/Text-Sentiment-Classification

PCML Project 2

Primary LanguagePython

Text Sentiment Classification

Authors: Yamane El-Zein, Camila Andrea González Williamson and Luis Emmanuel Medina Ríos.

Report

Please refer to the file Text_Sentiment_Classification.pdf

Disclaimer

The corresponding code as well as the report in this repository was developed by Yamane El-Zein, Camila Andrea González Williamson and Luis Emmanuel Medina Ríos from November 2016 to December 2016 for the course Pattern Classification and Machine Learning at EPFL, Switzerland.

Abstract

We discuss our findings for the Pattern Classification and Machine Learning project on text sentiment classification. Our work involves three main steps: a brief analysis of the text data, preprocessing of the text data, and testing different variations of text classification methods. These methods include the bag of words method, two word embeddings techniques, namely Word2Vec and GloVe, and finally the FastText method. We describe in detail all tested methods and compare their results. In the end, we find that an ensemble of 200 FastText classifiers produces the best results.

How it works

Note: the code was tested on MacOS

Required installations:

  • Numpy: to install Numpy execute: pip install numpy
  • Fasttext: to install Fasttext execute: pip install fasttext

Folder structure should be as follows:

  • data subdirectory: contains the following file:

    • data.zip: contains the three text files used to train and test the model, please UNCOMPRESS the file in the same folder (data) before continuing:
      • train_neg_full.txt: full train data set for negative tweets
      • train_pos_full.txt: full train data set for positive tweets
      • test_data.txt: test data set for predicting

  • Root directory:

    • run.sh: it is the bash script that has to be run to generate the submission file.
    • fasttext200.py: it is the python file called by run.sh that runs N=200 times the fasttext_test_probs.py file and average all the probabilities of the predictions that were generated.
    • fasttext_test_probs.py: it is the python file that runs FastText once, called by fasttext200.py
    • helpers.py: a python file that contains the functions to create the csv files.

The following directories are automatically created by running the run.sh script:

  • proc: contains the labeled files that were processed in order to use FastText.
  • mixed_pos_neg_labeled_FT.txt: final .txt file used to create the FastText model. it contains both negative and positive labeled tweets.
  • mixed_pos_neg_test.txt: test data set created from splitting the train data set in order to evaluate the model.
  • mixed_pos_neg_train.txt: train data set created from splitting the train data set in order to train the model.
  • negative_labeled_FT.txt: a .txt file that contains the labeled negative tweets.
  • positive_labeled_FT.txt: a .txt file that contains the labeled positive tweets.
  • test_data.txt: a .txt with the test data without the tweet ids.
  • results: contains both the model (model.bin) created by FastText and the csv file with the probabilities of one run of FastText:
    • FastText_output_probs.csv: Contains the probabilities of the predicted tweets.
    • model.bin: it is the model created by FastText.
    • final: contains the submission file:
      • FastText_ensemble_submission.csv: the final submission file.

In order to get the submission file, one has to run the bash script (run.sh) and there should be a subdirectory in the root directory called “data” containing the files “train_neg_full.txt”, “train_pos_full.txt” and “test_data.txt”.