Worksheet: Event Causal Relation Extraction

Background

You are annotating data for an AI system that extracts causal relations between events from news text.

A causal relation means that: One event (CAUSE) brings about or contributes to another event (EFFECT)

This task requires:

Identifying events
Deciding whether a causal relation exists
Assigning direction (what caused what)

Event Definition (for this worksheet)

An event is a real-world occurrence or action described in the text (e.g., attack, protest, collapse, evacuation, shutdown).

Causal Relation Labels

CAUSES — Event A directly causes Event B

ENABLED — Event A made Event B possible but did not directly cause it

NO-RELATION — No causal relation is expressed

UNCERTAIN — Causality is suggested but not clearly stated

Part 1: Warm-up — What Counts as Causation?

Question 1

Which of the following words or phrases often signal causality? (Check all that apply.)

because after led to resulted in during triggered followed by

Which one is most ambiguous? Why?

Part 2: Explicit Causation (Easier Case)

Sentence 1

"Heavy rainfall caused flooding across the region."

Event A: heavy rainfall | Event B: flooding

Question 2

What is the relation between Event A and Event B?

CAUSES ENABLED NO-RELATION UNCERTAIN

Why?

Part 3: Temporal ≠ Causal (Classic Trap)

Sentence 2

"Flooding occurred after heavy rainfall in the region."

Event A: heavy rainfall | Event B: flooding

Question 3

Is this sentence explicitly causal?

Yes No Unclear

Explain the difference between temporal order and causal relation:

Part 4: Implicit Causality (Harder Case)

Sentence 3

"The bridge collapsed, leaving dozens stranded."

Event A: bridge collapsed | Event B: people stranded

Question 4

Is Event A causally related to Event B?

CAUSES ENABLED NO-RELATION UNCERTAIN

What makes this decision difficult?

Part 5: Multi-Event Causal Chains

Sentence 4

"An explosion damaged the power station, causing widespread outages and forcing evacuations."

Event A: explosion | Event B: power station damaged | Event C: power outages | Event D: evacuations

Question 5

Which causal relations should be annotated? (Check all that apply.)

A → B B → C C → D A → D None

Explain one choice:

Part 6: Negation and Counterfactuals

Sentence 5

"The evacuation was avoided because the fire was quickly contained."

Event A: fire contained | Event B: evacuation

Question 6

How should this be annotated?

A CAUSES B A PREVENTS B NO-RELATION UNCERTAIN

Why is negation hard for causal annotation?

Part 7: Speculation and Modality

Sentence 6

"Officials said the strike may have triggered the supply shortage."

Question 7

How confident is the causal claim?

Certain Probable Speculative Not stated

What label would you assign?

CAUSES ENABLED UNCERTAIN NO-RELATION

Explain:

Part 8: World Knowledge vs Text Evidence

Sentence 7

"The hurricane hit the coast. Thousands of homes were destroyed."

Question 8

Should you annotate a causal relation?

Yes, based on world knowledge Yes, based on textual evidence No, causality is not stated UNCERTAIN

Why is this a hard judgment?

Part 9: Group Discussion

Discuss with your group.

Question 9

Where did you disagree most?

Event identification Relation existence Relation direction Certainty vs speculation

Give one concrete example:

Part 10: Annotation Design Decisions

Question 10

Which guideline would most reduce disagreement?

Only annotate explicit causal markers Allow implicit causation Separate temporal and causal relations Add confidence levels to relations Use expert annotators only

Explain your choice:

Part 11: Modeling Consequences (ML Perspective)

Question 11

If annotators disagree on causality, what might a model trained on this data learn?

Spurious correlations Overconfident causal predictions Conservative predictions Dataset-specific biases

Explain one outcome:

Final Reflection

Question 12

Why is causal relation extraction harder than event detection alone?

Requires interpretation beyond text Depends on world knowledge Involves directionality Often under-specified All of the above

Key Takeaway (for class discussion)

Causality is not directly observable in text—it is inferred, interpreted, and often underspecified.
Structured annotation tasks encode theories about the world, not just labels.

One-sentence takeaway: In structured annotation, you are modeling relationships between interpretations, not just words.