Text Classification with ConvNets (Basic Keras Practices)

CNN classifier for Formality in text.

Cross Validation

In order to run cross validation on a dataset, you can just run:

make cross-validate-formality.lahiri.dataset NUM_FOLD=10 PROBLEM_TYPE=regression

PROBLEM_TYPE defines your problem, either regression or classification. Labels for formality.lahiri.dataset between -3 to 3 and real-valued. So, you should give PROBLEM_TYPE=regression. NUM_FOLD determines how many fold you run for evaluation of the dataset. Important to note that dataset should be in datasets/ directory.

Example for a classification problem:

make cross-validate-formality.lahiri.classes.clean.dataset NUM_FOLD=10 PROBLEM_TYPE=classification

Build and Save Model

Assume that we have formality datasets for emails and it is located in datasets/formality-email. You can run the following command; build and save a model in pretrained-models directory.

make pretrained-models/formality-email USE_PRETRAINED_EMBEDDINGS=False

If you want to use pretrained word embeddings (e.g., Word2Vec) then you can run this:

make pretrained-models/formality-email

Prediction with Pretrained Models

Assume that you trained a regressor (or classifier) by using formality email datasets as above. By using this pretrained model, we can predict formality scores of the sentences in news domain. In the datasets directory, there is a dataset named formality-news. Let's use it as our test data.

make formality-news.pred MODEL_DIR=pretrained-models/formality-email USE_PRETRAINED_EMBEDDINGS=False

If you want to use Word2Vec and if you haven't built any model yet, you may remove providing USE_PRETRAINED_EMBEDDINGS=False (since default is True):

make formality-news.pred MODEL_DIR=pretrained-models/formality-email

Even you haven't run the command for build a model, Makefile finds the dependency path and run the necessary commands for you. Also, it is important that training data for the model building should be somewhat similar with the test data. News and email domains can be significantly different with each other. I used these two datasets just to illustrate the commands.