---
kind: concept
slug: stance-detection
title: Stance detection
summary: >-
  Classify whether a text favours, opposes or ignores a target or claim, then use that to flag false content: as evidence stance in fact-checking or as crowd and media stance in rumour pipelines. One survey is the only source here; its scores come from other papers and are not comparable. Trade-off: label mismatch and small data against evidence access and accuracy.
tags: [misinformation, fact-checking, rumours, nlp, text-classification]
papers: ["2103.00242"]
related: [automated-fact-checking, multi-task-learning, sentiment-analysis]
status: draft
updated_at: 2026-11-12
---

## Overview

Stance detection asks whether a text agrees with a target, disagrees with it,
or takes neither side. The target may be a claim, a headline or a topic, and
it may be implied rather than stated. When the target is implicit the task
looks much like sentiment analysis [arxiv:2103.00242 §3], so keep the two
apart when you pick a dataset [arxiv:2103.00242 §2].

In mis- and disinformation work, stance becomes a veracity signal in one of
two ways. Either the stance of the evidence towards a claim decides whether
the claim holds [arxiv:2103.00242 §3.1], or the stance of the crowd and the
media towards a rumour feeds a decision about it [arxiv:2103.00242 §3.2].

The invariant all variants share is a classification of a (text, target)
pair into one label from a fixed set. The label sets differ by dataset:
Favor, Against and Neither in the general definition; positive, negative,
discuss and unrelated for FNC-1; SUPPORTED, REFUTED and NOT ENOUGH INFO for
FEVER; Support, Deny, Query and Comment for rumours [arxiv:2103.00242 §2]
[arxiv:2103.00242 §3.1] [arxiv:2103.00242 §3.2]. Merge labels only with an
explicit mapping.

## Variants

The variants below all come from one survey, the [Hardalov et al. survey](/papers/2103.00242).
The survey runs
no experiments of its own, so each variant is described by the results it
reports. The underlying papers have no cards on this wiki, so their content
is not used here.

### Stance classification on a labelled dataset

Given a pair such as a headline and its article body, predict the label. The
survey's main example is FNC-1, with 75K instances [arxiv:2103.00242 tab. 1].
Early systems used hand-made features: words, embeddings and sentiment lexica.
In one analysis, lexical features mattered most and sentiment did not help
[arxiv:2103.00242 §4]. Pre-trained Transformers changed this. On FNC-1,
RoBERTa raised Disagree from 18 to 58 F1 and Agree from 50 to 70 F1
[arxiv:2103.00242 §4]. Card: [Hardalov et al. survey](/papers/2103.00242).

### Fact-checking as stance (evidence retrieval and reasoning)

Retrieve evidence, then judge the claim against it. In FEVER, the evidence is
Wikipedia sentences [arxiv:2103.00242 tab. 1]. The survey lists these top
FEVER scores: 64.2 for a neural semantic matching network, then 67.1 and about
70 for graph neural networks over the evidence, with a better retriever and
graph propagation [arxiv:2103.00242 §4]. The survey notes that FEVER-style
checking differs from stance detection: it scores evidence against a claim,
not an author's stance [arxiv:2103.00242 §3.1]. Card: [Hardalov et al. survey](/papers/2103.00242).

### Fact-checking without retrieval

A pre-trained language model scores the claim directly, through masked-token
scoring or through the perplexity of the claim given the document. There is
no retrieval step, but the approach is prone to pattern biases and performs
below supervised approaches. Few-shot training helps [arxiv:2103.00242 §4].

### Threaded stance in rumour pipelines

For replies in a conversation, the survey covers tree-structured models over
threads, branch-wise LSTMs, and Tree and Hierarchical Transformers for rumour
debunking. Pre-training on stance data, and per-class pre-training data built
from related tasks, improved RumourEval results [arxiv:2103.00242 §4]. Position
in the thread and timing carry signal: denying posts appear later, and the
share of querying posts rises over time while their absolute number falls
[arxiv:2103.00242 §4].

