Introduction

Stance detection is the study of how computational systems infer a writer’s or speaker’s position toward a target. The target may be a person, policy, event, scientific claim, product, political candidate, social movement, or any other object of evaluation. A short text such as “This proposal will hurt small clinics” may express opposition toward a healthcare proposal; “Finally, a policy that takes rural patients seriously” may express support; “What exactly would this change?” may express inquiry rather than support or opposition. The basic problem sounds simple: given a text and a target, decide whether the author is in favor of, against, neutral toward, or unrelated to that target. Yet almost every word in that sentence hides a serious research difficulty.

The first difficulty is that stance is target-dependent. A text can support one target while opposing another. Consider:

“I support stricter emissions rules because the current policy protects polluters.”

Toward stricter emissions rules, the stance is favorable. Toward the current policy, the stance is unfavorable. The sentiment of the sentence is mixed, but the stance becomes clear only after we know the target. This distinction is one reason stance detection developed as a task separate from general sentiment analysis. Sentiment analysis usually asks whether a text expresses positive, negative, or neutral affect, while stance detection asks what position the author takes toward a specified target. The SemEval-2016 Task 6 formulation made this distinction explicit by defining stance detection in tweets as determining whether the author is in favor of, against, or neither in relation to a given target, including cases where the target is not explicitly mentioned in the tweet (Mohammad et al., 2016).

The second difficulty is that stance is often indirect. People do not always say “I support X” or “I oppose Y.” They quote others, ask rhetorical questions, use irony, rely on shared background knowledge, or imply a position through evidence. For example:

Target: banning single-use plastic bags
Text: “Because what the ocean really needs is more plastic.”

A literal reading of the sentence does not contain the phrase “I oppose plastic bags.” Nevertheless, many readers would infer that the author supports restrictions on plastic pollution. This inference depends on irony, world knowledge, and pragmatic reasoning. In linguistics, stance is not merely a label attached to a sentence; it is a social and communicative act through which speakers evaluate objects, position themselves, and align or disalign with others. Du Bois’s “stance triangle” describes stancetaking as a relation among a subject who evaluates, an object being evaluated, and other subjects with whom alignment is negotiated (Du Bois, 2007). This view is important for NLP because many stance signals are not contained in a single obvious keyword.

The third difficulty is that stance is context-sensitive. A reply such as “Exactly” may be strongly supportive or strongly opposing depending on the post it answers. A quote post may repeat a claim in order to mock it. A hashtag may indicate affiliation in one community but sarcasm in another. Even an apparently simple phrase such as “do your research” can function differently across topics and discourse communities. For this reason, stance detection cannot be reduced to isolated sentence classification. It often requires modeling a target, a conversation, a platform, a user history, a discourse community, and a time period.

This book is about how to build, evaluate, and reason about such systems at a graduate level.

What this book means by stance

We begin from first principles. A text is a linguistic artifact: a tweet, comment, paragraph, news sentence, forum post, transcript segment, or document. An author is the person, institution, or account to whom the text is attributed. A target is the entity, proposition, event, or issue toward which the author may take a position. A stance label is a category assigned by a task definition, such as favor, against, neutral, or none.

A stance detection problem can therefore be expressed as a mapping:

given a text and a target, infer the author’s position toward that target.

For example:

The last example shows why stance detection is not a trivial keyword task. The word “great” is positive in form, but the sentence may be sarcastic. The text is also not simply about “electric vehicles” as machines; it is about affordability, policy, and social distribution. A careful system must represent more than surface sentiment.

In this book, we treat stance detection as a family of related computational tasks rather than a single fixed benchmark. Some datasets use three labels: favor, against, and none. Others distinguish support, deny, query, and comment, especially in rumor or claim-centered settings. Some tasks classify a short post toward a named target; others infer the relation between a document and a claim; still others estimate a user’s stance from many posts over time. Surveys of the field emphasize this diversity of task formulations, domains, targets, and modeling approaches (Küçük and Can, 2020).

