Text Classification

The goal of this repository is to implement text classification in traditional machine learning methods and deep learning methods (in Pytorch).

Note: Since I have to spend more time in conducting research, I stop reproducing classic papers on the task of text classification. However, I will still maintain this project. Any contributions lke issues and pull requests are welcome.

Models

Feature Extraction

BOW (Bag of Words)
TFIDF (Term Frequency-Inverse Document Frequency)
N-gram

Traditional Machine Learning Methods

Deep Learning Methods

DAN: Deep Unordered Composition Rivals Syntactic Methods for Text Classification
TextCNN: Convolutional Neural Networks for Sentence Classification
TextRNN: Recurrent Neural Network for Text Classification with Multi-Task Learning
~~RCNN~~
~~Capsule Network~~
~~Transformer~~
~~Elmo~~
~~BERT~~

Experiments

Effects of Feature Extraction Methods

Methods	KNN	Decision Tree	Perceptron	Bagging	Random Forest	AdaBoost	Gradient Boosting	Naive Bayes	SVM	Logistic Regression
BoW (ngram-range=(1, 1))	0.665	0.707	0.742	0.749	0.757	0.724	0.713	0.800	0.823	0.826
BoW (ngram-range=(1, 2))	0.666	0.699	0.750	0.744	0.751	0.724	0.712	0.795	0.819	0.823
BoW (ngram-range=(1, 3))	0.667	0.700	0.712	0.748	0.757	0.724	0.712	0.795	0.818	0.824
BoW (ngram-range=(2, 2))	0.579	0.628	0.652	0.646	0.652	0.584	0.600	0.669	0.671	0.692
BoW (ngram-range=(2, 3))	0.578	0.625	0.625	0.636	0.648	0.584	0.600	0.662	0.667	0.684
BoW (ngram-range=(3, 3))	0.536	0.572	0.525	0.578	0.581	0.532	0.539	0.561	0.576	0.590
TFIDF	0.714	0.705	0.594	0.759	0.760	0.723	0.714	0.807	0.804	0.824

Effects of Word2Vec Methods

Methods	DAN	TextCNN	TextRNN
Trained Using Corpus	0.880	0.893	0.903
Glove	0.886	0.900	0.890

To-do List

Word2Vec Methods (Pre-trained and Trained Using Corpus) Comparision
Class Vectorizer lazy initialization
Cross Validation
Grid Search
Visualize text feature
Shared Classifier Module for Deep Learning

tienanh28122000/text_classification