| A simple recurrent network (SRN) identified the systematic patterns in sequences of speech act categories in two-party conversations. A speech act categorization scheme included 8 categories: question, assertion, reply to question, response, directive, indirect directive, evaluation, and non-verbal response. The systematicity was examined across a range of conversation corpora, including child-dyad exchanges, telephone directory assistance exchanges, and tutoring sessions. The induced patterns were consistent with a dynamic replanning position in which the patterns were local and reactive, rather than a conversational scripts position which would predict longer patterns. Two or three speech acts of prior context provided all the context that was needed to make the predictions for the next speech act category. This local span of systematicity was also demonstrated by a forecast analysis in which the networks' predictions were accurate within the first 1-4 speech acts into the future. The performance scores for the SRN model decreased substantially when applied to conversational corpora outside of the training corpus. The pattern of results across the three separate corpora favored a register specific position rather than a context-free position with respect to the generalizability of the dialogue systematicity. However, enough transfer did occur among the different conversational corpora that a strict situated cognition position was ruled out. Finally, an analysis of the specific speech act categories revealed that the systematicity of the SRN was mainly captured by the high-frequency speech act categories. |