A useful graduate-level habit is to ask four questions whenever you see a stance detection task:

1. What exactly is the target?
   Is it an entity, a proposition, a policy, a claim, or a broad topic?

2. Whose stance is being inferred?
   Is it the writer of one post, the author of an article, a social media account, or a community?

3. What label space is being used?
   Are labels binary, ternary, ordinal, multilabel, or structured?

4. What evidence is allowed?
   Is the model given only the text, or also the target, conversation history, user information, retrieved evidence, or social graph context?

These questions will return throughout the book because they determine what counts as a correct model, a valid dataset, and a meaningful evaluation.

Why stance detection matters

Stance detection matters because many social, scientific, and political questions are not only about what people mention, but how they position themselves toward what they mention. A public health agency may need to distinguish questions about a vaccine from rejection of vaccination. A journalist may want to map how different actors respond to a policy proposal. A researcher may study polarization by examining how groups align around claims. A content moderation team may need to prioritize posts that endorse harmful instructions rather than merely discuss them.

In computational social science, stance detection can help analyze public discourse at scales that manual reading alone cannot handle. In argument mining, it can help identify whether a text supports or attacks a claim. In misinformation research, it can help distinguish spreading, denying, questioning, and commenting on a rumor. In product and policy analysis, it can separate satisfaction, dissatisfaction, skepticism, and conditional support. The same technical machinery can be useful across these settings, but the ethical and methodological risks differ sharply.

For example, suppose a model is used to estimate public support for a proposed law from social media posts. Several problems arise immediately. Social media users are not a representative sample of the population. Some posts are jokes, quotes, or coordinated messages. A stance label assigned to a single post may not represent the user’s stable belief. The model may perform worse for dialects, languages, or communities underrepresented in training data. Thus, a stance detection system can be technically impressive and still scientifically misleading if the data-generating process and evaluation design are weak.

This book therefore treats stance detection not only as a modeling problem, but also as a measurement problem. A measurement problem asks whether our observable data and labels validly capture the concept we claim to study. In stance detection, the concept is an author’s position toward a target; the observable evidence is text and context; the label is an operational decision made by annotators or by a dataset designer. If these pieces do not align, high accuracy can give a false sense of success.

Stance, sentiment, opinion, argument, and factuality

Because stance detection overlaps with several neighboring tasks, we need a careful first orientation.

Sentiment analysis usually concerns affective polarity: whether language is positive, negative, or neutral. If someone writes “I hate this policy,” sentiment and stance point in similar directions. But if someone writes “I hate that people are attacking this policy,” the negative sentiment is directed toward the attackers, not necessarily the policy. Stance requires knowing the target of the position.

Opinion mining is broader than sentiment analysis. It may include extracting opinion holders, opinion expressions, targets, and polarities. Stance detection can be viewed as related to opinion mining, but it often focuses on controversial claims, debates, or target-specific positions rather than product-style attributes.

Argument mining studies argumentative structure: claims, premises, evidence, support relations, and attack relations. Stance detection may use argument structure as evidence. For example, if a text gives a premise supporting a claim, the author may also support that claim. But the tasks are not identical. A text can have a stance without a fully developed argument, as in “Absolutely not,” and a text can describe an argument without endorsing it.

Factuality concerns whether an event or proposition is presented as real, possible, uncertain, negated, or hypothetical. In “The committee might approve the bill,” the author expresses uncertainty about an event. This is not the same as supporting or opposing the bill. Factuality, modality, and evidentiality can nevertheless influence stance. The linguistic study of stance has long emphasized lexical and grammatical markers such as evidentiality, affect, and evaluation, which help speakers communicate commitment and attitude (Biber and Finegan, 1989).

Misinformation detection asks whether content is false, misleading, or harmful according to some standard of evidence. Stance detection may help misinformation research by identifying whether a post endorses, rejects, questions, or merely reports a claim. But stance detection does not by itself determine truth. A person can support a true claim, oppose a true claim, support a false claim, or oppose a false claim.

