DIT++ Taxonomy of dialogue acts
Release 4, February 2010.
Differences with release 3.2 (February 2008) are listed here.)
The DIT++ taxonomy is a comprehensive system of dialogue act types obtained by extending the taxonomy of Dynamic Interpretation Theory (DIT),
originally developed for information dialogues (Bunt, 1994), with a number of dialogue act types from DAMSL (Allen & Core, 1997) and other schemas.
The DIT++ taxonomy forms a multidimensional system not only in the sense that it supports the assignment of multiple tags to markables, but also in
the sense that dimensions have a conceptual status, independent of annotation activities, as different aspects of communication
that may be addressed independent of each other by dialogue acts
(see Bunt, 2006)).
Applied to annotation, the multidimensionality of the schema means that a markable (i.e. a functionally relevant segment of dialogue behaviour)
may have more than one communicative function -- maximally one in each dimension.
In the taxonomy, dimensions are represented in boldface italic.
For convenience, the taxonomy is structured not only in dimensions but also in some additional groupings that do not have a theoretical significane, but
that are convenient for understanding the structure of the set of communicative functions, as well for referring to certain groups of funcions. Such groupings
are represented in italics.
Some communicative functions are specific for a certain dimension, the so-called "dimension-specific" functions.
Other functions are "general-purpose" in the sense that they can be used in every dimension.
In the presentation of the DIT++ communicative functions below, first the general-purpose functions are shown and subsequently the dimension-specific functions.
This document consists of three parts.
The first part shows the taxonomy of communicative functions, beginning with the general-purpose functions.
The hierarchical relations in the taxonomy, indicated by indentation, represent relative degrees of specificity of dialogue acts,
in the sense that a more specific act has stronger preconditions than a less specific act (which dominates it in the taxonomy);
in other words, the preconditions of more specific dialogue act logically entail those of any dominating act in the hierarchy.
A communicative function inherits all the preconditions of its ancestors in the hierarchy.
For instance, a Check Question is more specific than a Propositional Question because it has an additional precondition, concerning the speaker's expectation of the answer.
Similarly, a Confirm(ation) is more specific than a Propositional Answer.
This is reflected in the taxonomy by Check Question being dominated by Propositional Question, and Confirm by Propositional Answer.
The second part contains the definitions of all the communicative functions;
you can consult the definition of a communicative function by clicking on its name in the taxonomy.
The third part gives examples of the linguistic and/or nonverbal expression of these functions; to see examples, click on a definition.
The concepts from the DIT++ taxonomy have been applied and evaluated in a number of annotation efforts and in the design of a dialogue manager module that is capable of
generating multifunctional contributions to a dialogue. For the latter application, see
Keizer and Bunt, 2006 and
2007.
For its application to annotation,
see Geertzen and Bunt (2006)
and
Petukhova and Bunt (2007)
and Geertzen et al. (2007). For its use in other studies of dialogue see
Bunt (2007),
Roser Morante's PhD thesis (November 2007) and Bunt, Morante & Keizer (2007).
See also the
Annotation guidelines
for applying the DIT++ descriptors in the annotation of dialogues.
General-Purpose Communicative Functions
Answer Functions
Dimension-Specific Communicative Functions
- Activity-Specific Functions
- Functions, expressible either by means of performative verbs denoting
actions for performing activities in a specific domain, or by means of graphical actions such as
highlighting, or pointing to something in a picture. For example:
- Open Meeting, Suspend Meeting, Resume Meeting, Close Meeting (in meeting situations)
- Bet, AcceptBet
- Congratulation, Condolance
- Hire, Fire, Appoint,... (in a human resource management domain)
- Show, Highlight, Point, List,... for performing graphical/multimodal dialogue acts
- Dialogue Control Functions
Allo-Feedback
Interaction Management Functions
Turn Management
Time Management
Contact Management
Own Communication Management
Partner Communication Management
Discourse Structure Management
Social Obligations Management Functions
DIT definitions
(Click to see examples)
General-Purpose Communicative Functions
are functions that can be applied to any kind of semantic content. In particular, they can be applied not only to content information concerning a certain activity,
but also to information concerning the communication. In the latter case they form a `dialogue control act'.
For example, the utterance I did not hear what you said has the communicative function Inform, and in view of the type of is semantic content, it provides (negative) feedback about the speaker's perception of the previous utterance.
- Information Transfer Functions
In the definitions of Information Transfer functions, the hierarchical relations will be exploited by only specifying the way in which the preconditions of a communicative function strenghten those of its ancestors.
- Information-seeking functions: All varieties of questions, checks etc.
These all have a goal condition in common, namely that the speaker wants to know something.
Direct questions (such as Where is Harry's office?) carry the additional assumption that the addressee
knows the answer to the question;
indirect questions (such as Do you where Harry's office is?, or I would like to know where Harry's office is) do not carry this assumption. The various types of checks carry additional expectations (expressed by `weak beliefs') that the speaker have about the answer to his question, and about whether that expectation agrees with the addressee's beliefs about what is the correct answer.
