dennybritz/cnn-text-classification-tf

This repository is a Python implementation of a specific machine learning technique for sorting text into categories. The technique, called a convolutional neural network, was described in a 2014 academic paper by Yoon Kim, and this code is a simplified version of that paper's approach, built using TensorFlow. It was created as companion code for a blog post on the site WildML. The idea behind text classification is to train a program to read a piece of text and decide what category it belongs to. For example, you could teach it to label movie reviews as positive or negative, or sort support tickets by topic. The model learns by reading many examples, finding patterns in the words, and adjusting itself until it gets good at making those judgments. To use this code, you run a training script that feeds it data and produces a saved model. There are several settings you can adjust before training, like how many passes to make over the data, how large to make the batches of examples processed at once, and how aggressively to apply regularization (a technique that helps the model avoid memorizing the training data too closely). Once training finishes, a separate evaluation script lets you test how well the resulting model performs. The repository is fairly minimal in scope. It does not include a graphical interface or a hosted demo. Using it requires some comfort with running Python scripts from a terminal, managing dependencies like TensorFlow and Numpy, and preparing your own text dataset if you want to classify something beyond the default example data. This code dates from 2015 and targets an older version of TensorFlow. Researchers and students studying the foundations of text classification models still find it useful as a clean, readable reference for how this type of neural network is structured.