What a thought is doing in a conversation
Analyzed through the 63 communicative functions of ISO 24617-2, the international standard for dialogue acts
The question
An utterance in a conversation does not always convey just one thing. Speakers might express an opinion while asking a question, or agree with someone while moving on to the next topic.
The first question AWAI takes on is how to describe “what one thought within an utterance is doing in the conversation.”
Thoughts and points
AWAI splits each utterance into “thoughts.” A thought is the part of an utterance that says one thing, and it is the basic unit of analysis. A thought can contain several “points”: individual facts, proposals or issues.
Functions are analyzed at the level of the thought.
The standard: ISO 24617-2
How people speak and express themselves varies endlessly, and the function of an utterance cannot simply be captured by a single classification. Rather than creating its own taxonomy from scratch, AWAI builds on criteria established by experts through data-grounded discussions: ISO 24617-2, the international standard for dialogue act annotation.
Behind the standard lie numerous earlier annotation schemes, such as TRAINS, Verbmobil, and ICSI-MRDA and AMI for meeting analysis. Because each was built for a particular purpose or domain, inconsistencies arose where the same concept went by different names and the same term could refer to different concepts. ISO 24617-2 reorganized these concepts on the basis of DIT++, a domain-independent multidimensional scheme (Bunt 2006, 2009). Its usability and coverage were validated through manual annotation of dialogues in English, Dutch, and Italian. The first edition (2012) was developed by a ten-member project team led by Bunt alongside a 35-member expert consultancy group, and it was revised into a second edition in 2020.
Participants in a conversation constantly exchange information not only to advance the task, but also about whether they understood each other, who speaks next, how time is used, and social interactions such as greetings and thanks. The standard organizes these into “ten dimensions.” Three criteria were established for selecting a dimension: that it corresponds to a well-studied, distinct kind of activity; that it is independent of the others; and that both humans and machines can recognize it. Their validity was confirmed through an analysis of 18 existing annotation schemes alongside statistical and machine-learning tests (Petukhova & Bunt 2009).
Even more important is multidimensionality: the idea that a single utterance can function in several dimensions at once. For example, “Right, so, on to the next point” signals that the speaker has understood while simultaneously shifting the topic.
The 63 functions
The functions AWAI assigns to a thought are the “63 communicative functions” defined in Annex F of the standard. These 63 functions are the sole basis for judgment; AWAI adds no functions of its own.
Of these 63 functions, 27 are general-purpose functions that can be used in any dimension and appear mainly in exchanges that advance the task. The remaining 36 are specific to the nine dimensions other than Task.
| Dimension | What it covers | Functions in the standard |
|---|---|---|
| Task* general-purpose functions, usable in every dimension | Advancing the purpose of the conversation itself: asking, informing, answering, agreeing or disagreeing, requesting, suggesting, committing | questionsetQuestionpropositionalQuestioncheckQuestionchoiceQuestiontestQuestioninformagreementdisagreementcorrectionanswerconfirmdisconfirmofferpromiseaddressRequestacceptRequestdeclineRequestaddressSuggestacceptSuggestdeclineSuggestrequestinstructaddressOfferacceptOfferdeclineOffersuggest |
| Auto-Feedback | Showing how the speaker processed what the other person said | autoPositiveautoNegative |
| Allo-Feedback | Responding to, or checking, how the other person processed what the speaker said | alloPositivealloNegativefeedbackElicitation |
| Turn Management | Taking, keeping and giving up the turn | turnTaketurnAcceptturnGrabturnKeepturnAssignturnRelease |
| Time Management | Taking a moment to think; asking the other person to wait | stallingpausing |
| Contact Management | Checking whether the other person is ready to talk, or showing that the speaker is | contactCheckcontactIndication |
| Own Communication Management | Noticing, retracting, or correcting one's own speech errors | selfErrorretractionselfCorrection |
| Partner Communication Management | Completing the other person's words or correcting their speech errors | completioncorrectMisspeaking |
| Discourse Structuring | Opening, structuring the interaction, shifting the topic | openinginteractionStructuringtopicShift |
| Social Obligations Management | Greetings, introductions, apologies, compliments, congratulations, sympathy, thanks, goodbyes | initGreetingreturnGreetinginitSelfIntroductionreturnSelfIntroductionapologyacceptApologycomplimentcongratulationsympathyExpressionthankingacceptThankinginitGoodbyereturnGoodbye |
How AWAI applies the standard
AWAI interprets and applies the standard as follows.
