Tokens and labels

The question

The parser reports spansspanA contiguous range of characters or tokens in the input string, tagged with an address component type (street, locality, postcode, etc.). Parsed addresses are represented as collections of spans, possibly nested in a tree. of your input string: characters 35 to 45 are a localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighborhood in the hierarchy.. Underneath, it is making one decision per piece of text, and the piecesECE (Expected Calibration Error). A metric that measures how well a model's confidence scores align with its actual accuracy. Lower is better. Mailwoman's held-out ECE drops from 0.067 (raw) to 0.0035 (calibrated). are not words — analytics arrives at the modelneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.' as ▁Ana ly t ic s. Two questions follow from that. Why fragments rather than words? And how do five separate per-piece decisions add up to one spanspanA contiguous range of characters or tokens in the input string, tagged with an address component type (street, locality, postcode, etc.). Parsed addresses are represented as collections of spans, possibly nested in a tree. you can act on?

The analog

You have written the word-list version of this. A set of known street suffixesstreet affixA modifier on a street name indicating type or direction — Street, Avenue, rue, Calle, N, East. Mailwoman tags these as street_prefix / street_suffix, recognized via a morphology FST., a set of state abbreviations, a table of citylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighborhood in the hierarchy. names — and a tokenizertokenizerThe component that converts a raw address string into a sequence of numeric token IDs the model can process. Mailwoman's tokenizer is a SentencePiece unigram model trained specifically on postal addresses. that splits on whitespace so the tokenstokenOne word or subword in the tokenized input. For the neural classifier, tokens come from SentencePiece (subword units); for the rule classifiers, tokens are whitespace- and punctuation-separated words. line up with the table keys.

It works until it meets a word the table has never seen, and then it has nothing at all to say. Not a weaker opinion: no opinion, because the lookup either hits or misses. Pennsylvana, typed by a person in a hurry, is as unknown to a word table as Xzqjrt. And the table itself has no natural size — every new countrycountryThe top-level address component (an ISO country). Closed-vocabulary, so it is best handled by a deterministic matcher feeding a proposal rather than a retrained model head., every new spelling, every business name is another key, and a table with millions of keys is not a table any more, it is the corpuscorpusThe BIO-labeled training data used to train Mailwoman's neural classifier. Assembled from real sources (OpenAddresses, National Address Database) and synthetic shards (boundary stress, order variants, negative space). Managed by @mailwoman/corpus..

The fix is the printer's type case rather than the dictionary. Keep a fixed inventory of fragments, and spell every word out of them. Rare words cost more fragments; unseen words still get spelled. Nothing is unrepresentable.

Pieces, not words

Mailwoman spells input with SentencePiece, an open-source subword tokenizertokenizerThe component that converts a raw address string into a sequence of numeric token IDs the model can process. Mailwoman's tokenizer is a SentencePiece unigram model trained specifically on postal addresses.. The inventory holds 73,143 piecesECE (Expected Calibration Error). A metric that measures how well a model's confidence scores align with its actual accuracy. Lower is better. Mailwoman's held-out ECE drops from 0.067 (raw) to 0.0035 (calibrated)., chosen by a unigram language-modelneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.' fit over trainingtrainingThe process of adjusting a model's parameters so its predictions match labeled examples, by repeatedly measuring error and nudging the weights to reduce it. Distinct from inference, when the trained model is run on new input. text, and the ▁ prefix marks a piece that begins a new word. Common words are one piece; ▁Washington is a single unitunitA subdivision of a building — apartment, suite, floor — that refines a street address. Mailwoman's unit component; a designator plus identifier forms a subpremise.. Uncommon ones fragment, and the fragmentation carries information — a word that splits into five piecesECE (Expected Calibration Error). A metric that measures how well a model's confidence scores align with its actual accuracy. Lower is better. Mailwoman's held-out ECE drops from 0.067 (raw) to 0.0035 (calibrated). is a word the trainingtrainingThe process of adjusting a model's parameters so its predictions match labeled examples, by repeatedly measuring error and nudging the weights to reduce it. Distinct from inference, when the trained model is run on new input. text did not see much of.

