Must seed tuples be pairs?

No; the schema determines the number of fields.

Can a tiny seed set work?

It can start extraction, but quality and coverage still need evaluation.

Does adding more seeds always help?

No; incorrect or redundant seeds may add little value or introduce errors.

AI Glossary / Natural language and speech / Advanced

Seed tuples

Definition

Seed tuples are a small initial set of example records used to start an information-extraction or learning process. Each tuple supplies values in the roles of a target relation, such as manufacturer(product, organization). A bootstrapping system uses these examples to find relevant contexts, learn extraction patterns, and propose additional tuples.

What it means

A tuple is an ordered collection of values under a schema. The order tells the system which value fills which role. In manufacturer(product, organization), reversing the two values changes the record's meaning or makes it inconsistent with the schema.
Seed tuples provide examples of what to find. Seed patterns instead describe expressions or structures that may reveal those examples in text. A system can alternate between the two: known tuples reveal contexts, and useful contexts become patterns that identify new tuples.
A small starting set reduces annotation work but still needs supervision: relation definitions, checked seeds, and acceptance rules. Poor seeds can distort later extraction.

How it works

First define the relation's argument roles, entity identifiers, and inclusion criteria. Distinguish, for example, the organization that manufactures a product from the organization that sells or distributes it. Then choose a small set of trusted examples with evidence appropriate to that definition.
Search the target corpus for occurrences of the seed entities and inspect their surrounding contexts. Generalize promising contexts into candidate patterns. Apply accepted patterns to find new candidate tuples, assess them, and use selected additions to continue the process.
Brin's DIPRE work demonstrates this alternation between relation instances and extraction patterns. Its starting examples were author - title pairs. The broader idea also applies to relations with more arguments.
Entity co-occurrence is only a search cue. A sentence can mention the entities of a true seed tuple while discussing a different relationship. Candidate patterns and their outputs therefore need checks beyond “both names appeared.” Retaining the source passage and extraction history helps expose such errors.

Three practical examples

A product-catalog team starts with verified product - manufacturer examples and searches for alternative ways the relationship is expressed.
A bibliography project uses known author - title pairs to discover recurring citation formats.
A domain glossary project begins with checked term - category pairs and reviews patterns that suggest additional category members.

Think of it this way

Think of giving a researcher completed forms as examples. They clarify the desired fields, but nearby names do not automatically belong in them. A perfectly formatted example can still teach the wrong relationship.

Worked example

Consider an original fictional task using the schema manufacturer(product, organization). The two checked seed tuples are (Luma Lamp, Northcraft) and (Cove Desk, Westline). Assume resolved entities and correct seeds.
The corpus contains “Northcraft manufactures the Luma Lamp” and “Westline manufactures the Cove Desk.” These occurrences suggest a candidate pattern of ORGANIZATION manufactures PRODUCT. The pattern must output the product first and organization second to match the chosen schema.
Applying that pattern elsewhere retrieves “Eastmill manufactures the Brook Chair,” producing candidate tuple (Brook Chair, Eastmill). The tuple is a proposed instance; the pattern is a reusable extraction rule.
Now suppose another passage says “Northcraft distributes the Luma Lamp.” It contains both entities of a valid manufacturer seed, but its wording expresses distribution. Learning manufactures and distributes as interchangeable would be unjustified. A manufacturer may also distribute its own product, yet other distributors need not manufacture what they sell.
For a numerical audit, suppose one iteration proposes six unique tuples beyond the original two seeds. Human review accepts four and rejects two. The precision of these novel candidates is 4/6, approximately 66.7%. The accepted collection grows from two to six tuples, a net addition of four.
If the two already-known seeds were included among the evaluated outputs, the combined set would contain eight tuples with six correct, giving 6/8=75%. That figure answers a different question and makes the novel-candidate result look stronger. A report should say whether seed examples are included or excluded.
Assume the next iteration outputs five tuples, three of which are already in the accepted collection. Only two are novel. Review accepts one of those two. The accepted collection now grows from six to seven, not to eleven. Novel-candidate precision for this second iteration is 1/2=50%; five output records do not mean five discoveries.
Across the two iterations, there were eight novel candidates, five accepted and three rejected. The combined novel-candidate precision is 5/8=62.5%, calculated from counts rather than by averaging the two iteration percentages. The final accepted set contains the original two seeds plus five additions.
These invented figures assume complete, correct candidate review. Recall remains unknown because the total true tuple count is unspecified.

Illustrative case study

Fictional example: MakerLedger seeds a catalog extractor with trusted examples from one manufacturer's brochures. The initial patterns work on similarly formatted pages but miss descriptions written by smaller suppliers.
The team adds checked examples covering different wording and separates seller relationships from manufacturer relationships. It records novel accepted tuples per iteration and keeps a review sample outside the expansion process. This fictional scenario illustrates seed selection and evaluation design.

Common misconceptions and mistakes

Seed tuples are not random-number seeds, extraction patterns, or a complete training corpus. They are starting examples under a relation schema. Their order and entity identity matter.
A true relation does not make every sentence mentioning its entities a valid expression of that relation. Another mistake is counting duplicates or original seeds as new discoveries. Precision among proposed tuples also does not establish recall without a suitable reference for the true tuples being sought.

Business use

Choose seed examples that match the operational definition of the relation and retain their evidence. Track original seeds, proposed additions, accepted additions, duplicates, and rejected candidates separately. Review patterns that arise from ambiguous contexts. Evaluate coverage on representative material rather than assuming success on the seed sources transfers everywhere.

Check your understanding

What does the manufactures pattern output first in the example?

Answer: The product.

How many accepted tuples remain after both iterations?

Answer: Seven, including the two seeds.

Can the stated counts determine recall?

Answer: No; the total true relation set is unknown.

Related terms

Relation extraction
Tuple
Precision
Seed pattern
bootstrapping
entity resolution
semantic drift
deduplication

Learn next

Build a small seed set with documented evidence. Run one extraction iteration and compare raw outputs, unique novel candidates, and accepted additions before expanding again.

Sources

Jurafsky and Martin, Speech and Language Processing, chapter 21, seed tuples and bootstrapped relation extraction: https://web.stanford.edu/~jurafsky/slp3/21.pdf Sergey Brin, Extracting Patterns and Relations from the World Wide Web, DIPRE's seed instances and iterative expansion: https://www.cs.cmu.edu/~ReadTheWeb/handouts/SergeiBrinBootstrapping1999.pdf

Previous: Seed pattern in IE Next: Segmentation