### Multi-dataset and cross-domain learning

The survey covers multi-task learning over several stance datasets, and
cross-domain learning over 16 stance datasets with a domain-adaptation
architecture named MoLE. It also reports pattern-exploiting training across
12 languages [arxiv:2103.00242 §4] [arxiv:2103.00242 §5]. Combining datasets
with different label sets needs a way to relate the labels, such as label
embeddings [arxiv:2103.00242 §5]. See [multi-task learning](/concepts/multi-task-learning).

## When to use which

The survey gives no decision rules. The table reads its reported results
against common constraints and is guidance, not a tested recommendation.

| Situation | Use | Why |
|---|---|---|
| Labelled pairs in one label set, and fine-tuning is feasible | Pre-trained Transformer classifier | Large gains on FNC-1; the learning rate was the most important hyperparameter [arxiv:2103.00242 §4] |
| Labelled pairs, but few or skewed examples | Sampling, class weighting, auxiliary-task data, or several datasets | The survey's listed mitigations for small, imbalanced datasets [arxiv:2103.00242 §5] |
| Claim to check against a document collection | Retrieval, then a stance or inference classifier | Top FEVER-score systems use this pipeline [arxiv:2103.00242 §4] |
| No evidence index available | Language-model checking without retrieval | Needs no retrieval step, but below supervised approaches and prone to pattern biases [arxiv:2103.00242 §4] |
| Replies in a conversation thread | Thread-aware models | Thread depth and timing carry signal [arxiv:2103.00242 §4] |
| Few labelled examples in a language | Pattern-exploiting training, or multi-dataset learning | Reported useful in low- and full-resource settings [arxiv:2103.00242 §4] [arxiv:2103.00242 §5] |
| Need a reason for each decision | Partial explanations only | Memory networks, attention and topics give partial explanations; attention weights may mislead [arxiv:2103.00242 §5] |

## Implementation sketch

The survey gives no code. The sketch shows the two shapes described above.
The aggregation step is not stated in the survey.

```pseudocode
# Stance of one (text, target) pair: one label from a fixed set  [§2]
# Input:  text, target (the target may be implicit)  [§3]
# Output: a label from the dataset's own label set  [§2]

def stance(text, target, model):
    # Encode the pair; the survey does not fix the encoder  [§4]
    features = encode(text, target)
    # Return the most likely label from the dataset's label set  [§2]
    return argmax(model.predict(features))
```

```pseudocode
# Fact-checking as stance, formulation (a)  [§3.1]
# Input:  claim (text); evidence (list of texts)
# Output: veracity label

def verify_claim(claim, evidence, model):
    # Stance of each evidence text towards the claim  [§3.1]
    stances = [stance(claim, e, model) for e in evidence]
    # How the stances become one verdict is not stated  [§3.1]
    return aggregate(stances)
```

## Contradictions

- **Arabic FC and AraStance.** The text gives 52 F1 for AraStance and 78 F1
  for Arabic FC [arxiv:2103.00242 §4]. The table gives 52 to Arabic FC and 78
  to AraStance [arxiv:2103.00242 tab. 2]. The survey does not resolve this.
  Check the AraStance paper before you quote either number.
- **RumourEval '19.** The text gives 57.9 F1 for the best stance result
  [arxiv:2103.00242 §4]. The table gives 61.9 for the same paper and dataset
  [arxiv:2103.00242 tab. 2]. No explanation is given.
- **Comparability of the table.** The scores use different datasets, metrics
  and splits, and the survey says they cannot be compared directly
  [arxiv:2103.00242 §4] [arxiv:2103.00242 tab. 2]. Do not rank methods by
  that table.
- **Leaderboard row.** The FEVER row comes from a public leaderboard, not
  from a paper [arxiv:2103.00242 tab. 2].

## Papers

- [Hardalov et al. survey](/papers/2103.00242): maps stance formulations for false-content detection, tabulates 15 datasets and the methods tried on them, and lists open issues.
