Interpreting loss/accuracy

I’ve been trying to train a bidirectional LSTM on some hindi language text for sentiment analysis, not using ULMFIT, but with simple fasttext/glove word vectors. Not sure how to interpret the results I’m getting right now. My loss drops, accuracy oscillates, validation loss drops but validation accuracy just doesn’t move.

I’m guessing this is underfitting, and given another 10 epochs, it begins to overfit. Should I be tinkering with my learning rate or doing something else?

Currently this is using Adam, with no batch normalization, this is my model summary.


Layer (type) Output Shape Param #

embedding_5 (Embedding) (None, 90, 50) 93833250


bidirectional_2 (Bidirection (None, 40) 11360


dropout_7 (Dropout) (None, 40) 0


dense_6 (Dense) (None, 2) 82


dropout_8 (Dropout) (None, 2) 0


activation_3 (Activation) (None, 2) 0


dense_7 (Dense) (None, 1) 3


dropout_9 (Dropout) (None, 1) 0


activation_4 (Activation) (None, 1) 0


dense_8 (Dense) (None, 3) 6

Total params: 93,844,701
Trainable params: 93,844,701
Non-trainable params: 0


Train on 2638 samples, validate on 660 samples Epoch 1/10 2638/2638 [==============================] - 55s 21ms/step - loss: 1.0266 - acc: 0.4325 - val_loss: 0.9340 - val_acc: 0.5091 Epoch 2/10 2638/2638 [==============================] - 44s 17ms/step - loss: 1.0249 - acc: 0.4325 - val_loss: 0.9537 - val_acc: 0.4000 Epoch 3/10 2638/2638 [==============================] - 46s 17ms/step - loss: 1.0187 - acc: 0.4329 - val_loss: 0.9558 - val_acc: 0.4000 Epoch 4/10 2638/2638 [==============================] - 45s 17ms/step - loss: 1.0207 - acc: 0.4219 - val_loss: 0.9783 - val_acc: 0.4000 Epoch 5/10 2638/2638 [==============================] - 44s 17ms/step - loss: 1.0167 - acc: 0.4265 - val_loss: 0.9551 - val_acc: 0.4000 Epoch 6/10 2638/2638 [==============================] - 44s 17ms/step - loss: 1.0174 - acc: 0.4401 - val_loss: 0.9579 - val_acc: 0.5091 Epoch 7/10 2638/2638 [==============================] - 44s 17ms/step - loss: 1.0212 - acc: 0.4117 - val_loss: 0.9610 - val_acc: 0.5091 Epoch 8/10 2638/2638 [==============================] - 44s 17ms/step - loss: 1.0188 - acc: 0.4230 - val_loss: 0.9555 - val_acc: 0.4000 Epoch 9/10 2638/2638 [==============================] - 45s 17ms/step - loss: 1.0178 - acc: 0.4272 - val_loss: 0.9455 - val_acc: 0.5091 Epoch 10/10 2638/2638 [==============================] - 44s 17ms/step - loss: 1.0160 - acc: 0.4352 - val_loss: 0.9363 - val_acc: 0.5091