Backward Looking Functions (shown in figure 4) indicate how the current utterance relates to the previous discourse. For example, an utterance might answer, accept, reject, or try to correct some previous utterance or utterances. To capture these Backward Functions, we need to both identify the type of Backward Function and indicate the previous stretch of discourse that is being affected.
The previous utterance unit or set of units being responded to by the current utterance will be called the antecedent. The relations we consider here are local, and their antecedent typically is either the previous utterance unit or at least, in the previous turn. For the moment, we will assume that antecedents are single utterances near the current utterance, and in our examples, these will be marked in parentheses after the Backward Function tag. Note, if any Backward Looking Function is given to an utterance it must also be given a Response-to tag indicating its antecedent utterances. See section 5 for details on how to specify antecedent utterances in Response-to tags using the dat annotation tool. Constraints on what can be an antecedent and how to annotate complicated cases will be discussed after the basic set of tags is introduced.
There are four semi-independent aspects coded in this dimension. The first concerns relations affecting agreement about the task or whatever the topic of discussion is. The second concerns relations that signal understanding (or non-understanding). The third marks utterances that answer previous information requests. The fourth concerns the relationships between the informational content in utterances and is not developed further in this version of the annotation scheme. It provides a place holder for individual projects to define their own set of relations for their specific domains. The possibility of developing an abstract domain-independent level will be considered for later versions of this manual.
Figure 4: Backward Looking Function