How is emotional impact of words highlighted? about deepmoji

[bfelbo]

It's a bit hacky so I wouldn't recommend using it in production. We do the predictions w/o each word and see how much the distribution over predicted emojis change.

[amitness]

@bfelbo I see. So if I understand it correctly, is this how it's done?

• Say we have a sentence "I loved the application" and we run deepmoji to get top 5 predictions
• Then, we remove one word at a time and see the predictions
• The words whose removal changed the top 5 predictions are the important ones

If so, how do you give that score in the screenshot above? The importance reflected by the darker and lighter shades of purple.

[bfelbo]

Exactly. The shading is simply an interval over the difference. It was picked manually by trying out a couple of sentences and seeing what seemed to make sense.

Note that you can run all the predictions in a single batch, which will make this pretty fast even for long sentences.

[bfelbo]

I'm closing this for now as it's not related to the functionality of the repo, but to the demo. Hope my response was helpful!

[amitness]

It was. Thanks a lot.