-
We take the ApteMod subset of the Reuters-21578 corpus
-
For feature engineering, TF-IDF and Word2Vec are implemented
-
For classificiation model, Linear SVC, Feedforward Neural Network and Bi-LSTM are applied
-
The classification accuracy, Macro F1 and Weighted F1 are shown below:
| Feature Extraction | Model | Accuracy | Macro F1 | Weighted F1 |
|---|---|---|---|---|
| TF-IDF | SVC | 0.81 | 0.47 | 0.84 |
| Feedforward | 0.78 | 0.46 | 0.83 | |
| Word2Vec (Pre-trained) | SVC | 0.73 | 0.29 | 0.75 |
| Feedforward | 0.76 | 0.43 | 0.80 | |
| Bi-LSTM | 0.80 | 0.39 | 0.82 | |
| Word2Vec (Re-trained) | SVC | 0.79 | 0.40 | 0.81 |
| Feedforward | 0.80 | 0.45 | 0.83 | |
| Bi-LSTM | 0.81 | 0.50 | 0.84 |
- The Confusion matrix of Bi-LSTM based on Re-trained Word2Vec is as follows:
