1. Module name: Verbmobil dialogue acts [VM-DA].

  2. Module purpose: Dialogue act coding for appointment scheduling and hotel reservation dialogues.

  3. Coding level: Dialogue acts.

  4. Data source: BAS spoken dialogue corpora for Verbmobil.

  5. Module references: canonical level/phonetic transcription module.

  6. Markup declaration:

    7. ELEMENT   DAS
    ATTRIBUTES   id: ID   href: HREF(canonical level)
    ELEMENT   DA (child of DAS)
    ATTRIBUTES   id:ID   da: TEXT (label)   direction: TEXT (speaker, listener)

  7. Description:

    8. da value: "GREET" | "BYE" | "INTRODUCE" | "POLITENESS_FORMULA" | "THANK" | "DELIBERATE" | "BACKCHANNEL" | "INIT" | "DEFER" | "CLOSE" | "REQUEST" | "REQUEST_SUGGEST" | "REQUEST_CLARIFY" | "REQUEST_COMMENT" | "REQUEST_COMMIT" | "SUGGEST" | "INFORM" | "DIGRESS" | "DEVIATE_SCENARIO" | "REFER_TO_SETTING" | "EXCLUDE" | "CLARIFY" | "GIVE_REASON" | "EXPLAINED_REJECT" | "FEEDBACK" | "FEEDBACK_NEGATIVE" | "REJECT" | "FEEDBACK_POSITIVE" | "ACCEPT" | "CONFIRM" | "COMMIT" | "OFFER" | "NOT_CLASSIFIABLE"
    direction value: speaker id "," listener id

  8. Example:

    9. <DAS id="e032ach_RGM_DAS_9" href="e032ach_RGM_KAN.xml#id(e032ach_RGM_TRN5_x0)..id(e032ach_RGM_TRN5_x1)">
       <DA id="e032ach_RGM_DA_9" da="CLARIFY" direction="RGM,DNC"/>
    </DAS>
    Here assuming that the basic transcription is in file e032ach_RGM_KAN.xml and contains the lines:
    <KAN id="e032ach_RGM_TRN5_x0">Vv</KAN>
    <KAN id="e032ach_RGM_TRN5_x1">n@Uvemb=r</KAN>

  9. Coding procedure:

    10. This procedure describes how to train naive human annotators from the start to ensure relatively homogenous annotations. Time estimates are based on student worker contracts with 8 hours per week.

    First stage: 1/2 week
    Introduction to theoretical concepts (dialogue acts) and motivation (statistical recognition) providing some example annotations.

    Second stage: 1/2 week
    Annotators observe annotation process of each other together with supervisor. The supervisor explains unintuitive and special cases (e.g. EXPLAINED_REJECT, SUGGEST in negative mode, CLARIFY, GIVE_REASON before argument, etc.), the segmentation process, and introduces specialised terminology, like turn, utterance, and segment.

    Third stage: 1 week
    - supervised annotation of five selected dialogues
      (supervisor corrects instantly)
    - unsupervised annotation with correction of ten dialogues
      (as often as necessary)

    Fourth stage:
    Normal annotation with random checks. Each annotator annotates different dialogues. Discussion sessions on difficult annotation cases together with supervisor happen twice a month.

  10. Creation notes:

    11. Authors: Michael Kipp, Marion Klein.
    Version: 1 (1999/04/13)
    Comment: Created for use in the Verbmobil project. Serves as an example for the MATE project.
    Literature: [Alexandersson et al. 1998].