Comprehend incremental training

We trained custom model with manually annotated data set of documents but the accuracy is low, we want to annotate and train again. When I create a new version will it learn from the current set of inputs and also preserve the old training ? Do I need to give all the data for every incremental training ?

Since there is some low accuracy issue, I want to add a2i . How to do it in console UI for batch processing ? When the people make additional annotation in a2i, will the comprehend learn incrementally ? Or do we need to make run the training job again ?