Which network structure to use for multivariate time series data with unknown input and output size at each time step?

I have data that has is multivariate and can be thought of as many sentences of variable length. I’m trying to predict what the next sentence will be and it depends on multiple preceding sentences. I was thinking of using something like an encoder decoder neural network but I am not sure how to structure the network to have unknown input/output size at each time step
An example encoded input for consecutive time steps could be:

[ 0 , 1 , 3 , 5 ,6] t=0
[ 0 , 5 , 4]        t=1
[4 , 5 , 5 , 6]     t=2
[0]                 t=3

With example output:

[1 , 4 , 6 ]        t=0
[2 , 3 ]            t=1
[4 , 5, 6 , 7]      t=2
[2 , 4 , 6 ,7 , 8]  t=3