tgzw/text-formality-classifier

This is a text formality classifier project that examines how different machine learning models perform in classifying a text as formal or informal. Collaborators: Yonglin Wang, Xiaoyu Lu

Formality Classifier

Documentation

This is a text formality classifier project that examines how different machine learning models perform in classifying a text as formal or informal.

Detailed documentations of the project, such as the project proposal and the project report, can be found in doc directory.

Package Requirements

Here are the packages required to run our program

• TextBlob 0.15.3
• Scikit-learn
• readability 0.3.1
• NLTK
• spaCy
• Keras 2.3.1

To install the latest version of TextBlot, Keras, readability, and spacy, run the following commands in the exact order:

$ pip install textblob
$ pip install Keras
$ pip install readability
$ pip install spacy
$ python -m spacy download en_core_web_sm

Running the Code

Since all the feature vectors, vectorizers, and models are saved with joblib, you can skip right to the Model Training and Predicting sections.

Loading Corpus

As requested by the original owner of the corpus, the raw text version of the corpus will NOT be included in this public repository.

• Instead, we have pickled the dictionary object containing the corpus and corresponding tags. To load the corpus, run corpus.py under preprocessing folder.

Feature Extraction and encoding (very time consuming)

To perform all the feature extraction/load the vectorized features, run features.py under feature folder.

Model Training (slightly time consuming)

To see how the model training works, run the following Jupyter Notebooks in the toplevel of the repository:

• models_raw.ipynb: model training using the raw text of a sentence
  • Models Used: Naive Bayes (MultinomialNB), Logistic Regression (LogisticRegression), Decision Tree (DecisionTreeClassifier)
• models_features.ipynb: model training using different features extracted from a given sentence (not the sentence itself)
  • Features: entity types, entity length, simple numerical stats (fast number), n-gram, readability
  • Models Used: Naive Bayes (Multinomial NB), Logistic Regression (LogisticRegression), Decision Tree (DecisionTreeClassifier)
• model_GloVe.ipynb: model training using LSTM

See the write-up for an interpretation for the model evaluation.

Predicting (Beta Feature)

To see how the model makes a prediction for a few example sentences, open prediction.ipynb in the toplevel of the repository.

• Due to some compatibility issues, there's a portion of models that cannot be loaded for prediction. We are currently working on this.

Troubleshooting

1. Getting "OSError: [E050] Can't find model 'en_core_web_sm'. " when running get_subjectivity:

   • Follow the installation instruction in Package Requirements section in this file.

   • Use the following command to check where your spacy package is located.

     $ spacy info

     Then, make sure the project interpreter of your IDE is in the same directory.

   • If error persists, try other installation methods listed here