By Alexandre Papadopoulos, Pierre Roy, and François Pachet.
In the Proceedings of the 28th AAAI Conference on Artificial Intelligence (AAAI 2014), pp. 2731–2737, Québec City, Canada, 2014.
Read the paper (PDF)
Publisher record
About the paper
A Markov generator can reproduce long passages from the material it was trained on. This work treats the length of those copied passages as a constraint on generation.
Abstract summary
The authors introduce MaxOrder, a constraint represented by an automaton that limits the length of contiguous passages shared by a generated sequence and its training corpus. They illustrate the method with text generation.
Related work
See Markov Constraints for the broader family of sequence constraints and Flow Machines for their musical applications.