Two properties matter downstream.

The inventory is closed and small. A modelneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.''s output layerlayerOne transformer block — attention plus a feed-forward network, with normalization and residual connections — applied to every position. Stacking layers lets the model build up richer representations; Mailwoman's encoder has 6. and its embeddingembeddingA vector of numbers representing a token (or other item) so that similar items sit near each other in vector space. The first thing the model does is turn each token into an embedding. table are both sized by the inventory, so bounding it at 73,143 is what keeps the shipped modelneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.' a 39.4 MB file rather than a multi-gigabyte one.

Nothing is out of vocabularyvocabularyThe fixed set of tokens a tokenizer can produce. Mailwoman's SentencePiece vocabulary is tens of thousands of subword pieces, with byte fallback for anything outside it.. When a character is not covered by any piece, the tokenizertokenizerThe component that converts a raw address string into a sequence of numeric token IDs the model can process. Mailwoman's tokenizer is a SentencePiece unigram model trained specifically on postal addresses. falls back to encoding it byte by byte. A Cyrillic streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. name or a stray emoji produces piecesECE (Expected Calibration Error). A metric that measures how well a model's confidence scores align with its actual accuracy. Lower is better. Mailwoman's held-out ECE drops from 0.067 (raw) to 0.0035 (calibrated). rather than a failure. Those piecesECE (Expected Calibration Error). A metric that measures how well a model's confidence scores align with its actual accuracy. Lower is better. Mailwoman's held-out ECE drops from 0.067 (raw) to 0.0035 (calibrated). carry no learned meaning — byte fallbackbyte fallbackA tokenization strategy where characters not seen during training are encoded as raw UTF-8 byte sequences rather than mapped to an unknown token. Mailwoman's tokenizer uses byte fallback so non-Latin scripts (CJK, Cyrillic) produce real token sequences even though training data was Latin-dominant. is what stops the pipelinestaged pipelineMailwoman's runtime architecture: a sequence of pure-function stages (normalize → query-shape → locale-gate → kind-classifier → phrase-grouper → classifier → decoder) connected by typed handoffs. Each stage is published as its own npm package. from crashing, and it is not coveragecoverageThe fraction of a population or region for which a data source has real, non-placeholder entries — e.g. 47% rooftop coverage on Texas addresses. Distinct from accuracy on the rows that are present. of the script. What the model cannot do draws that line.

Pieces are not what you get back

The piece boundaries are an implementation detail of the modelneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.', and the parseaddress parsingThe process of decomposing a free-text postal address string into structured components — house number, street name, locality, region, postcode, and country — so a geocoder can resolve them to coordinates. result never exposes them. Each piece carries the character range it came from, reported by the tokenizertokenizerThe component that converts a raw address string into a sequence of numeric token IDs the model can process. Mailwoman's tokenizer is a SentencePiece unigram model trained specifically on postal addresses. itself rather than reconstructed afterward, and the decoderdecoderIn a transformer encoder-decoder model, the part that produces output sequences. Mailwoman's classifier is encoder-only (no decoder); the 'CRF decoder' is a different thing — a structured-prediction layer that picks the best label sequence from the encoder's outputs. projects labelscomponent tagOne of the 25 labels in Mailwoman's address schema — street, locality, region, postcode, house_number, unit, po_box, country, venue, intersection, and others. Each parsed span carries exactly one component tag. back onto those ranges. ▁Ana ly t ic s covers characters 13 through 22 of the input, so a venuevenueA named, non-address place — a business, building, park, or stadium. Mailwoman's free-text point-of-interest component, added as a Tier 2 fine label. spanspanA contiguous range of characters or tokens in the input string, tagged with an address component type (street, locality, postcode, etc.). Parsed addresses are represented as collections of spans, possibly nested in a tree. that runs from ▁Beacon to s is reported as characters 0 through 22 — one spanspanA contiguous range of characters or tokens in the input string, tagged with an address component type (street, locality, postcode, etc.). Parsed addresses are represented as collections of spans, possibly nested in a tree., in your coordinate system.

