Rule-based classifiers

A rule classifier is a small piece of hand-written code that labels tokens. Pelias and Mailwoman v1 are built almost entirely on rule classifiers. Mailwoman v2 keeps them and adds the neural classifier alongside — see How it works now for the hybrid.

This article walks through how rule classifiers are structured and what they are good and bad at.

One classifier per component

In Mailwoman's design, there is one rule classifier per address component type, and they all run in parallel. The interface a classifier conforms to looks roughly like:

interface RuleClassifier {
	id: string // 'postcode', 'house_number', …
	component: ComponentTag
	classify(tokens: Token[], context: Context): ClassificationProposal[]
}

Each classifier:

Looks at the full token sequence (so it has context — knowing the next or previous token matters).
Returns zero or more proposals (each saying "I think tokens i..j are a postcode with confidence 0.95").
Does not depend on what any other classifier returns. The merging happens in the solver.

A worked example — the postcode classifier

Probably the simplest classifier. Postcodes are a regex problem, not a machine learning problem.

const US_POSTCODE = /^\d{5}(-\d{4})?$/  // 10118 or 10118-1234
const FR_POSTCODE = /^\d{5}$/            // 75008
const UK_POSTCODE = /^[A-Z]{1,2}\d[A-Z\d]? ?\d[A-Z]{2}$/i  // SW1A 1AA

classify(tokens: Token[], context: Context): ClassificationProposal[] {
	const out: ClassificationProposal[] = []
	for (const token of tokens) {
		if (US_POSTCODE.test(token.text)) {
			out.push({
				span: token.span,
				component: "postcode",
				confidence: 0.95,
				source: "rule",
				source_id: "rule:postcode:us",
				penalty: 0,
				metadata: { country_hint: "US" },
			})
		}
		// FR + UK regexes similarly
	}
	return out
}

Three things to notice:

Confidence is not 1.0 even though the regex matches. A 5-digit number could also be a house number (rare) or a year. The solver sees the high score but is allowed to reject it if a better story emerges from the other classifiers.
Country hint as metadata. The classifier sets country_hint: "US" when the US pattern fires. The solver can use this to bias the country classification.
Multiple patterns coexist. If the same token matches both US and UK regexes, the classifier emits two proposals. The solver picks.

A harder example — the locality classifier

This one is harder because there is no regex for "is this word a city name?". The Mailwoman v1 locality classifier looks up the token (and short n-grams of consecutive tokens) in a Who's On First dictionary file. Pseudo-code:

const WOF_LOCALITIES: Set<string> = loadDictionary("wof/locality.txt")

classify(tokens: Token[], context: Context): ClassificationProposal[] {
	const out: ClassificationProposal[] = []
	for (let i = 0; i < tokens.length; i++) {
		// Try 1-, 2-, and 3-token spans
		for (let len = 1; len <= 3; len++) {
			const candidate = tokens.slice(i, i + len)
			const text = candidate.map((t) => t.text).join(" ")
			if (WOF_LOCALITIES.has(text)) {
				out.push({
					span: { start: candidate[0].span.start, end: candidate.at(-1)!.span.end },
					component: "locality",
					confidence: 0.6,
					source: "rule",
					source_id: "rule:whos_on_first:locality",
					penalty: 0,
				})
			}
		}
	}
	return out
}

Notes:

Confidence is moderate (0.6) because dictionary hits can be misleading. "New" is the start of many city names, so a 1-token match on "New" is weaker than a 3-token match on "New York City".
N-gram sweep is how multi-word cities work. The classifier tries spans of length 1, 2, and 3 and emits a proposal for each match.
No exclusion logic. If "New York" matches as a locality and "New" matches as a different component, both proposals exist; the solver picks one self-consistent combination.

What rule classifiers are good at

Determinism. Same input, same output, every time. No retraining, no random initialization.
Easy to fix. A bug in a regex is one line. A bug in a neural model is a retraining run.
No data required. A rule classifier ships in code. The dictionary classifiers ship a few MB of dictionaries, not a multi-GB model.
Fast. Pattern matching is microseconds per address. The whole parse runs in milliseconds.
Explainable. Every proposal carries a source_id. A consumer that sees rule:postcode:us on a wrong proposal knows exactly where to look.

What rule classifiers are bad at

The long tail. Real-world inputs do not fit predictable patterns. For every regex you write, there are inputs the regex misses.
Context. A regex sees one token at a time. "Lincoln" might be a city, a person, or a venue. A regex cannot tell.
Languages. Every locale needs its own rules, hand-written, by someone who knows that language.
Multi-word things. N-gram sweeps work for 2- and 3-token cases. "Saint Petersburg, FL" is fine; "Mt Tabor Forest Preserve" (a 4-token venue) is harder.
Negative evidence. Rules can say "this looks like X". They cannot easily say "this looks like X but only in the absence of Y".

These weaknesses are the motivation for the neural classifier. See How the model reasons.

How rule classifiers and the neural classifier coexist

In Mailwoman v2, both produce ClassificationProposals in the same shape. The policy registry decides whose proposal wins for each component:

const policy: ClassifierPolicy[] = [
	{ component: "postcode", mode: "rule_only" }, // regex wins, model not even consulted
	{ component: "country", mode: "rule_preferred" }, // rule wins unless rule confidence is very low
	{ component: "venue", mode: "neural_only" }, // rules can't classify venues at all
	{ component: "house_number", mode: "neural_preferred" }, // model is stronger here in v3.0.0
]

This is the Ship-of-Theseus pattern. The neural classifier earns each component one at a time. If a neural model regresses, flipping back to rule_only is one line of config.

Where this lives in the code

core/classifiers/ (in the v1 tree) — every rule classifier
core/policy/ — the per-component policy registry
core/solver/ — the merger that picks self-consistent combinations

One classifier per component​

A worked example — the postcode classifier​

A harder example — the locality classifier​

What rule classifiers are good at​

What rule classifiers are bad at​

How rule classifiers and the neural classifier coexist​

Where this lives in the code​

See also​