- Indirect Propositional Question: S wants to know
the truth of a given proposition; S does not know whether A knows the truth of that proposition
- Propositional Question: S assumes that A knows the truth of that proposition
- Check Question: S weakly believes that the proposition is true
- Indirect Alternatives Question: S wants to
know which one from a given list of alternative propositions is true;
S believes that exactly one of the alternatives is true;
S does not know whether A knows which of the alternative propositions is true
- Information-providing functions: All information-providing acts
have in common that the speaker provides the addressee certain information which he believes the addressee
not to know or not to be aware of, and which he believes to be correct.
The
various subtypes of information-providing functions differ in the speaker's motivation for providing
the information; in different additional beliefs about what the addressee knows; and in differences in strength
of the speaker's trust in the correctness of the information that he provides.
In particular, if the speaker has only a weak belief concerning the correctness that information, then we have
'uncertain informs' and 'uncertain answers'.
- Informing functions are motivated by the speaker S having the goal that the addressee A
comes to know or to be aware of something.
- Uncertain Inform:
Speaker S wants to make the information that forms the semantic content of the inform known to addressee A;
S weakly believes that the information he provides is correct.
- Inform:
S believes that the information he provides is correct.
- Agreement:
S assumes that A weakly believes the semantic content to be true
- Disagreement:
S assumes that A weakly believes the semantic content to be false
- Correction: S wants the semantic content,
which he believes to be correct, to replace a belief by A that S believes to be incorrect.
(Such a belief that A holds may have its origin in S saying something wrong, so in a sense he corrects himself.)
- Informs with a Rhetorical, Emotional, or Evaluative Function: S is motivated by a semantic
relation of the kind defined in Rhetorical Structure Theory, or by certain emotions or evaluations such as fear and danger.
Rhetorical:
- Elaborate: Speaker believes that it is appropriate to provide additional or more detailed information about something that he
mentioned before.
- Exemplify: Speaker believes that it is helpful to give an example of something that he mentioned before.
- Justify: Speaker believes that he needs to provide information to support what he just said or to explain why he said it.
Emotional/evaluative:
- Warning: Speaker evaluates the situation, described in the semantic content, to be dangerous or
potentially harmful for Address ee
- ...
- Answers: All Answer functions have in common that the speaker provides certain information because
he believes that the addressee wants to have it, and which he believes the addressee assumes the
speaker to have. If the speaker is not certain of the correctness of the information that he
provides, then the answer is qualified as `uncertain'.
- Uncertain Set Answer: S believes that A wants to know which elements of a certain
domain have a certain property; S believes that A believes that he (S) knows that;
S believes that the set of entities, described by the semantic content, is the complete set
of all elements of that domain that have that property; S's belief is not certain.
- Set Answer: Same as Uncertain Set Answer, but now S's belief is not uncertain.
- Uncertain Propositional Answer: S believes that A wants to know the truth of the proposition that forms
the semantic content; S believes that A believes that S knows that; S believes
that his answer is true, but S is not certain about that.
- Uncertain Confirm: S believes that A weakly believes that
the propositional content is true.
- UncrtnDisconfirm:S believes that A weakly believes that
the propositional content is false.
- Propositional Answer: Same as Uncertain Propositional Answer, but now S's belief is not uncertain.
- Confirm: Same as Uncertain Confirm, but now S's belief is certain.
- Disconfirm: Same as Uncertain Disconfirm, but now S's belief is certain.
- Action Discussion Functions:
Action Discussion functions have a semantic content consisting of (a) an action; (b) a predicate describing a manner
or frequency of performing the action (a).
This frequency may be zero, so e.g. an Instruct to perform an action with frequency zero is the same as prohibiting
that action, and committing oneself to perform
an action with zero frequency is the same as committing oneself to not perform the action.
- Commissives:
S is committed to performing a certain action in a certain manner
or with a certain frequency,
possibly dependent on certain conditions,
and possibly dependent on A's consent that S do so.
- Offer:
S is committed to perform the action in the manner or with the frequency,
described in the semantic content, if A woud like S to do so
- Promise:
S is committed to
perform the action in the manner
or with the frequency, described in the semantic content
- AddressRequest:
S knows that A wants S to perform the action;
S is committed to conditionally perform the action, with conditions
(for instance concerning the manner or frequency of performing the action),
described in the semantic content
- AcceptRequest:
S is committed to unconditionally perform the action
described in the semantic content
- DeclineRequest:
S is committed to not perform the action described in the semantic content
(i.e., S is committed to perform the action with frequency zero)
- Address Suggestion:
S knows that A believes that the action,
described in the semantic content, is potentially promising for
achieving a certain goal;
S also knows that A believes that S is able to perform the action
(possibly together with A);
S is committed to conditionally perform the action, with conditions
(for instance concerning the manner or frequency of performing the action),
described in the semantic content
- AcceptSuggestion:
S is committed to unconditionally perform the action
described in the semantic content
- DeclineSuggestion:
S is committed to not perform the action described in the semantic content
(i.e., S is committed to perform the action with frequency zero)
- Directives:
S wants A to consider a certain action which A might carry out (possibly together
with S), potentially wanting to put pressure on A to do so
- Instruct:
S wants A to perform the action in the manner or with the frequency described
in the semantic content;
S assumes that A is able to do so
- Address Offer:
S believes that A is committed to perform the action described in the
semantic content
dependent on S's consent that A do so
- AcceptOffer:
S wants A to perform the action described in the semantic content
- DeclineOffer:
S wants A to not perform the action described in the semantic content
(i.e., S wants A to perform the action with frequency zero)
- Indirect Request: S wants A to perform the requested action
in the manner or with the frequency described, conditional on A's consent
- Request: S wants A to perform the requested action
in the manner or with the frequency described, conditional on A's consent;
S assumes that A is able to do so
- Suggestion: S wants A to know that the action in the manner or with the frequency described in
the semantic content, is potentially promising for achieving a certain goal, which
either S believes A to have, or which is specified as part of the semantic content;
S assumes that A (possibly toegether with S) is able to perform
the action in the manner or with the frequency described.
