Skip to main content

BIO labels

BIO labelling is the trick that lets a tokentokenOne 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. classifier (which decides one tokentokenOne 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. at a time) emit 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. (groups of consecutive 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. that mean one thing together). It is the standard approach for sequence labelling tasks in NLP — Named Entity Recognition, part-of-speech tagging, address parsingaddress 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..

This article explains the scheme, why it works, and a failure mode (the orphan-Iorphan-IA label-sequence bug where I-X appears without a matching preceding B-X (e.g. O, I-locality). Structurally invalid in BIO; the CRF prevents it.) that motivated the decode-time fix covered in How Mailwoman parses an address. For the full walkthrough of 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., start there; this page is the deep dive on the labelcomponent tagOne of the 33 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 itself.

The scheme​

Each tokentokenOne 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. gets exactly one labelcomponent tagOne of the 33 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 labelcomponent tagOne of the 33 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. is either:

  • O — the tokentokenOne 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. is outside any tagged 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..
  • B-X — the tokentokenOne 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. is the beginning of an X 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..
  • I-X — the tokentokenOne 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. is inside a continuing X 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..

So a 3-tokentokenOne 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. "Saint Petersburg, FL" labelled correctly looks like:

Saint          → B-locality
Petersburg     → I-locality
,              → O
FL             → B-region

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. "Saint Petersburg" is signalled by one B-locality followed by one I-locality. 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. reconstructs 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. by walking the labelscomponent tagOne of the 33 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.: when it sees B-X, it starts a new 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.; while it keeps seeing I-X with a matching tag, it extends 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.; on O or a different tag, it closes 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..

The full Mailwoman vocabulary​

The current active set (StagestageOne of the dataflow stages in the runtime pipeline (normalize, locale gate, kind classify, phrase group, token classify, sequence correct, reconcile, resolve). Distinct from tier (model vocabulary) and phase (plan milestone). 3, shipped since v0.6.0) has 16 component tagscomponent tagOne of the 33 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., each split into {B-, I-} plus one shared O — 16 × 2 + 1 = 33 BIO labelsBIO tagging (Begin-Inside-Outside). A token-level labeling scheme where each token is tagged as B-X (beginning of an entity of type X), I-X (inside an entity of type X), or O (outside any entity). Mailwoman uses BIO over SentencePiece tokens to annotate address components., matching neural-weights-en-us/model-card.json's num_labels: 33:

labelcomponent tagOne of the 33 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.example tokentokenOne 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.
O",", "in"
B-country, I-country"United" "States"
B-region, I-region"New" "York" (stateregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality., when verbose)
B-locality, I-locality"Saint" "Petersburg"
B-dependent_locality, I-dependent_locality"Greenpoint"
B-postcode, I-postcode"10118", "75008"
B-subregion, I-subregion"Brooklyn"
B-cedex, I-cedex"CEDEX" "08" (FR-specific)
B-venue, I-venue"Wrigley" "Field"
B-street, I-street"5th" (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.'s base name)
B-house_number, I-house_number"350", "10" "bis"
B-street_prefix, I-street_prefix"SE" (directional, before the name)
B-street_suffix, I-street_suffix"St", "Ave" (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. type)
B-unit, I-unit"Apt" "3"
B-po_box, I-po_box"PO" "Box" "4"
B-intersection_a, I-intersection_a"5th" "Ave" (first cross streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels.)
B-intersection_b, I-intersection_b"42nd" "St" (second cross streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels.)

The neural 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 final classifier 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. has 33 outputs; 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.' picks a score for every labelcomponent tagOne of the 33 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. at every position (an emission), 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. — not raw per-tokentokenOne 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. argmax — turns those scores into a labelcomponent tagOne of the 33 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. sequence.

