Mohamed209/text_classification_assessment

Text Classifier API

this repo may be considered as text classification API

it describes text classification pipeline from text preprocessing until model evaluation

folder description

src >>> contain all source code / class definitions fro text preprocessing model training and evaluation
model_weights >>> contain weights of trained models and vectorizer objects
data >>>
├── model_predictions
│   └── predictions.txt # taskA results
├── processed # processed data ready for training
│   ├── train.csv
│   ├── x_testrandom_forest.txt
│   └── y_testrandom_forest.txt
└── raw # received non processed data
    ├── test_data.txt
    └── train_set.txt
taskA.ipynb and taskB.ipynb >>> driver code for training and evaluating models 
for task A and B

hyparameters tuning

all models parameters fitted in this repo are found using grid search technique we will discuss them in detail for every task

Task A

• text preprocessing pipeline :

  1. remove punctuation , special chars , extra spaces and numbers

  2. augment text to increase data set variance

  3. text stemming

• model training

  1. TF-IDF vectorizer used to represent text where best parameters found by grid search are :

  ngram_range=(1, 2)
  max_features=3000
  keeping other parameters to default
  all stop words are removed / excluded from training
  

  2. random forest classifier used to train the model

  n_estimators=200
  

• model evaluation check accuracy , confusion matrix and classification report for trained classifiers

  also find predections from "test_data.txt" in

  data --> model_predections --> predictions.txt
  

Note

in taskA , a text augmentation technique was used as original dataset contains only about 1750 training examples , after text augmentation i got 7500 training example this trick increases the model accuracy with about 14%

Task B

I used a LSTM neural network with hyparameters tuned using also gridsearch and with architecture as

Embedding (500,32) >> LSTM (100) >> Dense (1)

embedding is an advanced technique for word representations , as it makes the model knows words meaning and similarity between them so it is clear that why neural networks outperform classical machine learning algorithms