP, computational social science, and Web science, comparing stance with sentiment and covering text, network, and contextual features.
  • It reviews social-media applications, emerging trends, research gaps, and potential future directions for the field.

2 DEFINITIONS OF STANCE AND STANCE DETECTION

Stance is a speaker’s standpoint or judgment toward a proposition, shaped by evaluation, self-positioning, and alignment with other social actors. Detecting it is difficult because textual stance may be implicit and influenced by personal, cultural, and social factors.

  • Stance is the expression of a speaker’s attitude, standpoint, or judgment toward a proposition.
  • The field remains in its infancy, with unclear roles for language and social interaction in inferring users’ stances.
  • Stance-taking is affected by experiential behavior, previous knowledge, personal opinions, cultural norms, and social aspects.
  • The stance-taking process involves evaluating objects, positioning the self, and aligning with other social actors.
  • Social-media posts link the stance subject to the user, while linguistic features include adjectives, adverbs, and lexical items.
  • A stance is often not transparent in text and may require inference from interaction and historical context.

3 STANCE VS. SENTIMENT

Stance detection and sentiment analysis address related but distinct aspects of social-media opinion. The survey defines their task differences and shows that sentiment polarity is insufficient for reliably inferring stance.

  • Task definitions: Sentiment analysis determines textual emotional polarity, often without requiring a predefined target.Its usual outputs are positive, negative, and neutral labels.
  • Task definitions: Stance detection identifies a person’s standpoint toward a specified target, typically as favor, against, or none.Unlike sentiment analysis, it requires a clear target of evaluation.
  • Task definitions: Stance detection may use text, the social actor, or both, including behavioral, network, profile, and contextual features.Text-only approaches can frame stance detection as textual entailment, while broader approaches incorporate user information.
  • Qualitative differences: Positive sentiment can accompany an opposing stance, neutral sentiment can accompany an opposing stance, and negative sentiment can occur without a supporting stance.Examples include abortion legalization and Hillary Clinton, demonstrating that sentiment and stance are not always aligned.

4 EXPRESSED VS UNEXPRESSED STANCE DETECTION

The literature distinguishes stance classification, which detects expressed views, from stance prediction, which infers views before explicit expression or events. Prediction operates at micro and macro levels using textual, behavioral, network, and aggregation-based approaches.

  • Expressed versus unexpressed stance: Stance classification detects expressed positions in text, whereas stance prediction infers stances before an event or without explicit expression.Classification represents most existing work, while prediction addresses pre-event or unexpressed attitudes.
  • Data collection: Stance datasets often collect event-related tweets with target-specific keywords, hashtags, and predefined annotation guidelines.The SemEval dataset used favor, against, and ambiguous hashtags for topics such as Hillary Clinton.
  • Expressed stance detection: Expressed-stance studies commonly model whether text entails a favor or against relation to a target.Some studies augment textual models with signals from users’ social networks.
  • Prediction levels: Micro-level prediction estimates an individual user’s future stance from prior history, while macro-level prediction aggregates individual predictions into public opinion.The micro-level formulation is compared with recommending new items from a user’s purchase history.
  • Prediction methods: Prediction methods use user interactions, voting patterns, heterogeneous links, historical tweets, latent group beliefs, collaborative filtering, and matrix factorization.These approaches can target users, topics, ideology, or events not yet encountered by the user.
  • Prediction scope: Macro-level stance prediction has received relatively little study and has often been customized to specific cases such as Islamophobia or debate forums.Some work predicts post-event stances from prior tweets and user interactions, while other work addresses cold-start users.

5 STANCE DETECTION ACCORDING TO TARGET

Stance detection is categorized by target into single defined targets, multiple related targets, and claim-based targets. These settings differ in whether models predict orientation toward one target, jointly model related targets, or classify responses to claims.

  • Stance detection targets are grouped into single defined targets, multi-related targets, and claim-based targets.
  • Single defined targets: Single-target detection predicts stance toward one predefined target from text or user information.A separate classification model is generally built for each target, unlike target-independent sentiment analysis.
  • Multi-related targets: Multi-target detection jointly learns users’ orientations toward two or more related targets by exploiting information shared among them.For example, support for Hillary Clinton may imply opposition to Donald Trump.
  • Multi-related targets: A dataset of 3,450 tweets was annotated for stance toward both Trump and Clinton, while another study introduced 4,455 tweets concerning the 2016 US elections.
  • Multi-related targets: A memory-based model jointly modeling multiple targets achieved the current state-of-the-art result on the multi-target benchmark dataset.
  • Claim-based targets: Claim-based stance detection classifies whether text or conversation supports or denies a claim, such as a rumor post or news headline.The claim is the main input, while associated text may be an article body or replies.

