Efficient Strategy and Language Modeling in Human-Machine Dialogues

摘要

Each time an Interactive Dialogue System (IDS) is adapted to a new domain, the language modeling and dialogue strategy modules must be modified to fulfil the new requirements. In this paper we present an algorithm for creating Stochastic Finite-State Netwoks (SFSN) for language modeling of dialogue states in an IDS. The resulting SFSNs are evaluated in terms of perplexity and recognition performance. Moreover, we present a method that enables the designer of the dialogue strategy to investigate system performance by employing diagnostic evaluation during the initial phases of a system's development. The recognition success rate taken from the previous language model evaluation combined with the proposed dialogue mathematical modeling, can be used to predict an IDS's behaviour by relating dialogue parameters (e.g. recognition success rate, number of turns, dialogue strategy) with the final system's performance. Thus the effort during global system assessment is reduced since we have diagnostic measures in advance.