BanglaMCT7 : Bangla Multi-Class Text Classification

Click to download the dataset

Click to download trained models

Installation and training

Installing, Downloading Prerequisites and Traning

. install_and_train.sh

Only Traning

python __main__.py

Preprocessing

Train data

Before removing null : 95134

After removing null : 85800

Before removing duplicates : 85800

After removing duplicates : 70028

837 Documents removed which length less than equal 20

Total Remaining Documents: 69191

Test data

Before removing null : 23784

After removing null : 21523

Before removing duplicates : 21523

After removing duplicates : 17806

EDA

Maximum Length of a Document: 4442

Average Length of a Document: 223

EDA images saved to : REPORT/IMAGES/EDA

Trained models accuracy

Trained models accuracy & loss history images saved to : REPORT/IMAGES/HISTORY

Trained models performence report saved to : REPORT/PERFORMENCE

By considering overfitting and stable performance of pre-trained model I am selecting SIMPLE ANN (2 dense layers) with fasttext tokenaizer model as the best model.

Best Model

SIMPLE ANN (2 dense layers) with fasttext tokenaizer

Model file path : resources/simple_fasttext_best_model.pkl

Model summary

Model history

Note : The callback function will save the model if accuracy is improved during the epoch.

Report - train data

___________________ confusion_matrix _____________________
[[ 9708    71    19   173    67    12   234]
 [   96  7616    20    88    81     3   294]
 [   36    23 10220    79    49    23    25]
 [  130    34    84  9213    32    22   154]
 [   62    59    20    31 10862     8     9]
 [   22     8    57    69    12 14356    10]
 [  259    96    42   298    14    10  5118]]


___________________ classification report _____________________
               precision    recall  f1-score   support

      economy       0.94      0.94      0.94     10284
    education       0.96      0.93      0.95      8198
entertainment       0.98      0.98      0.98     10455
international       0.93      0.95      0.94      9669
     politics       0.98      0.98      0.98     11051
       sports       0.99      0.99      0.99     14534
   technology       0.88      0.88      0.88      5837

     accuracy                           0.96     70028
    macro avg       0.95      0.95      0.95     70028
 weighted avg       0.96      0.96      0.96     70028

Report - test data

___________________ confusion_matrix _____________________
[[2582   21    5   40   27    4   70]
 [  24 1898    6   24   25    3  144]
 [  10    9 2444   31   21   11   16]
 [  43   13   28 2307    7    6   48]
 [  16   11    8    5 2768    6    5]
 [   7    2   18   12    8 3629    2]
 [  71   50   12   80    8    2 1219]]


___________________ classification report _____________________
               precision    recall  f1-score   support

      economy       0.94      0.94      0.94      2749
    education       0.95      0.89      0.92      2124
entertainment       0.97      0.96      0.97      2542
international       0.92      0.94      0.93      2452
     politics       0.97      0.98      0.97      2819
       sports       0.99      0.99      0.99      3678
   technology       0.81      0.85      0.83      1442

     accuracy                           0.95     17806
    macro avg       0.94      0.94      0.94     17806
 weighted avg       0.95      0.95      0.95     17806

Flask Web APP

app.py