6 STANCE MODELING ON SOCIAL MEDIA

Social-media stance models use content signals and network signals to represent users and their stances. Across the surveyed comparisons, network-based representations generally outperform textual modeling, although effective feature choices vary by setting.

  • Stance models use two broad signal types: content signals from tweets and network signals from users’ connections and interactions.Figure 5 summarizes features in these categories, followed by comparisons across multiple datasets.
  • Content features: Content representations may collect topic-related keywords or unrelated content to model user behavior rather than only topic-level stance.
  • Content features: Textual stance features include n-grams, sentiment polarity, and latent semantics such as topic representations.
  • Content features: Word and character n-grams obtained the best F-score among participating systems in the SemEval-2016 stance detection task.
  • Content features: Sentiment alone was generally insufficient for predicting stance, because stance and sentiment do not necessarily align.
  • Network features: Network features include interaction, preference, and connection networks, with interaction networks producing the best performance among the three.
  • Comparative performance: Combining network and content features with a simple linear SVM was more effective than using word embeddings with RNNs and LSTMs.
  • Comparative performance: Comparative studies report that network modeling outperformed textual modeling across multiple datasets and stance settings.The surveyed results also describe network features as overcoming limitations of textual-entailment-based modeling.

7 STANCE DETECTION ALGORITHMS

Stance detection algorithms span supervised, weakly supervised or transfer-learning, and unsupervised approaches. The surveyed evidence favors supervised and network-informed models in several benchmarks, while unsupervised clustering also shows strong results in selected datasets.

  • Stance detection algorithms are divided into supervised, weakly supervised or transfer-learning, and unsupervised approaches.
  • Supervised learning: Supervised learning is the most common approach and uses predefined stance labels such as In-Favor, Against, and None.SemEval-2016 contains these three labels across five topics.
  • Supervised learning: Reported supervised models range from classical NB, SVM, and decision trees to neural RNN and LSTM systems.
  • Supervised learning: 69.0% F1 for content-only n-grams exceeded the approximately 66.8% F1 obtained when sentiment was used as a feature.
  • Supervised learning: 72.3% was the reported overall F score for a multitask bidirectional gated recurrent unit model using tweet sentiment.
  • Weakly supervised and transfer learning: Transfer-learning methods leverage labelled data from one target or task and unlabelled or related data from another target.SemEval Task B includes 78,000 unlabelled tweets related to Trump and motivated transfer-learning research.
  • Weakly supervised and transfer learning: Target-oriented distant supervision was not generalized to unseen targets such as Macron and Putin, partly because its labelling heuristics may vary across time periods.
  • Unsupervised learning: Unsupervised models use clustering over user and topic representations, often incorporating content and interaction networks.

8 STANCE DETECTION APPLICATIONS

Stance detection is applied to public-opinion analysis, political and social events, platform polarization, and news-veracity tasks. The survey also describes how stance signals can expose controversy and support misinformation assessment.

  • Stance detection measures attitudes toward entities and events for public-opinion analysis across political and social topics.
  • Analytical studies: Applications include immigration, political elections and debates, Brexit, Catalan independence, mass shootings, and reactions to disruptive events.
  • Analytical studies: In one immigration study, about 5% of users were against immigration and 23% supported it.
  • Analytical studies: Image-based stance analysis has been used to distinguish ideological leanings, with liberal images linked to economic equality and conservative images to state and economic power.
  • Social-media phenomena: Stance detection supports platform-level analysis by exposing users to contradicting views and measuring controversy.
  • Social-media phenomena: It has also been used to study echo chambers and homophily, where like-minded connections amplify aligned beliefs and reinforce users’ view biases.
  • Veracity checking: For news-veracity tasks, comments confirming or denying a news claim provide stance evidence used to distinguish rumors from authentic news.
  • Veracity checking: Short, context-limited social-media content makes truthfulness difficult to estimate, motivating stance-based approaches to fake-news and rumor verification.

9 STANCE DETECTION RESOURCES

The survey organizes stance-detection resources by task type, target structure, data source, and annotation framework. It reviews target-specific, multi-related-target, claim-based, and stance-prediction datasets, while noting limited benchmarks for prediction tasks.

