(work in progress - for details, see the proposal.)
Install requirements:
$ pip install -r requirements.txt
Setup the config as necessary:
$ cp config-sample.py config.py; vi config.py
Download the necessary corpora:
$ python -m textblob.download_corpora
You will need to prepare some data:
$ python prep.py train_phrases
$ python prep.py train_idf
You can use any corpus to train these on; I used the body text of about 120k NYT articles and it has worked well. The more, the better, most likely.
These are used to better identify phrases in text and to have some notion of salience (inverse document frequency).
Run the server:
$ python server.py
Then try out the demo:
localhost:5001/
If you are developing and need to reload Geiger a lot, you are in for a bad time. The phrase, IDF, and Word2Vec models take a very long time to load.
Fortunately, things are setup so that you can run each of these in their own separate processes, which don't need to be reloaded.
If you set remote=True in config.py, the functions which rely on the phrase and Word2Vec models will call out to these
separate processes instead of loading the models directly. Just make sure you set remote=False when deploying for production.
Then you can run these processes separately like so:
$ python dev.py word2vec
$ python dev.py phrases
$ python dev.py idf
The downside is that calling out to separate processes like this slows the usage of these models, but you'll likely be saving time overall.