In this project, we use the power of machine learning techniques to classify the sentences into active or passive.

To perform the classification task, we use the Naive Bayes algorithm. The training data comes from the "2nd_Gore-Bush.txt" file. Since the data is not labeled, in the first stage, we use the parts of speech tags. The sentence is tagged passive if a past participle verb succeeds by a "to be" form.

Naive Bayes algorithm requires an input feature that specifies the required information. Since we only need grammatical representation for this task, for each sentence parts of speech are extracted. If a verb is in the "to be" form, that tag is replaced with the "BE" tag.

The model consists of a count vectorizer followed by a Bayesian classifier.

The experiments are conducted over five folds of cross-validation To select the best hyperparameters. Then the best model based on the f1 measure is stored to be used for the prediction.

The model is then evaluated on the test data "3rd_Bush-Kerry.txt".

Users can use the API endpoint /predict/sentence with the sentence to get the prediction result back.

pip install -r requirements.txt

Run the flask app by:

python app.py

Example usage Call API with

curl -X POST -d  '{"sentence":"It is told to be true."}'  http://localhost:5000/predict/sentence

Its output: {"sentence":"It is told to be true.","type":"passive"}

pytest

Use this command to see the evaluation of the model on test data. If the model does not exist, it creates the model.

python ml_backend.py data/2nd_Gore-Bush.txt data/3rd_Bush-Kerry.txt

Sample output where model.pkl exist.

precision:0.945054945054945, recall:0.86, accuracy:0.9910672308415609, f1:0.9005235602094239

Sample output when model.pkl does not exist.

[Parallel(n_jobs=1)]: Using backend SequentialBackend with 1 concurrent workers.
[Parallel(n_jobs=1)]: Done 1575 out of 1575 | elapsed:  2.2min finished
Best score: 0.862
Best parameters set:
	clf__alpha: 0.4
	vect__max_df: 0.15
	vect__ngram_range: (2, 2)
precision:0.945054945054945, recall:0.86, accuracy:0.9910672308415609, f1:0.9005235602094239