Can a seed pattern be manually written?

Yes; it can encode a carefully selected expression.

Is a seed tuple the same thing?

No; it is an example relation instance.

Must the seed pattern cover every wording?

No; expansion can discover additional expressions.

AI Glossary / Natural language and speech / Advanced

Seed pattern in IE

Definition

A seed pattern in information extraction (IE) is an initial rule or template used to find examples of a target relation or category in text. It supplies starting evidence for extraction or learning. In a bootstrapping system, its matches can help discover additional patterns and examples, gradually expanding beyond the original rule.

What it means

A pattern describes a form of expression, such as a person followed by founded and an organization. Its variable positions specify what to extract and in which roles. A seed tuple instead supplies a particular relation instance, such as a named person and organization. The two can start related learning procedures, but they are different inputs.
Seed describes the starting role; it does not guarantee correctness, full coverage, or verified knowledge.
A useful initial pattern is often selective. Early errors can become inputs for later learning, so a broad rule that collects many irrelevant examples may be worse than a narrow rule with limited coverage. The balance depends on the task.

How it works

Define the target relation and argument types first. For a founder relation, distinguish the person who founded an organization from its current manager, investor, or spokesperson. Then specify a pattern over words, entity labels, syntax, or another representation that provides evidence for that relation.
Apply the pattern to a corpus and retain candidate extractions with their supporting passages. Accepted candidates can be used to locate other contexts expressing the relation. Generalizing those contexts produces candidate patterns, which are evaluated before they expand the extraction process.
Confidence estimates and review can limit error propagation. Repeated appearances are not necessarily independent support: copied passages or recurring ambiguous wording can make an unreliable candidate look strong. A learned pattern should be assessed for what it extracts, not merely how often it appears.
Bootstrapping alternates between patterns and instances. Brin's work on extracting patterns and relations illustrates this mutual expansion. Representations and acceptance rules vary; a seed pattern is a component, not a unique algorithm.

Three practical examples

A company-directory project begins with a carefully checked founder-expression pattern and reviews the person - organization pairs it retrieves.
A terminology project starts from a selective phrase indicating category membership and tests alternative expressions found around accepted examples.
A document-processing team uses a structured announcement phrase to collect initial examples before training a broader extractor.

Think of it this way

Think of searching with a reliable clue. Examples suggest further clues, but each expansion needs checking. A misleading clue can redirect later searching. More material does not necessarily mean more correct material.

Worked example

Consider a fictional task: extract affirmative founder(person, organization) claims. Assume entity spans and their types are already supplied. The initial pattern P0 matches a PERSON mention, immediately followed by founded, immediately followed by an ORGANIZATION mention. It preserves the person - organization argument order.
Four sample sentences are “Olin founded MossLab,” “Rhea founded ValeWorks,” “MossLab was founded by Olin,” and “Rhea joined ValeWorks.” P0 extracts the first two pairs, misses the passive sentence, and does not match joined. These decisions follow the narrow rule.
The first and third sentences describe the same founder pair. Counting matched sentences and counting unique relation tuples therefore answer different questions. P0 has two matched sentences and two unique tuples in this sample; adding the passive pattern would raise matched sentences to three while leaving the unique tuple count at two.
Suppose searching around the accepted Olin - MossLab pair reveals the passive expression. A candidate pattern P1 matches ORGANIZATION followed by was founded by and PERSON. Its output must reverse surface order to produce founder(person, organization). Treating the first surface mention as the founder would create the incorrect tuple founder(MossLab, Olin).
Now consider an invented held-out audit with ten true founder-claim tuples. Assume the combined accepted patterns produce eight unique tuples: seven agree with the reference and one is incorrect. Precision is 7/8=87.5%, recall is 7/10=70%, and F1 is 14/18, approximately 77.8%.
A broader candidate P2 treats PERSON joined ORGANIZATION as founder evidence. On the same audit, suppose it adds four previously unseen predictions, all incorrect under the founder schema. The system now has twelve predictions but still only seven correct ones. Precision falls to 7/12, approximately 58.3%; recall remains 70%; F1 becomes 14/22, approximately 63.6%.
More predictions have not improved true-claim coverage. If those four errors are accepted as training instances, they can encourage still more patterns about employment instead of founding. That is a possible route to semantic drift.
The exercise assumes supplied entities, deduplication, and reference labels. Real passages containing quotation, negation, questions, or hypothetical language require additional scope handling. Even a correctly extracted affirmative statement remains a claim from its source rather than independent verification of the founding event.

Illustrative case study

Fictional example: OriginIndex builds a directory from company histories. Its seed rule retrieves founding statements, but a later expansion begins admitting employment announcements. Reviewers trace the change to a broad joined pattern.
The team removes that candidate, reviews affected tuples, and retains the supporting passage and producing pattern for each extraction. It evaluates additions on held-out examples before another expansion. This fictional scenario illustrates error tracing.

Common misconceptions and mistakes

A seed pattern is not a random-number seed or a particular extracted tuple. It is also not necessarily a regular expression: patterns can use syntactic or entity information.
High precision on a small seed sample does not establish broad coverage or future precision. A pattern can fail under changed wording, domain, or entity recognition. Do not count repeated copies as independent confirmation, and do not assume that accepting a new pattern is safe merely because it retrieves more candidates.

Business use

Record the target schema, seed rules, supporting passages, and iteration history. Review new patterns before allowing their outputs to shape later learning. Measure unique correct extractions and relevant coverage alongside raw volume. Preserve provenance for revising affected records.

Check your understanding

Why must P1 reverse surface order?

Answer: The organization appears before the person in the passive wording.

Did P2 improve recall in the audit?

Answer: No.

Why can a wrong early pattern matter later?

Answer: Its outputs may become evidence for further learning.

Related terms

Information extraction
Seed tuples
Relation extraction
Precision
bootstrapping
semantic drift
extraction pattern
provenance

Learn next

Implement a narrow pattern on a small annotated corpus. Compare mention counts with unique tuples, then review one proposed expansion for argument order and false positives.

Sources

Jurafsky and Martin, Speech and Language Processing, chapter 21, semisupervised relation extraction via bootstrapping: https://web.stanford.edu/~jurafsky/slp3/21.pdf Sergey Brin, Extracting Patterns and Relations from the World Wide Web, original paper: https://www.cs.cmu.edu/~ReadTheWeb/handouts/SergeiBrinBootstrapping1999.pdf

Previous: Security Next: Seed tuples