locationformatter 0.1.3

A dual-head NER-based parser for location strings

decide-location-formatter

A Python package for parsing and structuring ocation strings into their individual address components. Built around a dual-head NER model (svercoutere/abb-dual-location-component-ner) fine-tuned on top of XLM-RoBERTa base.

How it works

Raw location strings like "Scaldisstraat 23-25, 2000 Antwerpen" or "Cafe den Draak, Lovegemlaan 7, 9000 Gent" are common in municipal decision text but are inconsistently formatted and often contain multiple distinct locations in a single string.

The pipeline has three steps:

1. Text cleaning — normalises whitespace, unicode, and newlines.
2. Dual-head NER inference — the model runs two independent CRF-decoded classification heads over every token simultaneously:
   • Component head — tags each token as one of 12 address component types (street, city, postcode, …).
   • Location head — groups tokens that belong to the same physical location into B-LOCATION / I-LOCATION spans, allowing multi-location strings to be split.
3. Post-processing — component spans are nested inside their parent location spans, housenumber ranges/sequences (e.g. 23-25, 7 en 9) are expanded into individual entries, and bus numbers are split into a separate field.

Architecture

Component	Detail
Base encoder	xlm-roberta-base (12 layers, 768 hidden)
Component head	Linear(768 → 256) · GELU · Dropout(0.1) · Linear(256 → 25) + CRF
Location head	Linear(768 → 256) · GELU · Dropout(0.1) · Linear(256 → 3) + CRF
Tokenisation	Word-level regex tokeniser; sub-word alignments via fast tokenizer word_ids()
Max input length	256 sub-word tokens

Entity types (component head)

Label	Description
STREET	Street name (no house number)
ROAD	Road or route name
HOUSENUMBER	House/building number(s), ranges or sequences
POSTCODE	Postal or ZIP code
CITY	City or municipality name
PROVINCE	Province or region name
BUILDING	Named building, site or facility
INTERSECTION	Crossing or intersection of roads
PARCEL	Land parcel, section or lot number
DISTRICT	District, neighbourhood or borough
GRAVE_LOCATION	Plot/row/number within a cemetery
DOMAIN_ZONE_AREA	Domain, zone or area name

Evaluation

Evaluated on a held-out 10 % split of ~10 000 Belgian municipal decision location strings.

Metric	Score
Combined F1	0.9435
Component F1	0.9295
Location F1	0.9576

Installation

From source (recommended during development)

git clone https://github.com/semantic-ai/decide-location-formatter.git
cd decide-location-formatter
pip install -e .

Dependencies only

pip install torch>=2.0 transformers>=4.35 pytorch-crf>=0.7.2

The model weights (~1 GB) are downloaded automatically from the Hugging Face Hub on first use.

Usage

Quick start

from locationformatter import LocationFormatter

lf = LocationFormatter()   # loads model once; reuse for many calls

result = lf.parse("Scaldisstraat 23-25, 2000 Antwerpen")
print(result)

{
  "original": "Scaldisstraat 23-25, 2000 Antwerpen",
  "locations": [
    {
      "location": "Scaldisstraat 23-25, 2000 Antwerpen",
      "street": "Scaldisstraat",
      "housenumber": "23",
      "housenumber_type": "single",
      "postcode": "2000",
      "city": "Antwerpen"
    },
    {
      "location": "Scaldisstraat 23-25, 2000 Antwerpen",
      "street": "Scaldisstraat",
      "housenumber": "25",
      "housenumber_type": "single",
      "postcode": "2000",
      "city": "Antwerpen"
    }
  ]
}

Multi-location strings

Strings that contain several distinct locations are automatically split:

result = lf.parse("Lovegemlaan 7, 9000 Gent en Dorpstraat 12, 9240 Zele")
for loc in result["locations"]:
    print(loc)

Raw prediction (no housenumber expansion)

predict() returns spans straight from the model without expanding ranges or splitting bus numbers:

raw = lf.predict("Heikeesstraat 2-4, 9240 Zele")
# raw["locations"][0]["housenumber"] == "2-4"
# raw["locations"][0]["housenumber_type"] == "range"

One-shot helper

For a single call without keeping the model in memory:

from locationformatter import parse_location

result = parse_location("Grote Markt 1, 2000 Antwerpen")

Note: parse_location reloads the model on every call. Use LocationFormatter for repeated parsing.

Custom model or device

lf = LocationFormatter(
    repo="your-org/your-model",   # any compatible HF Hub repo
    device="cuda",                # "cpu" or "cuda"; auto-detected when omitted
)

API reference

LocationFormatter

class LocationFormatter:
    def __init__(self, repo: str = "svercoutere/abb-dual-location-component-ner",
                 device: str | None = None): ...

    def parse(self, text: str) -> dict: ...
    # Full pipeline: clean → NER → expand housenumbers.
    # Returns {"original": str, "locations": list[dict]}

    def predict(self, text: str) -> dict: ...
    # NER only, no housenumber expansion.
    # Returns {"original": str, "locations": list[dict]}

Helper functions

from locationformatter import clean_string, clean_house_number, extract_house_and_bus_number

clean_string("  Grote   Markt\n1  ")
# → "Grote Markt 1"

clean_house_number("3 t.e.m. 7")
# → ["3", "4", "5", "6", "7"]

clean_house_number("10-14")
# → ["10", "11", "12", "13", "14"]

extract_house_and_bus_number("5 bus 3")
# → {"housenumber": "5", "bus": "3"}

Output schema

Each entry in the locations list is a flat dict. Only fields detected by the model are included.

Field	Type	Description
location	str	The substring corresponding to this location
street	str	Street name
road	str	Road/route name
housenumber	str	Individual house number (after expansion)
housenumber_type	str	"single", "range", or "sequence"
bus	str	Bus/apartment number (when present)
postcode	str	Postal code
city	str	City or municipality
province	str	Province
building	str	Named building or facility
intersection	str	Road intersection
parcel	str	Land parcel identifier
district	str	District or neighbourhood
grave_location	str	Cemetery plot/row/number
domain_zone_area	str	Zone or area name

Development

Running tests

pytest tests/

The unit tests for the helper functions (clean_string, clean_house_number, extract_house_and_bus_number) do not require the model to be loaded and run offline.