Why does the author treat 'where value' as a classification problem?

[shaomai00]

The start and end token are treated as a classification problem and use cross-entropy loss. Have you tried nll loss and treated it as a tagging problem? Thank you for the great work. ^^