Testing dialogue principles in task-oriented dialogues: An exploration of cooperation, collaboration, effort and risk

This paper takes four behavioural principles which have been suggested as explanatory models for human conversation and tests them on a corpus of task-oriented dialogues (the HCRC Map Task Corpus). The principles chosen are Grice’s Cooperative Principle, a folklinguistic notion of ‘cooperation’ (which we argue is often confused with the Gricean notion), Clark’s Collaborative Theory, and Shadbolt’s Principle of Parsimony. The aim of the study is to compare the explanatory power of each of these principles when they are applied to real language data.

Each of the principles was converted into a set of representative hypotheses about the types of behaviour which they would predict in dialogue. Then, a way of coding dialogue behaviour was developed, in order that the hypotheses could be tested on a suitably sized dataset. In particular, the coding system tried to distinguish between the levels of effort which participants used in their utterances. Finally, a series of statistical tests was undertaken to test the predictions of the hypotheses on the information generated by the coding system.

The strongest support was found for the Principle of Parsimony and its associate Principle of Least Individual Effort, at the expense of the Collaborative Principle and the Principle of Least Collaborative Effort. There is certainly evidence that speakers try to minimise effort, but this seems to be occurring on an individual basis – which can be to the cost of the overall dialogue and task performance – rather than on a collaborative basis. Some support was also found for Gricean Cooperation, although this is weakened by the difficulty in transforming the underspecified nature of Grice’s work into a precise and unarguable set of predictions. However, a clear distinction can be drawn between Gricean Cooperation and the folklinguistic notion: even a broad definition of Grice is manifestly different from the predictions made for ‘cooperation’, and these indicators of ‘cooperation’ were not supported by the data.