Incremental PDFA learning for conversational agents

When finite-state machines are used for dialogue models of a conversational agent, learning algorithms which learn probabilistic finite-state automata with the state merging method are useful. However these algorithms should learn the whole data every time the number of example dialogues increases. Therefore, the learning cost is large when we construct dialogue models gradually. We proposed a learning method which decreases the number of compatibility checks by caching the merging information, and evaluated it and the perplexities of learned models. From the comparison among the dialogue models, the method which caches only the compatibility-changed states reduced the total number of compatibility checks by 13%. We also applied the algorithm to an actual conversational agent.