Worksheet: Relation Extraction

Background

Relation extraction identifies semantic relationships between entities mentioned in text, creating structured knowledge from unstructured data.

Relation Triple: (Subject Entity, Relation Type, Object Entity)

Example: "Steve Jobs founded Apple" → (Steve Jobs, FOUNDED, Apple)

Common Relation Types

Relation	Description	Example
WORKS_FOR	Employment	(John, WORKS_FOR, Google)
LOCATED_IN	Physical location	(Paris, LOCATED_IN, France)
BORN_IN	Birthplace	(Einstein, BORN_IN, Germany)
FOUNDED	Created organization	(Gates, FOUNDED, Microsoft)
SPOUSE	Married to	(Obama, SPOUSE, Michelle)
SUBSIDIARY_OF	Owned by	(Instagram, SUBSIDIARY_OF, Meta)

Entity Types

PERSON ORGANIZATION LOCATION DATE MISC

Part 1: Basic Relation Identification

Elon Musk is the CEO of Tesla, which is headquartered in Austin, Texas.

Question 1

Identify all relation triples in the sentence:

Triple 1: (, , )

Triple 2: (, , )

Is "Austin, Texas" one entity or two?

Part 2: Implicit vs. Explicit Relations

Dr. Sarah Chen, a professor at MIT, published her findings in Nature last month.

Question 2

Which relations are explicitly stated vs. inferred?

Relation	Explicit or Inferred?
(Dr. Sarah Chen, WORKS_FOR, MIT)
(Dr. Sarah Chen, PUBLISHED_IN, Nature)
(Dr. Sarah Chen, PROFESSION, professor)

Should annotation include inferred relations?

Part 3: Relation Directionality

Amazon acquired Whole Foods in 2017.

Question 3

What is the correct direction for the acquisition relation?

(Amazon, ACQUIRED, Whole Foods) (Whole Foods, ACQUIRED_BY, Amazon) Both are valid representations Use symmetric relation MERGER

How should guidelines handle relation directionality?

Part 4: Nested and Complex Relations

Bill Gates and Paul Allen co-founded Microsoft in Albuquerque.

Question 4

How should co-founding be represented?

Two separate triples: (Gates, FOUNDED, Microsoft) and (Allen, FOUNDED, Microsoft) Joint entity: ([Gates, Allen], CO_FOUNDED, Microsoft) Qualified relation: (Gates, FOUNDED, Microsoft, with=Allen) Event-based: (FoundingEvent, FOUNDER, Gates), (FoundingEvent, FOUNDER, Allen)

What about the location relation?

(Microsoft, FOUNDED_IN, Albuquerque) (FoundingEvent, LOCATION, Albuquerque) Both representations are valid Don't annotate - not a core relation

Part 5: Negative and Uncertain Relations

John used to work at Google but now works at Meta. Mary might join Apple next year.

Question 5

How should these relations be annotated?

Statement	Relation Status
(John, WORKS_FOR, Google)
(John, WORKS_FOR, Meta)
(Mary, WORKS_FOR, Apple)

Should relation extraction capture temporal and modal information?

Part 6: Annotate This Passage

Tim Cook, who succeeded Steve Jobs as CEO of Apple, announced that the company would invest $1 billion in North Carolina. Apple, based in Cupertino, has been expanding its operations across the United States.

Question 6

Extract all relation triples (aim for at least 5):

Which relations were hardest to decide?

Part 7: Group Discussion

Question 7

Compare your annotations with your group. Where did you disagree?

Entity boundaries Relation type selection Relation directionality Implicit vs explicit Temporal aspects Missing relations

Describe one specific disagreement:

Part 8: Reflection

Question 8

Why is relation extraction difficult?

Many relations are implicit Schema design is subjective Real sentences have complex structures Context is needed for disambiguation Temporal/modal aspects are complex All of the above

Key Takeaway

Relation extraction turns unstructured text into structured knowledge, but requires many annotation decisions.

The relation schema determines what can be captured
Implicit relations require inference and increase annotator burden
Directionality and temporal aspects need clear guidelines
Entity identification and relation extraction are interdependent