1. Introduction - Coding Purpose

Co-operativity is a central issue both in human-computer dialogue and in human-human dialogue.  Non-co-operative behaviour easily leads to miscommunication and an unnecessarily long, complicated and perhaps even ultimately unsuccessful dialogue.  In particular in human-computer interaction the consequences may often be a totally failed dialogue because of the system’s limited abilities to detect, handle and recover from non-co-operative (or seemingly non-co-operative) dialogue [Bernsen et al. 1996].

Research into aspects of communication, including various communication problems and co-operativity, is not new (see e.g. [Grice 1975]).  However, there is no exhaustive theory on these, and much of the research has focused on human-human dialogue.

With an increasing number of advanced spoken language dialogue systems available as viable means of supporting people in carrying out ordinary tasks (such as automatic flight/train timetable information, ticket booking, directory enquiries, etc.), there is also an increasing demand for rigorous methods and tools for the analysis of problems (latent and actual) in communications so as to pre-empt them if possible, or failing this, to initiate some form of repair.  Thus, the study of communication problems is of obvious importance for the development of adequate system interaction models.

If detected, communication problems in everyday conversation typically lead to clarification or repair meta-communication.  Human-human dialogue both allows for and is greatly assisted by the possibility of this kind of repair and maintenance.  However, for spoken language dialogue systems the situation is different.  The possibility of real-time handling of clarification and repair meta-communication using current technology is seriously limited.  User needs for clarification meta-communication that arise from the way the system addresses the domain, can easily surpass the systems meta-communication skills.  Thus co-operative communication is important because it facilitates smooth interaction and prevents unnecessary user-initiated clarification and repair meta-communication, as well as other kinds of unexpected user behaviour with which the system cannot cope.

In spite of this, the detection of communication problems in spoken language dialogue systems has so far usually been carried out in a rather unsystematic manner, and on ad hoc basis; it has generally only been performed at a fairly late stage, as part of the evaluation, if at all.  In order to support a more cost-efficient development process for interaction models of spoken dialogue systems we need a solid understanding of communication problems, both their nature and why they occur.  A straightforward way of achieving this is to annotate them for analysis from the transcripts of SLDS interaction.  Annotation of communication problems in spoken dialogue corpora may not only help developers and researchers to extract information on the deficiencies of a system, but may also yield clues as to how it might be improved.  It could also provide the insight required for creating methods or tools to enable the efficient, systematic development and evaluation of a system, especially during early analysis and design.  This would result not only in the improvement of interaction model quality, but also in the reduction of development costs.

Communication problems are in several respects different from most other phenomena usually annotated and studied in a corpus.  Most notably they need not necessarily be present in a corpus at all.  In fact the fewer there are of them the better.  This is in direct contrast with prosodic and morpho-syntactic phenomena and dialogue acts, which are present in any spoken dialogue corpus.  The same is also true to some extent for coreference (although unlike the aspects of language listed previously, mark-up for coreference is neither continuous nor contiguous).  All these phenomena form part of the building blocks of a dialogue.  Communication problems, on the other hand, are disruptive to a dialogue and co-operative human interlocutors usually try to avoid them.  However, in order to avoid them in human-computer interaction we need to study the nature and mechanisms of communication problems.  The computer is much less flexible than a human which makes it much more likely that communication problems will occur in a dialogue unless great care is taken to design the dialogue co-operatively within the limitations dictated by the spoken language dialogue system in question.