Language Identifier among English, French and Italian with 1-layer neural network
Part 1:
I take all the 5 adjacent letters as my input to train the NN. When predicting a sentense, I first break it down to the 5-letters and then predict the outcome for all the 5-letters. The prediction for the sentense is the simple majority. Accuracy of test set after 3 epochs: 0.67
Part 2:
I tried to increase d to see whether it improves, and I tried a high and a low learning rate as well:
d = 200, eta = 0.1, epoch = 4. dev accuracy: 0.991 d = 100, eta = 0.01, epoch = 4. dev accuracy: 0.988 d = 100, eta = 0.5, epoch = 4. dev accuracy: 0.983 d = 200, eta = 0.01, epoch = 4. dev accuracy: 0.978 d = 200, eta = 0.5, epoch = 4. dev accuracy: 0.988
The 1st and 4th model produce the highest dev accuracy. Using the 4th model to predict the test set, the final accuracy rate for the test set is 0.68