Pros and Cons of localization based on a key or on the actual string

I am interested to find good pros and cons on implementing localization based on keys or the actual strings.

With keys, the implementation is a function where the parameter is a key for that string:

translate(DeleteFileConfirmationId)

With actual strings, the implementation is a function where the parameter is the actual string in the default\neutral language:

translate('Are you sure you want to delete the file?')

The issue I see with the latter is if the application needs to display a different element (for example, an image) based on the value returned from the backend, if the value is a string and not a key, it has to check with the string:

if(x === 'Red dragon') 
{
    displayImage('RedDragon.png');
}

Any other pros and cons?