Keep the distinction in mind when you read a confidence score, because the score is per piece before it is per spanspanA contiguous range of characters or tokens in the input string, tagged with an address component type (street, locality, postcode, etc.). Parsed addresses are represented as collections of spans, possibly nested in a tree.. A five-piece word contributes five numbers that get aggregated into the one you see.

There is one place the piece-versus-word distinction becomes visible in behavior. Because piecesECE (Expected Calibration Error). A metric that measures how well a model's confidence scores align with its actual accuracy. Lower is better. Mailwoman's held-out ECE drops from 0.067 (raw) to 0.0035 (calibrated). are labeled independently, the five piecesECE (Expected Calibration Error). A metric that measures how well a model's confidence scores align with its actual accuracy. Lower is better. Mailwoman's held-out ECE drops from 0.067 (raw) to 0.0035 (calibrated). of one word can disagree — three of them venuevenueA named, non-address place — a business, building, park, or stadium. Mailwoman's free-text point-of-interest component, added as a Tier 2 fine label., two of them streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. — which is a reading no person would produce and no consumer wants. After decoding, the parser groups piecesECE (Expected Calibration Error). A metric that measures how well a model's confidence scores align with its actual accuracy. Lower is better. Mailwoman's held-out ECE drops from 0.067 (raw) to 0.0035 (calibrated). back into their ▁-delimited words and, for any word whose piecesECE (Expected Calibration Error). A metric that measures how well a model's confidence scores align with its actual accuracy. Lower is better. Mailwoman's held-out ECE drops from 0.067 (raw) to 0.0035 (calibrated). disagree about which kind of component they belong to, holds a confidence-weighted vote and applies the winner across the whole word. It runs by default, and it operates strictly inside a word: Saint Petersburg is two words and is never merged by this pass.

BIO: the bridge from "matched" to "labeled"

The rule-world version: a regular expression returns a spanspanA contiguous range of characters or tokens in the input string, tagged with an address component type (street, locality, postcode, etc.). Parsed addresses are represented as collections of spans, possibly nested in a tree. directly. It matched from here to there, and the spanspanA contiguous range of characters or tokens in the input string, tagged with an address component type (street, locality, postcode, etc.). Parsed addresses are represented as collections of spans, possibly nested in a tree. is the match.

What happens here: the modelneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.' produces one labelcomponent tagOne of the 25 labels in Mailwoman's address schema — street, locality, region, postcode, house_number, unit, po_box, country, venue, intersection, and others. Each parsed span carries exactly one component tag. per piece, and a labelcomponent tagOne of the 25 labels in Mailwoman's address schema — street, locality, region, postcode, house_number, unit, po_box, country, venue, intersection, and others. Each parsed span carries exactly one component tag. alone cannot express a boundary. If every piece of Saint Petersburg, FL were labeled just locality, locality, region, there would be no way to tell one two-word town from two one-word towns sitting side by side.

BIO labeling is the standard fix, and it is older than any of this — it is what named-entity recognition has used since the 1990s. Each labelcomponent tagOne of the 25 labels in Mailwoman's address schema — street, locality, region, postcode, house_number, unit, po_box, country, venue, intersection, and others. Each parsed span carries exactly one component tag. carries a position as well as a type:

• B-locality — a localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighborhood in the hierarchy. spanspanA contiguous range of characters or tokens in the input string, tagged with an address component type (street, locality, postcode, etc.). Parsed addresses are represented as collections of spans, possibly nested in a tree. begins at this piece.
• I-locality — this piece is inside the localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighborhood in the hierarchy. spanspanA contiguous range of characters or tokens in the input string, tagged with an address component type (street, locality, postcode, etc.). Parsed addresses are represented as collections of spans, possibly nested in a tree. already open.
• O — this piece is outside every component.