- The unit is the thoughtThe standard annotates dialogue by dividing it into functional segments, the smallest stretches that carry a function. AWAI instead assigns functions to the thoughts it extracts from an utterance. Because a thought is the part that says one thing, a single thought can carry several functions, aligning with the standard's principle of multidimensionality.
- Several functions per dimensionFor each thought, zero or more functions are judged in each dimension.
- An LLM makes the judgmentThe standard's definitions are provided to an LLM in the prompt, and the LLM judges functions against them. From its inception, ISO 24617-2 included automatic annotation from transcripts among its intended use cases. Today's LLMs make that automatic annotation possible by applying the standard's definitions directly as written.
Six groups
Listing all 63 functions makes it difficult to see overall trends across a conversation. To make the results easier to take in at a glance, AWAI displays the functions assigned in the Task dimension (the 27 general-purpose functions) organized into “six groups.” The grouping follows the four families the standard defines for general-purpose functions (information-seeking, information-providing, commissive, and directive). The judgment itself is always made on the 63 functions; the six groups serve only for display.
| Group | Meaning | Functions included |
|---|---|---|
| Question | asks for information or an answer | questionsetQuestionpropositionalQuestioncheckQuestionchoiceQuestiontestQuestion |
| Inform | states information or a view | inform |
| Response | answers or agrees / disagrees with the other side | agreementdisagreementcorrectionanswerconfirmdisconfirm |
| Request / proposal | asks or suggests that the other side do something | requestinstructsuggestaddressOfferacceptOfferdeclineOffer |
| Commitment | promises or offers to do something | offerpromiseaddressRequestacceptRequestdeclineRequestaddressSuggestacceptSuggestdeclineSuggest |
| Other | keeps the conversation going (greetings, acknowledgements) | A thought with no function in the Task dimension |
In the standard, Inform and Response both belong to the information-providing family. AWAI treats them as separate groups because whether a thought responds to what the other person said is essential for reading the flow of a conversation.
A thought with several functions can also appear in two or more groups.
Where misreadings still occur
Misreading the function of an utterance occurs even among human participants in a conversation. An intent that is clear when reviewing the entire conversation in retrospect can be difficult to judge in the moment because the necessary context is not yet available. Furthermore, the elements of a function are often present without appearing clearly on the surface of the words.
AWAI faces the same difficulty. Functions that sustain conversation flow, such as backchannels and turn-taking, are especially difficult to read from text alone, and the LLM that makes the judgment still misreads some of them. We expect this to improve as more cues become available.
References for this page
- ISO 24617-2:2020. Language resource management — Semantic annotation framework (SemAF) — Part 2: Dialogue acts. www.iso.org/standard/76443.htmlChecked against the ISO/DIS 24617-2 text published by its authors (2019; the version approved in the international ballot of February 2020): dit.uvt.nlNote — what this page draws on:The ten dimensions for describing utterance functions (§8.2) and the criteria for selecting them; the definitions of the 63 communicative functions in Annex F (which AWAI provides to the LLM); multidimensionality, where a single utterance functions across several dimensions (§7); the standard's background (§4.1); and the inclusion of automatic annotation among its use cases (§4.2).
- Bunt, H., Petukhova, V., Gilmartin, E., Pelachaud, C., Fang, A., Keizer, S. and Prévot, L. (2020). The ISO Standard for Dialogue Act Annotation, Second Edition. Proceedings of LREC 2020, pp. 549–558. aclanthology.org/2020.lrec-1.69Note — what this page draws on:The addition of the Contact Management dimension in the second edition, bringing the total to ten, and the continued validity of annotations made under the first edition.
- Bunt, H. (2006). Dimensions in Dialogue Act Annotation. Proceedings of LREC 2006, Genoa, pp. 919–924.Note — what this page draws on:The principle of organizing dialogue act categories multidimensionally across dimensions, which ISO 24617-2 builds upon. Referenced on this page as cited in the standard itself (§7.2).
- Bunt, H. (2009). The DIT++ taxonomy for functional dialogue markup. Proceedings of the AAMAS 2009 Workshop EDAML, Budapest, pp. 13–24.Note — what this page draws on:The DIT++ taxonomy on which ISO 24617-2 is based. Referenced on this page as cited in the standard (§4.1) to explain the background of the standard.
- Petukhova, V. and Bunt, H. (2009). The independence of dimensions in multidimensional dialogue act annotation. Proceedings of NAACL-HLT 2009, Short Papers, pp. 197–200. aclanthology.org/N09-2050Note — what this page draws on:Validation of the mutual independence of dimensions. The standard (§8.2) states that its ten dimensions were selected on the basis of the series of analyses that includes this work.