  • Dataset taxonomy: Available resources are organized into stance classification and stance prediction datasets, with classification further divided by target structure.The taxonomy includes target-specific, multi-related-target, and claim-based datasets.
  • Data sources: Existing data sources include Twitter, Wikipedia, debate websites, news sites, and multilingual comment collections.Some resources target cross-lingual stance detection using German, French, and Italian comments.
  • Target-specific datasets: Target-specific resources include SemEval datasets supporting supervised and weakly supervised frameworks.The supervised data annotate tweet-target pairs for stance and sentiment, while the weakly supervised data concern Donald Trump and include 78,000 unlabeled tweets.
  • Claim-based datasets: Claim-based datasets evaluate the stance of information sources toward claims, including rumor-supporting or rumor-denying tweets.The Rumours dataset contains 10,417 tweets, while Emergent and Fake-News datasets support fact-checking or claim-detection settings.
  • Multi-related-target datasets: Multi-related-target datasets annotate stances toward multiple candidates within the same tweets.Examples include Trump versus Hillary and three candidate pairs in the Multi-targets dataset.
  • Stance prediction datasets: Stance-prediction benchmarks remain scarce, leading researchers to construct their own datasets for predicting stance before events.Prediction studies use chronological pre-event and post-event data and may draw on Twitter and other sources.

10 DISCUSSION

The discussion distinguishes stance from sentiment, reviews textual, supervised, transfer, unsupervised, and network-based approaches, and identifies benchmark, domain, efficiency, and privacy gaps. It reports that network features often outperform textual features but require additional information and computation.

  • Stance and sentiment: Sentiment and stance are orthogonal, so sentiment polarity should not be treated as a direct indicator of support or opposition.The survey identifies a misconception in studies that use negative or positive sentiment as a linear proxy for stance.
  • Learning approaches: Stance detection commonly uses supervised classification, but dependence on human-annotated data creates scarcity concerns.Transfer learning, unsupervised learning, and distant supervision address annotation scarcity, although they receive less attention than supervised methods.
  • Network information: Network features generally improve stance-detection performance compared with using post content alone.These models emphasize users’ online behavior and social attributes, while textual-entailment modeling is described as sub-optimal relative to network representations.
  • Network information: Network-based models are more computationally expensive because they require collecting additional user information.The discussion calls for models that retain network information while becoming more effective and efficient.
  • Benchmark gaps: The SemEval 2016 benchmark has limited training data, is aging, and can lose tweets needed for retrieving user-network information.The discussion reports an average of about 500 records per topic in training and calls for a larger benchmark.
  • Future directions: Future resources should expand beyond predominantly social, political, and religious topics and include additional domains and languages.The survey also highlights privacy measures for social-media users as an important consideration.

11 CONCLUSION

The conclusion synthesizes stance-detection tasks, targets, features, methods, applications, datasets, and research directions. It reports that network features are usually superior to textual features, SVM-based supervised methods are highly effective, and transfer and unsupervised learning remain promising.

  • Scope of the survey: The survey distinguishes stance detection from sentiment analysis and covers target-specific, multi-related-target, and claim-based tasks.It also separates expressed stance detection from prediction of unexpressed stance toward future events.
  • Features: Network features outperform content features in most reviewed stance-detection studies.The conclusion presents this comparison as a central finding across the surveyed literature.
  • Methods: SVM-based supervised methods are reported as the most effective across different datasets.This conclusion follows the survey’s comparison of machine-learning approaches.
  • Emerging methods: Transfer-learning and unsupervised-learning attempts for stance detection have produced promising results.The conclusion identifies these approaches as continuing directions rather than established replacements for supervised methods.
  • Applications and directions: The survey reviews applications including fake-news detection and summarizes datasets and effective approaches.It also identifies new datasets and combined NLP and social-computing approaches as future priorities.

A APPENDIX A: STANCE DETECTION WORK

The appendix catalogs recent stance-detection and stance-prediction studies, their datasets, features, and algorithms. It presents separate work lists for prediction and classification.

  • Appendix scope: The appendix provides a list of recent stance-prediction and stance-detection work on social media.The accompanying tables separately identify work in stance prediction and stance classification.
  • Tasks and datasets: The listed studies cover target-specific, candidate-focused, claim-based, and breaking-news settings.Examples include SemEval targets, Hillary Clinton and Donald Trump, Wikipedia claims, and breaking-news data.
  • Features: Reviewed studies use textual, sentiment, linguistic, structural, contextual, and network-related features.Examples include n-grams, part-of-speech tags, word embeddings, hashtags, replies, user attributes, and engagement measures.
Loading 2006.03644v5…