Sentence Classification

The goal of this project is to classify sentences, based on type:

• Statement (Declarative Sentence)
• Question (Interrogative Sentence)
• Exclamation (Exclamatory Sentence)
• Command (Imperative Sentence)

Each of the above broad sentence categories can be expanded and can be made more indepth. The way these networks and scripts are designed it should be possible expand to classify other sentence types, provided the data is provided.

This was developed for applications at Metacortex and is accompanied by a guide on building practical/applied neural networks on austingwalters.com.

Please, feel free to add PRs to update, improve, and use freely!

To Install

• Install CUDA and CuDNN if you have a GPU (on your system of choice)
• Install requirements (on python 3, python 2.x will not work)

pip3 install -r requirements.txt --user

To execute:

Pretrained model:

python3 sentence_cnn_save.py models/cnn

To build your own model:

python3 sentence_cnn_save.py models/<model name>

The models/ will load any pretrained model with said name, or retrain it.

See supplemental material for full guide.

Supplemental Material

This repository was created in conjunction with a guide titled Neural Networks to Production, From an Engineer.

Below is the guides table of contents:

• Acquiring & formatting data for deep learning applications
• Word embedding and data splitting
• Bag-of-words to classify sentence types (Dictionary)
• Classify sentences via a multilayer perceptron (MLP)
• Classify sentences via a recurrent neural network (LSTM)
• Convolutional neural networks to classify sentences (CNN)
• FastText for sentence classification (FastText)
• Hyperparameter tuning for sentence classification

Additional, (more complex models) are available in the advanced_modeling directory. Eventually, posts should come out of them.

Dataset

The dataset is created from parsing out the SQuAD dataset and combining it with the SPAADIA dataset.

The samples in the dataset:

• Command 1111
• Statement 80167
• Question 131001

Note: Questions in this case are only one sentence, statements are a single sentence or more. They are classified correctly, but don’t include sentences prior to questions.

Results

With the above, we are able to get the following accuracy:

With some hyperparameter tuning:

Computer Configuration:

• GTX 1080
• 32 Gb RAM
• 8x 3.6 Ghz cores (AMD)
• Arch Linux, up to date on 12/16/2018

CNN Hyperparameter tuning