This post follows the google AI blog.
https://ai.googleblog.com/
introduction
Last week at Google I/O, they introduced Smart Compose, a new feature in Gmail that uses machine learning to interactively offer sentence completion suggestions as you type, allowing you to draft emails faster.
This is build upon Smart Reply using TPU.
Key Challenges
In developing Smart Compose, there were number of challenges, including
1. Latency: To interactively suggest the sentence while user types, the latency play important role. Indeed, since it predicts the sentence on a key-stroke basis, ideally it should be able to respond within 100ms.
2. Scale: Currently Gmail is used almost 1.4 billion diverse users. So the language they use is diversified as well. Hence they had to prepare the multi-language model.
3. Fairness and Privacy: During investigating the dataset, they have realised the difficulty in handling personal information since any emails containing certain information which must not be exposed anywhere. So they seriously made sure that they don't leak or expose the user's personal information.
Finding the Right Model
They have combined the BoW(Bag-of-Word) and RNN-LM(RNN-Language Model) which is similar to FastText from FaceBook.
In fact, the novelty arises in the data pre-processing. They have embedded previous message and subject line averaging them respectively. So the RNN-LM take as input this word embedding and predict next word step-by-step.
Future Work
They mean to constantly improve the quality of this auto-suggestion of sentence completion. Especially they will work with Transformer RNMT+ to improve the performance.