These distinctions matter in research design. If a dataset labels negative sentiment as opposition without checking the target, it may not actually measure stance. If a misinformation system treats every mention of a false claim as endorsement, it may penalize debunking. If an argument mining system assumes that quoted evidence is endorsed by the author, it may misread critical quotation. Graduate-level work in stance detection begins by refusing these shortcuts.

Why the task is hard for machines

A modern NLP model receives text as input and produces a prediction. Under the hood, a neural model represents words or subword units as vectors, combines them through learned transformations, and estimates probabilities over labels. Transformer models, introduced through the self-attention architecture of Vaswani et al. (2017), and later pretrained language models such as BERT (Devlin et al., 2019), have greatly influenced how NLP systems represent context. These models can learn rich patterns from large corpora and can be fine-tuned for stance detection tasks.

But strong representation learning does not remove the conceptual difficulty of stance. A model may exploit spurious cues: hashtags, named entities, topic words, or annotation artifacts. If most training posts mentioning “climate hoax” are labeled against climate action, the model may learn a useful pattern for one dataset but fail on a community that uses the phrase to criticize denialism. If a benchmark splits data randomly, near-duplicate posts or target-specific lexical cues may leak across train and test sets. If annotators disagree because the stance is genuinely ambiguous, a single “gold” label may hide uncertainty.

This leads to a central theme of the book: stance detection requires both modeling skill and experimental discipline. A good stance system is not merely one that achieves a high leaderboard score. It should be clear about the task definition, honest about uncertainty, evaluated under meaningful generalization conditions, and documented well enough that another researcher or engineer can understand its limits.

Consider the following example:

Target: “remote work”
Text: “My manager says productivity is ‘better than ever’ while scheduling three extra meetings a day.”

A surface model might notice “better than ever” and predict favorable stance. A more context-sensitive model might notice quotation marks and the complaint about meetings, and infer a negative or skeptical stance toward how remote work is being managed. But even this interpretation depends on the target. Is the stance toward remote work itself, toward the manager’s claim, toward meeting culture, or toward company policy? The example illustrates why annotation guidelines must define the target carefully and why models must be evaluated against those definitions.

The path through the book

The book begins with the problem itself. Chapter 1 defines stance detection in relation to neighboring NLP tasks and introduces major use cases and task formulations. The goal is to give you a precise vocabulary before introducing more technical machinery.

Chapter 2 builds the linguistic and pragmatic foundation. We examine deixis, modality, hedging, presupposition, negation, quotation, implicature, irony, and discourse context. These phenomena explain why stance is often indirect. For example, “I’m not saying the policy is corrupt, but…” uses denial and implication in a way that a literal keyword system may mishandle.

Chapter 3 reviews the mathematical foundations needed for modern NLP models: probability, optimization, vector spaces, information theory, gradients, regularization, calibration, and uncertainty. These tools are not decorative. They explain how classifiers estimate label probabilities, how training objectives work, and why a system can be accurate but poorly calibrated.

Chapters 4 through 6 move from data design to classical baselines. You will learn how to define targets, write annotation guidelines, measure inter-annotator agreement, handle ambiguity, and implement strong non-neural systems. Classical models such as logistic regression and support vector machines remain valuable because they are interpretable, efficient, and often difficult to beat on small datasets when engineered carefully.

Chapters 7 through 9 introduce neural representations, transformers, and target-aware modeling. We study word embeddings, contextual encoders, attention, pretrained language models, cross-encoders, dual encoders, and conversation-aware architectures. A recurring question will be: how does the model represent the relation between text and target?

Chapter 10 focuses on evaluation. We study accuracy, macro-F1, class-wise F1, AUC, calibration, significance testing, confidence intervals, cross-validation, out-of-domain testing, and error analysis. Stance detection is especially vulnerable to misleading results because topics, targets, and lexical cues can leak across splits.

Chapters 11 and 12 address generalization and data scarcity. Real systems often face new targets, new communities, new time periods, and limited labels. We examine transfer learning, domain adaptation, contrastive learning, weak supervision, self-training, active learning, and synthetic data generation.

