Recently received a task is to do a service robot, in order to let the robot a better understanding of visitors, we use BiLSTM algorithm to train answer model,
But there is a problem, is the enterprise will add corpora, and required to publish online as soon as possible, time best can control within half an hour,
But the training model alone needs at least 2 ~ 3 hours,
I'm just involved in this field, so the plan in the industry know it is not much, don't know what a great god can raise a point,

My idea is
Solution a: adopt the way of incremental training, reduce the time of the model training, the plan to test, because the model is that at the moment, I all amount of training,
Scheme 2: using another matching degree is low, but can be immediately implemented text matching scheme, set up the front model, namely once has the change, change the first to join her in front of the corpus, guest request to switch to the front model, the model also began training at the same time, the training test is completed, switch back,

Don't know if you have a better way?