In code:

const STAGE3_TAGS = [
	"country",
	"region",
	"locality",
	"dependent_locality",
	"postcode",
	"subregion",
	"cedex",
	"venue",
	"street",
	"house_number",
	"street_prefix",
	"street_suffix",
	"unit",
	"po_box",
	"intersection_a",
	"intersection_b",
]
const STAGE3_BIO_LABELS = ["O", ...STAGE3_TAGS.flatMap((t) => [`B-${t}`, `I-${t}`])]

The last 6 tags — street_prefix through intersection_b — are the StagestageOne of the dataflow stages in the runtime pipeline (normalize, locale gate, kind classify, phrase group, token classify, sequence correct, reconcile, resolve). Distinct from tier (model vocabulary) and phase (plan milestone). 3 expansion, shipped in v0.6.0: decomposing the monolithic street tag into prefix/base/suffix and adding unit, po_box, and paired intersectionintersectionAn address that names a location by two crossing streets ('5th & Main') rather than a number and street. Mailwoman tags the two streets as intersection_a and intersection_b — a negative-space format that starved the early model. tags. labels.py keeps the earlier StagestageOne of the dataflow stages in the runtime pipeline (normalize, locale gate, kind classify, phrase group, token classify, sequence correct, reconcile, resolve). Distinct from tier (model vocabulary) and phase (plan milestone). 1 and StagestageOne of the dataflow stages in the runtime pipeline (normalize, locale gate, kind classify, phrase group, token classify, sequence correct, reconcile, resolve). Distinct from tier (model vocabulary) and phase (plan milestone). 2 vocabularies (7 and 10 tags) defined alongside StagestageOne of the dataflow stages in the runtime pipeline (normalize, locale gate, kind classify, phrase group, token classify, sequence correct, reconcile, resolve). Distinct from tier (model vocabulary) and phase (plan milestone). 3 so historical checkpointscheckpointA saved snapshot of the model weights and optimizer state during training. Mailwoman saves a checkpoint periodically so training can resume after a GPU hang. and evalevalRunning the model against a held-out golden dataset and computing per-component F1, exact-match, calibration, and resolved-coordinate error. reports can still be diffed against older labelcomponent tagOne of the 33 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. sets — but StagestageOne of the dataflow stages in the runtime pipeline (normalize, locale gate, kind classify, phrase group, token classify, sequence correct, reconcile, resolve). Distinct from tier (model vocabulary) and phase (plan milestone). 3 is what every current 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. config and 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.' uses.

The orphan-I problem​

A naive tokentokenOne 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. classifier picks the highest-probability labelcomponent tagOne of the 33 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. for each tokentokenOne 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. independently. This produces sequences like:

Saint    → O                        ← the model wasn't sure, picked O
Petersburg → I-locality             ← the model was confident, picked I-locality

The result is structurally invalid. An I-locality is by definition "inside a localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood 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.", and the previous tokentokenOne 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. is not in a localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood 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.. There is no B-locality to be inside of. This is called an orphan-Iorphan-IA label-sequence bug where I-X appears without a matching preceding B-X (e.g. O, I-locality). Structurally invalid in BIO; the CRF prevents it..

What happens when 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. reconstructs 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.? Depending on how it handles the orphan-Iorphan-IA label-sequence bug where I-X appears without a matching preceding B-X (e.g. O, I-locality). Structurally invalid in BIO; the CRF prevents it.:

  • Strict mode — drop the orphan. "Saint Petersburg" becomes "Petersburg" (a 1-tokentokenOne 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. localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy. starting at Petersburg). This is the "Saint Petersburg → Petersburg" bug visible in Mailwoman v0.2.0.
  • Forgiving mode — treat the orphan-Iorphan-IA label-sequence bug where I-X appears without a matching preceding B-X (e.g. O, I-locality). Structurally invalid in BIO; the CRF prevents it. as a B-X. "Saint Petersburg" becomes two adjacent localities. Worse.

Neither is what the data wants: B-locality, I-locality.

How Mailwoman fixes this​

The fix is decode-time, not learned. A fixed BIO-validity table pins the impossible transitions — O → I-X, B-X → I-Y where X ≠ Y — to negative infinity, and the Viterbi algorithmViterbi decodingA dynamic programming algorithm that finds the most likely sequence of hidden states (labels) given a sequence of observations (token emissions). Mailwoman uses Viterbi over a linear-chain CRF to produce globally coherent BIO label sequences from per-token model scores. searches for the highest-scoring labelcomponent tagOne of the 33 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. sequence that respects the table, rather than picking each tokentokenOne 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.'s labelcomponent tagOne of the 33 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. independently. The orphan-Iorphan-IA label-sequence bug where I-X appears without a matching preceding B-X (e.g. O, I-locality). Structurally invalid in BIO; the CRF prevents it. is structurally excluded: there's no valid path through it.