Chapter 13 turns to large language models. Prompting and instruction-following models can be useful for zero-shot or few-shot stance detection, but they introduce prompt sensitivity, reproducibility concerns, cost constraints, and risks of hallucinated rationales. We will treat these models as powerful tools that still require careful evaluation.

Chapters 14 and 15 expand the unit of analysis. We connect stance to argumentation, evidence, natural language inference, conversation structure, social networks, user histories, and temporal dynamics. Many real stance signals appear not in one sentence, but across interactions.

Chapter 16 studies multilingual and cross-lingual stance detection. We consider multilingual representations, translation pipelines, code-switching, language-specific pragmatics, and cultural variation. A stance system trained primarily on English political data may not transfer reliably to another language or discourse tradition.

Chapter 17 examines robustness, bias, fairness, and ethics. Stance detection can be used for beneficial analysis, but it can also support surveillance, manipulation, or unfair moderation. We will discuss demographic bias, dialect variation, platform bias, annotation bias, and governance questions.

Chapter 18 covers interpretability and diagnostics. You will learn what saliency maps, attention analysis, counterfactual examples, influence functions, probing, calibration plots, and confusion analysis can and cannot tell us. Interpretability methods are useful, but they do not automatically reveal a model’s “reasoning.”

Chapter 19 guides you through a research-grade stance detection system from dataset construction to transformer fine-tuning, hyperparameter search, evaluation, error analysis, documentation, and reproducible experiment tracking. Chapter 20 then turns to deployment: latency, compression, monitoring, drift, human review, privacy, audit trails, continual learning, and open research problems.

The conclusion synthesizes the path and suggests directions for deeper study.

What you should be able to do by the end

By the end of this book, you should be able to read a stance detection paper critically, reproduce a baseline, design a dataset, fine-tune a neural model, evaluate it under rigorous conditions, diagnose its errors, and explain its limitations. You should also be able to recognize when stance detection is the wrong framing or when a proposed application is ethically dangerous.

More concretely, you should be able to look at an example like this:

Target: “expanding nuclear energy”
Text: “If we were serious about decarbonization, we would stop pretending solar panels alone can carry the grid.”

and ask the right questions. Is the target an energy policy, a technology, or a proposition? Does the text explicitly support nuclear energy, or only criticize an alternative? What background knowledge is required? How should annotators label it? What would count as model evidence? Would a model trained on climate activism discourse interpret it differently from a model trained on energy policy hearings? How should uncertainty be represented?

Those questions are not obstacles to progress. They are the beginning of good research.

Stance detection sits at a productive intersection of linguistics, machine learning, social computing, argumentation, and responsible AI. It is technically demanding because language is indirect and context-dependent. It is scientifically demanding because labels must measure the intended construct. It is ethically demanding because predictions about people’s positions can affect reputation, governance, and power. This book will proceed carefully through all three dimensions.

References

Biber, Douglas, and Edward Finegan. 1989. “Styles of Stance in English: Lexical and Grammatical Marking of Evidentiality and Affect.” Text 9 (1): 93–124.

Devlin, Jacob, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. “BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding.” In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, 4171–4186. Association for Computational Linguistics.

Du Bois, John W. 2007. “The Stance Triangle.” In Stancetaking in Discourse: Subjectivity, Evaluation, Interaction, edited by Robert Englebretson, 139–182. Amsterdam: John Benjamins.

Küçük, Dilek, and Fazli Can. 2020. “Stance Detection: A Survey.” ACM Computing Surveys 53 (1), Article 12.

Mohammad, Saif M., Svetlana Kiritchenko, Parinaz Sobhani, Xiaodan Zhu, and Colin Cherry. 2016. “SemEval-2016 Task 6: Detecting Stance in Tweets.” In Proceedings of the 10th International Workshop on Semantic Evaluation (SemEval-2016), 31–41. Association for Computational Linguistics.

Vaswani, Ashish, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. “Attention Is All You Need.” In Advances in Neural Information Processing Systems 30.