A dialog is a spoken, typed or written interaction in natural language between two or more agents. This scheme was developed primarily for two-agent task-oriented dialogs, in which the participants collaborate to solve some problem. Terms such as speaker and listener will be used in this manual, however, those working with typed dialogs or with dialogs with more than two participants should substitute ``currently contributing participant'' for ``speaker''. ``listener'' refers to the person that the speaker directs their utterance toward or to the person who responds to the speaker's utterance in the case where the speaker does not direct his speech at a single person.
A dialog is divided into units called turns, in which a single speaker has temporary control of the dialog and speaks/writes for some period of time. Within a turn, the speaker may produce several spoken or typed utterance units. While there are many possible ways to define utterances, here we based the notion of utterance on an analysis of the intentions of the speaker (the speech act). For each utterance, the annotation involves making choices along several dimensions, each one describing a different orthogonal aspect of each utterance unit.
In the unannotated dialogs as well as the examples in this manual, the start of utterance units will be preceded by an utterance name such as utt1 or utt2, and the turns will be marked to the left of the utterances (space permitting). The dialog excerpt below shows the formatting; the excerpt contains three utterance units, two spoken by u (``I need a'' and ``okay'') and one by s (``so you've got the engines at Elmira and uh'').
T1 utt1: u: I need a | utt2: okay T2 utt3: s: so you've got the engines at Elmira and uh
In spoken dialogs, when one speaker interrupts the other, speakers will sometimes talk at the same time. Words of overlapping speech are marked with numbered square brackets with the number in parenthesis next to the right bracket. You can match up overlapping sections by finding all the bracketed text with the same index. Consider the two utterance units below. Here, ``assuming'' and ``okay'' are marked as overlapping since they are both are bracketed with number 1 brackets.
T1 utt1: s: uh would take two hours <sil> [ assuming ](1) you have | | an engine at Bath T2 utt2: u: [ okay ](1)
In the middle of a sentence, a speaker will sometimes pause long enough for a second speaker to interject a few words and then the first speaker will finish their sentence. This is marked by empty square brackets as shown below.
T1 utt1: s: so you've got the engines at Elmira and uh [](1) Avon T2 utt2: u: [Avon](1)
Cases where one speaker interrupts the sentence of another will be handled as shown above. Sometimes interruptions such as ``mm-hm'' or ``okay'' come between sentences that would be one utterance if not interrupted. In the example below, s makes a three utterance answer that is interrupted by an acknowledgment. You are allowed to group a continuous series of utterances together into a segment to accommodate cases where several utterances form one answer, request, etc.
Segments can receive any label that an utterance can. However, a segment must be composed of a continuous set of utterances, so in the example below ``okay'' and ``and we need'' must be included in the ans segment. Since the segment is labeled as an answer, the individual utterances of the segment (utt2-utt5) do not also need to be labeled as answers. However, utt3 and utt4 should be given Answer tags (see section 2.4) of ``no'' since unlike utt2 and utt5 they are not an answer to a question. Except for acknowledgments such as ``okay'' and ``mm-hm'' and repaired utterances such as utt4, utterances within a segment should have the characteristics with which the segment is labeled. Section 5 discusses how to tag dialogs using the dat annotation tool and how to form and tag segments with the tool.
T1 utt1: u: where are the engines
ans| T2 utt2: s: there's an engine at Avon
| T3 utt3 u: okay
| T4 utt4 s: and we need
| | utt5 s: I mean there's another in Corning
You will be classifying each utterance unit along several dimensions (using Utterance-Tags) in order to record the unit's purpose and role in the dialog. Generally, the dimensions are orthogonal, and you can find examples of any possible combination of labels. We will explicitly point out a few places where this does not seem to hold.
Sometimes you will be unsure of whether or not a label should be applied to an utterance. One option you will have is to apply the label with an ``uncertainty modifier''. Section 5 discusses the dat annotation tool and how to mark uncertainty using it.
A special dimension called Communicative-Status is used to mark utterances that appear to be abandoned in mid-stream or which are impossible to understand. In general, if utterances are marked as being defective in one of these ways, you don't need to complete the rest of the annotation for that utterance.