A complete example (contact extraction). See GH-24.

Author: kmike

.gitignore

@@ -51,4 +51,7 @@ notebooks/*.marisa
 notebooks/*.wapiti
 notebooks/*.crfsuite
 webstruct_data/corpus/random_pages/wa/*.html
-webstruct_data/corpus/us_contact_pages/cleaned
+webstruct_data/corpus/us_contact_pages/cleaned
+example/_data/*
+example/*.joblib
+example/*.html

example/README.rst

@@ -0,0 +1,44 @@
+Contact extraction using Webstruct
+==================================
+
+This repository contains code to train a model for contact and address
+extraction. The result is a .joblib file with pickled webstruct.NER object.
+
+Currently the example requires Python 3.5+
+
+Training
+--------
+
+To train a model, first build gazetteers using built_gazetteers script::
+
+    python3 -m ner.build_gazetteers
+
+It will create "_data" folder with city/state geonames data. The script uses
+several GBs or RAM.
+
+To train a CRF model run::
+
+    python3 -m ner.train
+
+The model uses training data from opensource webstruct package
+(mostly contact pages of US, CA and GB small business websites)
+and provides 'ORG', 'TEL', 'FAX', 'HOURS', 'STREET', 'CITY', 'STATE',
+'ZIPCODE', 'COUNTRY', and 'SUBJ' entities.
+
+Script should produce "contact-extractor.joblib" file with a saved
+webstruct.NER object and "crf-features.html" file with debugging
+information about the model.
+
+Usage
+-----
+
+To use the saved model code in this repository is not needed.
+Make sure joblib, sklearn-crfsuite and webstruct are installed,
+then load the model::
+
+    import joblib
+    ner = joblib.load('contact-extractor.joblib')
+    print(ner.extract_groups_from_url('<some URL>'))
+
+See https://webstruct.readthedocs.io/en/latest/ref/model.html#webstruct.model.NER
+for the API.

example/ner/__init__.py

@@ -0,0 +1 @@
+# -*- coding: utf-8 -*-

example/ner/build_gazetteers.py

@@ -0,0 +1,88 @@
+#!/usr/bin/env python3
+import argparse
+from pathlib import Path
+
+import requests
+
+from webstruct.gazetteers.geonames import read_geonames_zipped, to_dawg
+
+
+FILES = [
+    ('http://download.geonames.org/export/dump/allCountries.zip', 'allCountries.zip'),
+    ('http://download.geonames.org/export/dump/cities1000.zip', 'cities1000.zip'),
+    ('http://download.geonames.org/export/dump/cities5000.zip', 'cities5000.zip'),
+    ('http://download.geonames.org/export/dump/cities15000.zip', 'cities15000.zip'),
+]
+
+DATA_ROOT = Path('_data')
+
+
+def download_geonames():
+    """ Download geonames files if they don't exist in ./_data folder. """
+    DATA_ROOT.mkdir(exist_ok=True)
+    for url, name in FILES:
+        path = (DATA_ROOT / name)
+        if path.exists():
+            continue
+        print("downloading {}".format(url))
+        path.write_bytes(requests.get(url).content)
+
+
+def _compile_cities(path: Path, lowercase: bool=False):
+    out_path = path.with_suffix('.dafsa')
+    # if out_path.exists():
+    #     return
+    print("reading {}".format(path))
+    df = read_geonames_zipped(str(path))
+    if lowercase:
+        df = _to_lower(df)
+    print("compiling {}".format(out_path))
+    dawg = to_dawg(df)
+    dawg.save(str(out_path))
+
+
+def _to_lower(df):
+    return df.assign(
+        main_name=df.main_name.str.lower(),
+        asciiname=df.asciiname.str.lower(),
+        alternatenames=df.alternatenames.str.lower(),
+    )
+
+
+def _read_full():
+    path = DATA_ROOT / 'allCountries.zip'
+    print("reading {}".format(path))
+    return read_geonames_zipped(str(path))
+
+
+def _compile_adm(df):
+    codes = ['ADM1', 'ADM2', 'ADM3', 'ADM4']
+    out_paths = [DATA_ROOT / "{}.dafsa".format(code.lower()) for code in codes]
+    # if all(p.exists() for p in out_paths):
+    #     return
+    for code, out_path in zip(codes, out_paths):
+        # if out_path.exists():
+        #     continue
+        print("compiling {}".format(out_path))
+        df_adm = df[df.feature_code == code]
+        dawg = to_dawg(df_adm)
+        dawg.save(str(out_path))
+
+
+def compile_gazetteers_contacts(lowercase=False):
+    """ Compile geonames data downloaded by ``download_geonames``. """
+    for name in ['cities1000.zip', 'cities5000.zip', 'cities15000.zip']:
+        _compile_cities(DATA_ROOT / name, lowercase=lowercase)
+    df = _read_full()
+    if lowercase:
+        df = _to_lower(df)
+    _compile_adm(df)
+
+
+if __name__ == '__main__':
+    p = argparse.ArgumentParser()
+    p.add_argument('--lower', action="store_true")
+    args = p.parse_args()
+
+    download_geonames()
+    compile_gazetteers_contacts(args.lower)

