I'm looking at a fill-in-the-blanks type of problem and I have not found a lot on it, which is why I am asking. It may be that it exists (and it probably does), but it goes by a name I'm not familiar with.
What I want to do is predict the best (i.e. most likely) sequence of words (events/things/whatever) based on a learned model of these words. However, in most (Hidden) Markov Models you predict the next characters/words based on the past. What I want to do is fill in the blanks because I also have text (or events) after the blank and that ideally affect the most probable sequence in my opinion. What is unknown in my problem is the number of words that have to go in the blank. Ideally I'd like to compute the most likely sequence.
After all, if presented with "This is ..." the best may be "This is great!", but if presented with "This is ... ever!" the blank may best be filled by 'dumbest idea', depending of course on the model.
I was thinking about several ways this might be done:
- Predict the next words/events in the sequence and pick the one that generates the following words, so: predict next and line up. Constraining the original search space is likely the best alternative here because it's not known how many words/events would best fit in the blank spot.
- Compute the next in sequence based on the normal model and the previous based on a sequence-reversed model, then meet in the middle.
My question is: Is there any research on this topic? I have looked for (Hidden) Markov Models but the classic use case is to predict the next in sequence or the most probable sequence overall. I haven't found anything on a fill-in-the-blanks problems, but I may have overlooked it as I am not an expert. RNNs might be an option too, but also there I haven't read anything that seems to fit the bill.
Apologies if this is not the right place