Text Classification of Yelp’s Category Based on Text Review

---

Study Overview

Machine Learning and Deep Learning has become two of the most popular Artificial Intelligence topics across the globe in recent times, especially for being proved perform well targeting Natural Language Processing (NLP) Tasks. One foundational application of task is text mining. For the purpose of acquiring the solid ML and DL techniques, we choose to conduct the text classification experiment using Yelp Dataset. Not only we can apply our text mining theories to the actual business scenario, but to understand the model’s principles in-depth to enhance our skills. In our experiments, we have both trained many classic textual models using W2V(SG/CBOW) and GloVe and feed the document representation into the ML and DL models like LR, SVM, RF, CNN, and RNN. We have also tried implementing handcraft features and using validation accuracy to compare the experimental results.

Motivation

This study mainly focuses on Categorization in text mining topic, also as known text classification, which is one of the popular NLP tasks that is employed by most commercial entities. As the NLP tasks are applied in multiple individuals’ daily scenarios (i.e. user portrait classification, recommendation system, intelligent voice virtual assistant, etc.) This study is first set up in an Independent Study course, hence the goals not only seeking the solutions to the text classification task but also to practice the current text mining techniques combing with ML and NN techniques to resolve the problem. The study process includes collecting dataset, preprocessing dataset, researching, and practicing current ML and NN models, tunning the models, retrieving the results, and producing summaries. Meanwhile, the motivation also covers toolkit manipulation, coding, and engineering implementation. Hence, it is both endued the expectation in researching and engineering.

Repo Architecture

Directory	File	About
dataset	README.md	Introduing the dataset used in the study
	IS_new.csv	Lemmatized dataset
	IS.csv	Stemmed dataset
experiment	README.md	Giving the part of exact experiment details (more in README in ML_models and NN_models)
	run_ML.py, run_RNN.py
images		Containing all README images
ML_models	README.md	Giving the Machine Learning models details including the theory, parameters tuning, implementation.
	tuning_ML_models.py
NN_models (important)	README.md	1. Containing the detail information of the Neural Network models including Kim’s CNN model for text classification, RNN encoder model using GRU/LSTM memory unit implemented by TensorFlow Keras; 2. Illustrating models’ architectures, parameter’s tuning process and implementation summary
	create_CNN_model.py, create_RNN_model.py,tuning_CNN_model.py
plot		Containing the methods to draw figure
preprocessing	README.md	The view of the preprocessed dataset
	stem_lemmatize.py
textual_models (important)	README.md	1. Containing the report on the training/testing/tuning processes for W2V(SG/CBOW)/GloVe models using Gensim library; 2. Including the summary of the embedding models comparison and implementation; 3. Three self-trained models/Word Vectors and One pretrained Stanford GloVe 42b 300d word vectors
	cbow_model.bin	CBOW model trained and saved by Gensim Word2Vec library
	sg_model.bin	SG model trained and saved by Gensim Word2Vec library
	glove_wv.bin	GloVe word embeddings trained by Stanford GloVe program
	Stanford_glove.bin	Pretrained Stanford 42B300d Word Embeddings
	init_w2v_glove.py, tuning_glove.py,tuning_w2v.py,_lda.py,_tf-idf.py
regular_features	README.md	The method to extract the handcraft features
	hours.py
tools	old_report.md	A sumary on data reading and converting
	files_read.py, statistics.py