Worksheet: Stance Detection

Background

Stance detection identifies whether a text expresses a favorable, opposing, or neutral position toward a specific target (topic, entity, or claim).

Stance: The attitude or position expressed in text toward a specified target.

Labels: FAVOR (supports), AGAINST (opposes), NEUTRAL (no clear position)

Why Stance ≠ Sentiment

Sentiment is about positive/negative emotion. Stance is about agreement/disagreement with a target.

"Climate change is a serious problem we must address" → Negative sentiment, FAVOR stance toward climate action
"These climate activists are so inspiring!" → Positive sentiment, FAVOR stance toward climate action

Part 1: Basic Stance Detection

Target: Renewable Energy

Example 1

"Wind and solar are the future. We need to invest more in clean energy infrastructure."

FAVOR

AGAINST

NEUTRAL

Example 2

"Solar panels are getting cheaper, but they still can't match the reliability of natural gas."

FAVOR

AGAINST

NEUTRAL

Example 3

"The energy sector is undergoing significant changes as both renewable and traditional sources compete for market share."

FAVOR

AGAINST

NEUTRAL

Question 1

Example 2 contains both positive and negative points about renewables. How did you decide on the stance?

Part 2: Implicit vs. Explicit Stance

Target: Universal Basic Income (UBI)

Example 4 (Explicit)

"I strongly support UBI as a solution to technological unemployment."

FAVOR

AGAINST

NEUTRAL

Example 5 (Implicit)

"Finland tried UBI and ended the experiment after two years."

FAVOR

AGAINST

NEUTRAL

Example 6 (Implicit)

"Where exactly is the money for UBI supposed to come from?"

FAVOR

AGAINST

NEUTRAL

Question 2

Example 5 states a fact. Example 6 asks a question. Can facts and questions express stance?

Facts and questions should be NEUTRAL They can imply stance through framing Depends on context and intent

Should annotators infer implicit stance or stick to explicit statements?

Part 3: Target Ambiguity

Target: Immigration

Example 7

"Legal immigration strengthens our economy, but we need better border security."

FAVOR

AGAINST

NEUTRAL

Question 3

The target "immigration" is broad. This text is favorable toward legal immigration but suggests limits on illegal immigration. How should this be labeled?

FAVOR (supports some immigration) AGAINST (wants restrictions) NEUTRAL (mixed position) Split annotation (FAVOR legal, AGAINST illegal)

How should annotation handle nuanced positions on complex targets?

Part 4: Distinguishing Stance from Topic

Target: Gun Control

Example 8

"The Second Amendment guarantees our right to bear arms."

FAVOR

AGAINST

NEUTRAL

Question 4

This statement is about gun rights, not directly about gun control. But opposing gun control often means supporting gun rights. Is this AGAINST gun control?

Part 5: Sarcasm and Irony in Stance

Target: Electric Vehicles

Example 9

"Sure, let's all buy electric cars and pretend the electricity comes from magic instead of coal plants."

FAVOR

AGAINST

NEUTRAL

Question 5

The literal text could suggest support, but the sarcasm indicates opposition. Should stance annotation capture literal or intended meaning?

Literal meaning Intended meaning Flag as sarcastic, annotate intended

Part 6: Group Discussion

Question 6

Compare annotations with your group. Where did you disagree?

Implicit stance inference Target interpretation Nuanced/mixed positions NEUTRAL threshold Sarcasm handling

Describe one specific disagreement:

Part 7: Reflection

Question 7

Why is stance detection politically sensitive?

Annotator political bias affects labels Target framing affects interpretation Applications can be used for surveillance Complex views get oversimplified

Key Takeaway

Stance annotation requires understanding targets, inferring intent, and acknowledging that political views rarely fit neat categories.

Same text → different stance depending on target definition
Implicit stance requires inference that may introduce bias
Real opinions are often nuanced; forced-choice labels lose information
Annotator backgrounds affect how they interpret political content