Dialogue Modelling and Generation
Peter Kuhnlein, Paul Piwek · Discourse Processes · 2007
Traditionally, the tasks of understanding input and of generating output are markedly distinguished in natural language processing (nlp). This distinction methodologically becomes very clear in dialogue processing. Computational linguists who had to deal with understanding of humanly produced discourse quite naturally were urged to find general models of their subject matter. In a way, for the task of understanding, finding models of discourse is an extension of building grammars for sentence parsing. Needless to say, there are differences between the syntax at sentential level and the structurs found at discourse level, but the task of finding the right structural descriptions is ubiquitous. One early theory of discourse structure and interpretation that explicitly was based on observation of human behaviour was formulated in (Polanyi, 1986). This paper further developed earlier work (Polanyi & Scha, 1983a, 1983b) and remains influential. The most famous discovery with regard to human dialog that is reflected in this theory pertains to possible continuations in discourse. It has been observed that new contributions to discourse can only relate to certain earlier contributions. These are said to reside on the right frontier of the discourse parse tree. More formally, the node representing the last utterance and every node dominating this one are said to be on the right frontier. This property of human discourse was implemented in Polanyi’s LDM and is part and parcel of every theory of discourse structure since. The situation regarding the need to mirror naturally occuring phenomena is different for generation. Unless the additional requirement of generating humanlike (also called natural) responses is adopted, generation need not be based on models of human behavior. The pressure to adhere to human behaviour as the guideline is lower for generation than for understanding. This has led to the situation that only recently researchers strive to increase the naturalness of machine output by looking at models of human behaviour. Especially the quest for embodied conversational agents (ecas) sparked this development. This is witnessed in (Cassell, 2000, 70), where the author states that