Rule-Based Information Extraction and Relation Patterns

Turn Domain Language into Auditable Extraction Rules

• Write a token pattern for a narrow business relation.
• Return evidence spans and a rule identifier instead of a bare value.
• Prefer a small tested rule set over hundreds of interacting exceptions.
• Test each relation rule against a positive example, a near miss and a deliberate counterexample.

Choose a relation you can define

Purchase, employment and version references have different arguments. Write the output schema before the pattern.

Include counterexamples such as planned to buy or did not acquire so the rule does not confuse mention with fact.

Layer deterministic and statistical methods

Use rules for stable formats and precise language. Use a model for varied wording, then route low-confidence cases to review.

Negation and scope matter

The word not can modify a distant predicate. A local window rule may still misread reported speech or hypothetical language.

Treat extracted relations as claims supported by text, not verified facts about the world.

Version every rule

Store a rule ID with output and keep an example suite for each revision.

Compare rule-level precision before adding coverage.

Build the Rule-Based Information Extraction and Relation Patterns example

The rule matches a version cue followed by a constrained version token. It does not pretend to understand deployment status.

Return offsets so a reviewer can read the phrase in context. Add document ID and rule version in a real output record.

Install:

python -m pip install spacy
python -m spacy download en_core_web_sm

import spacy
from spacy.matcher import Matcher
nlp = spacy.load('en_core_web_sm')
matcher = Matcher(nlp.vocab)
matcher.add('PRODUCT_VERSION', [[
    {'LOWER': {'IN': ['version','release']}},
    {'TEXT': {'REGEX': r'^v?\d+(?:\.\d+){0,2}$'}}
]])
text = 'We deployed version 3.2.1, but version soon is not a valid match.'
doc = nlp(text)
for match_id, start, end in matcher(doc):
    span = doc[start:end]
    print(nlp.vocab.strings[match_id], span.text, span.start_char, span.end_char)

Expected Relation Records and Evidence Spans

PRODUCT_VERSION version 3.2.1 12 25

Test release v2, version 3, malformed versions and negated sentences. Decide which are valid for your schema.

If version formats are governed by a product registry, validate the candidate against that registry after extraction.

Diagnose failures in Rule-Based Information Extraction and Relation Patterns

Run positive and negative examples per rule and report results by rule ID.

Check your Rule-Based Information Extraction and Relation Patterns reasoning

Evaluate the rule boundaries before opening the explanations. The cases test negation, scope and versioned pattern behaviour.

Primary references for Rule-Based Information Extraction and Relation Patterns

• spaCy rule-based matching: official token and phrase matching guidance.
• Python regular expression documentation: confirm the syntax behind fallback relation patterns.

Review the extractor whenever domain terminology or spaCy matching behaviour changes. Keep the rule version and false-match ledger with the release.