Customization

Isolated registries

Expander(...) is a small mutable facade around a new, isolated registry:

from abbr2words import Expander

expander = Expander("en")
expander.add("Tech.", "Technology")
assert expander("Tech. works") == "Technology works"

get_expander(...) also returns a new isolated implementation registry. Changes to either form do not affect the shared convenience API.

Entries can be case-sensitive and can use regular-expression guards:

expander.add("in.", "inch", only_if_preceded_by=r"\d\s*$")
expander.add("KI", "Künstliche Intelligenz", case_sensitive=True)

Entries use fixed expansion text by default. For reviewed lexical phrases that should compose with sentence position, opt into sentence casing:

expander.add("Pág.", "página", case_policy="sentence")
assert expander("Pág. 12") == "Página 12"
assert expander("consulte la Pág. 12") == "consulte la página 12"

The sentence policy uppercases the first cased character only at input or after sentence-ending punctuation, including an opening quote or bracket that follows that boundary. A colon is not a sentence boundary. Dotted abbreviation matches preserve one final period when their consumed dot is sentence-final; commas, semicolons, and internal dots are not added or moved.

Ambiguous English dotted forms

The English registry prefers reversible letter readings when a dotted spelling could be either a semantic abbreviation or a person’s initials. The structural fallback recognizes standalone uppercase forms with two through eight dotted letters, such as E.D. and F.C.S.C.J., after registered abbreviations and reviewed units have had priority. Thus lowercase e.g. expands to for example, while uppercase E.G. is rendered as E G; I.D. is rendered as I D.

Single-letter N., S., E., and W. use letter defaults and require bounded address/street evidence before expanding to a direction. This protects personal initials and biological names such as S. aureus. D.C. expands to District of Columbia only with explicit Washington/place evidence; L.A. remains L A in this layer so author initials are not rewritten as a city name.

Followed-by guards are evaluated against the suffix immediately after the candidate abbreviation. In only_if_followed_by=r"^\s*\d", ^ therefore means “immediately after this abbreviation,” even when the candidate occurs after other source text. Preceded-by guards continue to use their bounded prefix window.

Registration is validated immediately. Abbreviations and expansions must be non-empty strings; context keys and values, POS labels, booleans, and guard patterns have stable type/value checks. Guard regexes are compiled once and are trusted application configuration; do not pass arbitrary untrusted regex text because the standard-library engine has no portable timeout.

Boundaries use (?<!\w) and (?!\w). Exact case-sensitive custom entries outrank bundled exact entries and case-insensitive fallbacks, so conflicts do not depend on registration order. A preceding guard must end immediately before the abbreviation after horizontal whitespace is removed.

POS guards are optional and are evaluated only when source-aligned annotations are supplied:

from abbr2words import Expander, TokenAnnotation

expander = Expander("en")
expander.add("Ref.", "Reference", only_if_pos="NOUN")
assert expander.expand("Ref.", annotations=[TokenAnnotation(0, 4, "NOUN")]) == "Reference"

Pass a collection when several labels are accepted. A deny constraint wins if both sets match:

expander.add("Code.", "Code", only_if_pos={"NOUN", "PROPN"}, not_if_pos="PROPN")

Structural guards and reviewed numeric unit matching run before POS guards. POS output is treated as an optional signal; missing labels do not veto an entry, and a general tagger cannot disable a valid numeric unit expression.

Abbreviation expansion remains lexical. Number/date/decimal/identifier parsing, article contraction, surrounding grammar, and TTS-oriented rendering belong to the downstream normalizer, such as spokenform.

Shared registries

get_shared_expander(...) returns the process-wide registry used by abbr2words(...). Shared custom entries therefore affect later calls in the same process. Shared registries are scoped by both normalized language and context mode, so context=True and context=False are separate mutable registries.

from abbr2words import get_shared_expander, reset_expanders

shared = get_shared_expander("de", context=True)
shared.add_custom_abbreviation("KI", "Künstliche Intelligenz")

# Remove shared customizations during test teardown or application reset.
reset_expanders("de")

reset_expanders() removes all shared language registries, while reset_expanders("de") limits cleanup to one language. Shared state is local to the current process and is not a synchronization mechanism between threads or processes; applications should coordinate concurrent mutation themselves.

Shared lookup and reset are atomic. At the start of each expansion, the expander captures a complete snapshot of its abbreviation entries, unit overrides, and suppressed-unit set; later mutations affect the next expansion, not the one already in progress. Applications should still avoid mutating a shared registry while a long expansion is running because registry mutation is not itself coordinated with other application-level state.

Unit customization

Units are a separate reviewed inventory, not ordinary abbreviation entries. Use the instance-local unit methods when an application needs an override:

expander.set_unit("kg", "custom kilogram")
expander.remove_unit("kg")

Calling add("kg", ...) or abbreviation removal for a known unit raises a clear error rather than silently changing a registry that expansion ignores. Unit overrides do not leak between isolated expanders.

set_unit() retains the reviewed canonical ID when replacing a bundled symbol. Pass canonical_id="..." to explicitly assign a different identity, or omit it for a new user-defined symbol. The optional category distinguishes custom units from currencies or magnitudes. remove_unit() accepts either a symbol or a canonical ID; removing an ID suppresses all bundled aliases for that identity. The same rules are available through Expander.iter_unit_matches().

For semantic consumers, prefer the structured matcher over parsing replacement strings. Numeric quantity matching has priority over lexical abbreviation matching, so numeric 1 Mio. produces one structured magnitude claim while standalone Mio. retains its ordinary German abbreviation behavior.

Finite aliases and exact replacements

Bundled entries can have finite aliases for reviewed formatting variants. For example, the German registry accepts z.B., z. B., z . b ., and zB with the same boundary policy and replacement metadata. Aliases are registry data, not global regular-expression substitutions, so attached strings such as pizzaB, ModellzB12, and du.a.test remain unchanged.

Use the replacement result when a caller needs semantic provenance instead of reconstructing edits with a text diff:

result = expander.expand_with_replacements("Prof. Klein, S. 12")
for replacement in result.replacements:
    print(replacement.source, replacement.start, replacement.end)