Intro

This is a simple classifier for recognizing names of People from random text. I try four different classification methods from Scikit and also provide an API to try out or use as a service.

Usage

The project contains three scripts

data.py for downloading and forming the data set
main.py for training the model
api.py for serving the model as a web service.

Other than that I have also added the data.csv file, which can also be downloaded from DBPedia using the data.py script, but it was IMHO small enough (33mb) to put up on github for other people to use as is.

Environment

To run the project for the first time, using python 2.7, run the following commands:

pip install virtualenv

cd project folder

virtualenv venv

source venv/bin/activate

pip install -r requirements.txt

mkdir model // For storing the models

After that you can deactivate venv by calling

deactivate

and re-activate it whenever you want by just running

source venv/bin/activate

Getting the data

I have uploaded the data.csv file that has the name data, and random text data comes from the newsgroups of sklearn. The name data was downloaded from DBPedia using the data.py script.

python data.py

Training the model

To train the model run

python main.py

This trains the NB classifier and saves the model in pickle files. The confusion matrix and accuracy score is printed in the terminal. You can uncomment more models inside the code to add them to the process.

Running the api

I have also added a small flask api to try out the model with user inserted strings.You can start it using:

python api.py

Then you can test it at http://127.0.0.1:5000/classify using simple get requests from a browser or a tool like Postman. Give it a query in the url query parameters and it will return the probability that the string you provided is the name of a person or not.

ex.

http://127.0.0.1:5000/classify?q="John Smith"
result: {"multi_nb": [0.020987041619013835, 0.9790129583809848]}

http://127.0.0.1:5000/classify?q="Hello World"
{"multi_nb": [0.9917522714060827, 0.008247728593917876]}

Issues:

Currently only the random forest classifier seems to be able to correctly classify gibberish (ex. "fdsafdsafdsafdsafdas"). I tried character n-grams and word n-grams. The character n-grams perform worse than random. Word n-gram do not help. One solution could be to introduce completely random text in the training set. Another could be to label non dictionary words as negative.