The result: 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.' that is uncertain about "Saint" between O and B-locality will still produce a structurally valid sequence at decode time. "Saint Petersburg" comes out as one localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood 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.. See CRF decoder for why this is narrower than a full learned CRFCRF (Conditional Random Field). A statistical modeling method that predicts structured outputs by modeling dependencies between adjacent labels. Mailwoman uses a linear-chain CRF as the Viterbi decoder at inference time to enforce BIO label consistency — a B-street must be followed by I-street or O, never I-locality., and what that page's name still refers to.

The other reason BIO works well​

BIO labelsBIO tagging (Begin-Inside-Outside). A token-level labeling scheme where each token is tagged as B-X (beginning of an entity of type X), I-X (inside an entity of type X), or O (outside any entity). Mailwoman uses BIO over SentencePiece tokens to annotate address components. are simple enough that:

  • 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.' architecture stays small. No special 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.-prediction headsattention 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. — just a per-tokentokenOne 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. classifier.
  • 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. data is easy to generate. Given (raw, components) from a 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. adapter, you align each component's text to its 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. and emit B- for the first tokentokenOne 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. and I- for the rest.
  • EvaluationevalRunning the model against a held-out golden dataset and computing per-component F1, exact-match, calibration, and resolved-coordinate error. is straightforward. Compare predicted 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. to gold 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.; compute precisionprecisionOf the spans the model labeled as a given tag, the fraction it got right. High precision means few false positives. Paired with recall to compute F1., recallrecallOf the spans whose gold label is a given tag, the fraction the model found. High recall means few misses. Paired with precision to compute F1., F1 per tag.

This pattern is one of the most-tested approaches in NLP. CoNLL-2003 (the canonical Named Entity Recognition benchmark) uses BIO. CoNLL-2000 (chunking) uses BIO. Every modern NER tool exposes BIO-style output as the default. Mailwoman inheriting this standard means our 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. data is interoperable and our evaluationevalRunning the model against a held-out golden dataset and computing per-component F1, exact-match, calibration, and resolved-coordinate error. tooling is familiar.

Where this lives in the code​

  • Labelcomponent tagOne of the 33 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. 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.: corpus-python/src/mailwoman_train/labels.py (STAGE3_BIO_LABELS, ACTIVE_BIO_LABELS)
  • TypeScript mirror: core/types/component.ts (BIO_LABELS)
  • 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 alignmentalignmentThe step in the corpus pipeline that takes a (raw, components) pair from an adapter and produces a (raw, tokens, BIO labels) row by finding each component's text inside the raw string and labeling the matching tokens.: corpus/src/align.ts (turns (raw, components) into per-tokentokenOne 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. BIO labelsBIO tagging (Begin-Inside-Outside). A token-level labeling scheme where each token is tagged as B-X (beginning of an entity of type X), I-X (inside an entity of type X), or O (outside any entity). Mailwoman uses BIO over SentencePiece tokens to annotate address components.)
  • InferenceinferenceRunning the trained model on new input to get predictions, as opposed to training, which produces the model. In Mailwoman that means a small transformer encoder reads an address string and classifies every token — house number, street, locality, region, postcode, and the rest. A Who's On First gazetteer can feed soft location hints into the pass, but the model makes the final call on every label. Where a generative model writes text token by token, Mailwoman's output is a retrieval-augmented token classification: one label per input piece.-time decoding: neural/viterbi.ts (buildBIOTransitionMask, viterbi) — see CRF decoder

See it in action​

Loading demo embed…

Expand the BIO labelsBIO tagging (Begin-Inside-Outside). A token-level labeling scheme where each token is tagged as B-X (beginning of an entity of type X), I-X (inside an entity of type X), or O (outside any entity). Mailwoman uses BIO over SentencePiece tokens to annotate address components. section to see the word-level BIO labelcomponent tagOne of the 33 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. breakdown for a live 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..

See also​

  • How Mailwoman parses an address — the full 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. this labelcomponent tagOne of the 33 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 plugs into
  • CRF decoder — the structural-validity 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.
  • Tokenization — what gets labelled
  • Training pipeline — how BIO labelsBIO tagging (Begin-Inside-Outside). A token-level labeling scheme where each token is tagged as B-X (beginning of an entity of type X), I-X (inside an entity of type X), or O (outside any entity). Mailwoman uses BIO over SentencePiece tokens to annotate address components. become 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. data