Generating and understanding dialogues without planning

While symbolic planning is useful in generating dialogues, it is not as useful in understanding them. Dialogue systems that are programmed with a set of fixed rules cannot parse plans that are outside the limited set of plan structures specified by those rules. Input can fall outside the rules because the user is misconceived or because the rules only cover a fraction of the domain, and so the user may attempt correct, but out-of-scope plans. It would be useful for dialogue systems to learn the sequences of actions that occur in a dialogue from given dialogue corpus data, without any use of prespecified rules. This approach would have a second advantage, in that it performs user modelling as well, being able to find out the patterns of actions that a particular user or stereotype group prefers.

An analogous problem occurs in recognition of not action sequences, but word sequences in sentences. Statistical language models are much better than hand-crafted grammars in predicting word sequences, illustrated by their usefulness in speech recognition systems and in machine translation. This has led to a branch of research in dialogue understanding that uses statistical information about dialogue act sequences to better understand the user's utterance. Each dialogue act has a type and a propositional content. Deciding the type of the act is a classification task to which common machine learning algorithms can be applied, using various features of the previous dialogue acts, and the current utterance [63]. The dialogue act type can then fill one slot in a semantic frame for the utterance. Dialogue act classification has also been used in speech recognition, where a small gain in recognition accuracy can be obtained by applying language models that have been trained on the expected act type of the given utterance [70].