Human dialog is a highly collaborative and interactive process, which involves, among other things, the ability to engage in meta-language--that is, the ability to talk about the dialog itself or its linguistic constituents--and to use the results of meta-linguistic interactions to help coordinate the ongoing conversation and understand otherwise problematic utterances. It is our contention that facility with meta-language is necessary for the ability to engage free and flexible conversation (which we call conversational adequacy), and more importantly, that a robust meta-linguistic ability can make up for weaknesses in other aspects of language use.
However, the frequency with which, and the precise conditions under which people use meta-language is not well known. We are investigating this general question by developing a markup scheme for meta-language, and applying it to the British National, Map-Task, TRAINS-91, and TRAINS-93 corpora. Although there exist annotation schemes for dialog clarifications, as well as schemes for annotating self-correction in spoken dialog, there are currently no schemes which focus on meta-language in particular. Given the importance of meta-language to human conversation, we believe that it is time to address this lacunae.
Having such an annotation scheme, and results from these corpora, will be extremely useful to language researchers. One study we intend to perform as part of the current proposal is to cross-index our annotations with the existing annotations for these corpora. We will be looking in particular for correlations between meta-language (and its types) and other linguistic structures, e.g. local syntax, speech-act classifications, and/or dialog moves. Such correlations will not only be interesting in their own right, but will be helpful in the development of automated methods for detecting and interpreting meta-language.
This work is part of a larger project involving the development of viable natural language computer interfaces with the ability to engage in meta-dialog, and thereby with some of the flexibility that meta-language provides to human conversation. We believe that the ability to engage in even simple meta-language can be used to fruitfully enhance the performance of interactive systems, even those having relatively limited speech recognition and language processing abilities. The systems we have developed are described in: Traum, et al. 2002; Anderson, Josyula and Perlis, 2003; Josyula, Anderson and Perlis 2003; Josuyla, Anderson and Perlis, 2004.