Evaluation metrics for Sequence-based tasks

In the last article, we discussed the Learning Algorithm for recurrent neural networks. In this article, details about the model evaluation are discussed:

Evaluation

For sequence classification tasks (binary classification whether a sentence conveys positive or negative sentiment), model evaluation could simply be the accuracy matrix/accuracy value wherein for each training example the actual outcome is compared with the predicted one and evaluated how many training instances were identified correctly

And is computed as the (No. of correct predictions divided by the total No. of training examples) or essentially the fraction of instances predicted correctly out of the complete dataset

For multi-class classification (say sequence labeling tasks wherein the model predicts an output for each time step), overall accuracy, as well as the accuracy per class, could be computed

The overall accuracy could be computed as the class is predicted for each word in every sentence and from this, the number of correct predictions is counted and is divided by the total number of words (across all the instances in…