example/ner/cv.py

@@ -0,0 +1,26 @@
+# -*- coding: utf-8 -*-
+import numpy as np
+
+
+def crf_cross_val_predict(pipe, X, y, cv, groups=None, n_folds=None):
+    """
+    Split data into folds according to cv iterator, do train/test prediction 
+    on first n_folds (or on all folds if n_folds is None). 
+    """
+    X, y = np.array(X), np.array(y)
+    y_pred = []
+    y_true = []
+
+    for idx, (train_idx, test_idx) in enumerate(cv.split(X, y, groups)):
+        if n_folds and idx >= n_folds:
+            break
+
+        X_train, X_dev = X[train_idx], X[test_idx]
+        y_train, y_dev = y[train_idx], y[test_idx]
+        pipe.fit(X_train, y_train, X_dev=X_dev, y_dev=y_dev)
+        y_true.append(y_dev)
+        y_pred.append(pipe.predict(X_dev))
+
+    y_pred = np.hstack(y_pred)
+    y_true = np.hstack(y_true)
+    return y_pred, y_true

example/ner/data.py

@@ -0,0 +1,56 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+from itertools import chain
+from pathlib import Path
+from typing import List, Tuple, Any, Set
+
+import webstruct
+
+from .utils import pages_progress
+
+
+WEBSTRUCT_DATA = Path(__name__).parent / ".." / "webstruct_data"
+GAZETTEER_DATA = Path(__name__).parent / "_data"
+
+
+KNOWN_ENTITIES = [
+    'ORG', 'TEL', 'FAX', 'HOURS',
+    'STREET', 'CITY', 'STATE', 'ZIPCODE', 'COUNTRY',
+    'EMAIL', 'PER', 'FUNC', 'SUBJ'
+]
+CONTACT_ENTITIES = [
+    'ORG', 'TEL', 'FAX', 'HOURS',
+    'STREET', 'CITY', 'STATE', 'ZIPCODE', 'COUNTRY',
+    'SUBJ',
+]
+ADDRESS_ENTITIES = [
+    'STREET', 'CITY', 'STATE', 'ZIPCODE', 'COUNTRY',
+]
+
+
+def load_webstruct_data() -> List:
+    """
+    Load training data from webstruct repository.
+
+    It is a mess: there are two folders which have OK data, one
+    is stored in WebAnnotator format, another is stored in GATE format.
+    """
+    wa_loader = webstruct.WebAnnotatorLoader(known_entities=KNOWN_ENTITIES)
+    gate_loader = webstruct.GateLoader(known_entities=KNOWN_ENTITIES)
+
+    trees1 = webstruct.load_trees(
+        str(WEBSTRUCT_DATA / "corpus/business_pages/wa/*.html"),
+        loader=wa_loader,
+    )
+
+    trees2 = webstruct.load_trees(
+        str(WEBSTRUCT_DATA / "corpus/us_contact_pages/annotated/*.xml"),
+        loader=gate_loader
+    )
+    trees = chain(trees1, trees2)
+    return list(pages_progress(trees, desc="Loading webstruct default annotated data"))
+
+
+def load_countries() -> Set[str]:
+    countries_path = WEBSTRUCT_DATA / 'gazetteers/countries/countries.txt'
+    return set(countries_path.read_text().splitlines())