Those of us who have a relatively popular YouTube channel have to spend a lot of time reading and responding to comments. If we, approaching 100,000 subscribers, have hundreds of comments per week, imagine those with millions of followers …
Now Google wants to give us a hand with the issue by expanding its smart response system, the one we use in Gmail, to YouTube comments, as shown in this animation:
As you can see, we will have, in the English and Spanish channels, a series of buttons with possible answers that we can give to the followers, so that the corresponding field will be filled in automatically.
Of course, as with emails, it will only be useful for responses such as: Thank you, more tomorrow, and things like that, so if there are followers asking questions about the content of the video, the answer will still have to be personalized.
Although it may seem basic, the system used is quite intelligent. They explain it in the Creators blog, where they indicate that in the email they use neural networks to process the text and extract words and short phrases. This preprocessing included, but was not limited to, language identification, tokenization, and normalization. Two neural networks simultaneously and independently encoded the message and the suggestion, thus enhancing the suggestions.
In the case of YouTube, they have had to expand the system to take into account emojis, ASCII art, change of language, etc. In light of this, and taking inspiration from their recent work on byte modeling and character language, they have encoded the text without any prior processing. This approach is supported by research showing that a deep transformative network is capable of modeling words and phrases from scratch simply by feeding text as a sequence of characters or bytes, with a quality comparable to word-based models.
So, for the next few hours, you will watch YouTube SmartReply, starting with comments in English and Spanish, the first SmartReply based on bytes and crossed characters.