So Saint Petersburg, FL labelscomponent tagOne of the 25 labels in Mailwoman's address schema — street, locality, region, postcode, house_number, unit, po_box, country, venue, intersection, and others. Each parsed span carries exactly one component tag. as B-locality, I-locality, O, B-region, and the spanspanA contiguous range of characters or tokens in the input string, tagged with an address component type (street, locality, postcode, etc.). Parsed addresses are represented as collections of spans, possibly nested in a tree. reconstruction is a single walk: open at B-, extend at a matching I-, close at anything else. Two adjacent towns would read B-locality, B-locality, and the second B- is the boundary.

The scheme has one structural weakness, and it is the reason the decoderdecoderIn a transformer encoder-decoder model, the part that produces output sequences. Mailwoman's classifier is encoder-only (no decoder); the 'CRF decoder' is a different thing — a structured-prediction layer that picks the best label sequence from the encoder's outputs. exists: I-locality with no B-locality before it is not a reading of anything. A per-piece argmax will produce that sequence whenever the modelneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.' is unsure about the first word and confident about the second, and the standard symptom is a town losing its first word. Decoding and the best path is where that gets ruled out.

Three counts, and why they differ

Three numbers get quoted around the labelcomponent tagOne of the 25 labels in Mailwoman's address schema — street, locality, region, postcode, house_number, unit, po_box, country, venue, intersection, and others. Each parsed span carries exactly one component tag. space, and they answer different questions.

The schema defines 25 component tagscomponent tagOne of the 25 labels in Mailwoman's address schema — street, locality, region, postcode, house_number, unit, po_box, country, venue, intersection, and others. Each parsed span carries exactly one component tag. — everything a parseaddress parsingThe process of decomposing a free-text postal address string into structured components — house number, street name, locality, region, postcode, and country — so a geocoder can resolve them to coordinates. result is allowed to carry, including tags for Japanese address structure and tags no shipped modelneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.' produces. At two labelscomponent tagOne of the 25 labels in Mailwoman's address schema — street, locality, region, postcode, house_number, unit, po_box, country, venue, intersection, and others. Each parsed span carries exactly one component tag. per tag plus one shared O, that is a capacity of 51 labelscomponent tagOne of the 25 labels in Mailwoman's address schema — street, locality, region, postcode, house_number, unit, po_box, country, venue, intersection, and others. Each parsed span carries exactly one component tag..

The shipped modelneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.''s headattention headOne of several parallel attention computations in a layer, each free to focus on a different kind of relationship between tokens. Their outputs are concatenated — 'multi-head attention'. Mailwoman uses 4 heads. is fixed at trainingtrainingThe process of adjusting a model's parameters so its predictions match labeled examples, by repeatedly measuring error and nudging the weights to reduce it. Distinct from inference, when the trained model is run on new input. time and carries 33 labelscomponent tagOne of the 25 labels in Mailwoman's address schema — street, locality, region, postcode, house_number, unit, po_box, country, venue, intersection, and others. Each parsed span carries exactly one component tag. over 16 tags. That is the alphabet this modelneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.' can write in. Nine tags in the schema are outside it: seven belong to a separate Japanese headattention headOne of several parallel attention computations in a layer, each free to focus on a different kind of relationship between tokens. Their outputs are concatenated — 'multi-head attention'. Mailwoman uses 4 heads. that is not shipped, and two are carried for annotation and formatting rather than produced by the parser.

The gap between the two is not a defect to work around, and it is not post-processing that fills in later. A tag outside the headattention headOne of several parallel attention computations in a layer, each free to focus on a different kind of relationship between tokens. Their outputs are concatenated — 'multi-head attention'. Mailwoman uses 4 heads. never appears in a parseaddress parsingThe process of decomposing a free-text postal address string into structured components — house number, street name, locality, region, postcode, and country — so a geocoder can resolve them to coordinates. from this modelneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.'. Component tags lists all 25 with meanings and examples, and states per modelneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.' which are reachable.

