To develop a statistical NER system the following steps are needed:
- define what entities you are interested in;
- get some training/testing data (annotated webpages);
- define what features should be extracted;
- develop a statistical model that uses these features to produce the output.
- read annotated HTML data produced by WebAnnotator or GATE;
- transform HTML trees into “HTML tokens”, i.e. text tokens with information about their position in HTML tree preserved;
- extract various features from these HTML tokens, including text-based features, html-based features and gazetteer-based features (GeoNames support is built-in);
- convert these features to the format sequence labelling toolkits accept;
- use Wapiti or CRFSuite CRF toolkit for entity extraction (helpers for other toolkits like seqlearn are planned);
- group extracted entites using an unsupervised algorithm;
- embed annotation results back into HTML files (using WebAnnotator format), allowing to view them in a web browser and fix using visual tools.
Unlike most NER systems, webstruct works on HTML data, not only on text data. This allows to define features that use HTML structure, and also to embed annotation results back into HTML.
Webstruct requires Python 2.7 or Python 3.3+.
To install Webstruct, use pip:
$ pip install webstruct
Webstruct depends on the following Python packages:
Note that these dependencies are not installed automatically
pip install webstruct is executed.
There are also some optional dependencies:
- Webstruct has support for Wapiti sequence labelling toolkit;
you’ll need both the
wapitibinary and python-wapiti wrapper (from github master) for the tutorial.
- For training data annotation you may use WebAnnotator Firefox extension.
- Code for preparing GeoNames gazetteers uses marisa-trie and pandas.