What it costs

Boundaries are implicit, so they inherit labelcomponent tagOne of the 25 labels in Mailwoman's address schema — street, locality, region, postcode, house_number, unit, po_box, country, venue, intersection, and others. Each parsed span carries exactly one component tag. errors. The modelneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.' has no separate boundary headattention headOne of several parallel attention computations in a layer, each free to focus on a different kind of relationship between tokens. Their outputs are concatenated — 'multi-head attention'. Mailwoman uses 4 heads.: a spanspanA contiguous range of characters or tokens in the input string, tagged with an address component type (street, locality, postcode, etc.). Parsed addresses are represented as collections of spans, possibly nested in a tree. begins where a B- labelcomponent tagOne of the 25 labels in Mailwoman's address schema — street, locality, region, postcode, house_number, unit, po_box, country, venue, intersection, and others. Each parsed span carries exactly one component tag. lands. That means it cannot hold a confident view of where a streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. ends while remaining unsure what the streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. is — boundary confusion is downstream of labelcomponent tagOne of the 25 labels in Mailwoman's address schema — street, locality, region, postcode, house_number, unit, po_box, country, venue, intersection, and others. Each parsed span carries exactly one component tag. confusion, always. When a house numberhouse numberThe numeric or alphanumeric identifier of a building on a street. Mailwoman's house_number component; its position relative to the street name flips between locales. gets absorbed into a streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. name, the fix is evidence about the labelcomponent tagOne of the 25 labels in Mailwoman's address schema — street, locality, region, postcode, house_number, unit, po_box, country, venue, intersection, and others. Each parsed span carries exactly one component tag., not a boundary rule.

Fragmentation is uneven across languages. A language whose text is thin in the trainingtrainingThe process of adjusting a model's parameters so its predictions match labeled examples, by repeatedly measuring error and nudging the weights to reduce it. Distinct from inference, when the trained model is run on new input. corpuscorpusThe BIO-labeled training data used to train Mailwoman's neural classifier. Assembled from real sources (OpenAddresses, National Address Database) and synthetic shards (boundary stress, order variants, negative space). Managed by @mailwoman/corpus. spells into more piecesECE (Expected Calibration Error). A metric that measures how well a model's confidence scores align with its actual accuracy. Lower is better. Mailwoman's held-out ECE drops from 0.067 (raw) to 0.0035 (calibrated). per word, and more piecesECE (Expected Calibration Error). A metric that measures how well a model's confidence scores align with its actual accuracy. Lower is better. Mailwoman's held-out ECE drops from 0.067 (raw) to 0.0035 (calibrated). per word means a longer sequence, more chances for two piecesECE (Expected Calibration Error). A metric that measures how well a model's confidence scores align with its actual accuracy. Lower is better. Mailwoman's held-out ECE drops from 0.067 (raw) to 0.0035 (calibrated). of one word to disagree, and less of the fixed 128-piece input window left for the rest of the address. The cost of thin coveragecoverageThe fraction of a population or region for which a data source has real, non-placeholder entries — e.g. 47% rooftop coverage on Texas addresses. Distinct from accuracy on the rows that are present. shows up here first, before it shows up as a wrong labelcomponent tagOne of the 25 labels in Mailwoman's address schema — street, locality, region, postcode, house_number, unit, po_box, country, venue, intersection, and others. Each parsed span carries exactly one component tag..

The vote is per word, not per name. The consistency pass keeps one word internally coherent and deliberately stops there. Multi-word names are left to the decoderdecoderIn a transformer encoder-decoder model, the part that produces output sequences. Mailwoman's classifier is encoder-only (no decoder); the 'CRF decoder' is a different thing — a structured-prediction layer that picks the best label sequence from the encoder's outputs. and the priors, because a rule that forced agreement across a whitespace boundary would merge a localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighborhood in the hierarchy. into the streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. in front of it.

Related

• How a model reads an address — the whole chain, on one input.
• Decoding and the best path — the constraint that rules out an orphan continuation.
• Component tags — the 25 tags, and which the shipped modelneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.' emits.
• Library API — where spansspanA contiguous range of characters or tokens in the input string, tagged with an address component type (street, locality, postcode, etc.). Parsed addresses are represented as collections of spans, possibly nested in a tree., offsets and options live in code.