- Dimension-specific communicative functions: By contrast with
general-purpose communicative functions, dimension-specific functions are
only applicable to information concerned with a specific dimension of
communication.
- Activity-Specific Functions: Functions, expressible
either by means of performative verbs denoting (partly) communicative
actions for performing activities in a specific domain, or by means of graphical
actions such as highlighting, or pointing to something in a picture.
- Dialogue Control Functions: The functions of communicative
acts that serve to create or maintain the conditions for successful interaction.
- Feedback Functions:
Feedback acts provide or elicit information about the processing of he previous utterance(s),
where at least five levels of attending to an utterance and processing it are distinguished:
- attention, i.e. paying attention to the dialogue partner sufficiently to fully enable the perception of the partner's contributions
(e.g. listening, looking).
- perception, i.e. the recognition of the auditive, visual, or tactile components of communicative behaviour.
- interpretation, i.e. the assignment of meaning to the recognized communicative behaviour. In terms of dialogue acts,
this is the assignment of semantic content and communicative functions to utterances.
- evaluation, i.e. comparing the information that an utterance encodes, due to its communicative functions and
semantic content, with what was already known. For instance, when a question was asked to which,
according to the addressee, the questioner already knows the answer, then the addressee cannot
accept the information conveyed by the question, as this would put him in an inconsistent belief state.
- execution, also called 'application' or 'dispatch'. For instance, execution of a request or instruct is performing the
requested or instructed action; execution of a question is gathering the information to answer; executing
an answer is integrating its semantic content with the belief state.
Auto-Feedback acts are about the speaker's own attention and processing of an utterance in the addressee's last turn;
Allo-Feedback acts are about the speaker's beliefs about the addressee's attention and processing of an utterance in the speaker's last turn.
Dimension-specific Auto-Feedback functions are intended to signal that the processing of the utterance in question failed at a certain level
or was successful up to a certain level, ranging from attending via perceiving, understanding, and evaluating to doing something
with the result of the processing at all these levels ("execution").
(More articulate feedback acts, signalling or requesting help for a specific processing problem, are
constructed with general-purpose functions and a specific processing-related semantic content.)
- Auto-feedback functions:
- Positive (= Unspecified Positive):
S successfully processed the previous utterance, but provides no
information about the level(s) of processing being reported
- Negative (= Unspecified Negative):
S was unsuccessful in processing the previous utterance, but provides no
information about the level(s) of processing being reported
- Execution Positive (= Overall Positive) Feedback: S's
perception, interpretation, evaluation, and execution of the previous utterance
were successful.
- Evaluation Positive Feedback:
S's perception, interpretation, and evaluation of the previous utterance were
successful.
- Interpretation Positive Feedback:
S's perception and interpretation of the previous utterance were
successful.
- Perception Positive Feedback:
S's perception of the previous utterance was successful.
- Attention Positive Feedback:
S is paying full attention
- Execution Negative Feedback: S's
perception, interpretation, and evaluation of the previous utterance were
successful, but that he encountered a problem in applying the information
from that utterances (for example, S was unable to carry out an instruction,
or to find the information needed for answering a question).
- Evaluation Negative Feedback:
S encountered a problem in evaluating the semantic content of the previous utterance
(for example, the utterance provided information that is in conflict with
information already available to S).
- Interpretation Negative Feedback:
S's perception of the previous utterance was successful, but he encountered
a problem in trying to assign an interpretation to the utterance (for example,
S was unable to make sense of the semantic content).
- Perception Negative Feedback:
S's perception of the previous utterance encountered a problem (S did not hear
the utterance well, or was unable to read it).
- Attention Negative (Overall Negative) Feedback:
S did not pay (full) attention to the previous utterance (e.g., S did not listen carefully).
- Allo-Feedback Functions:
- Interaction Management Functions
- Turn management functions:
Turn management acts are those dialogue acts which are performed in order to
keep or to reallocate the speaker role. The beginning and end of a
turn, defined as an instance of communicative behaviour bounded
by lack of activity or another communicator's activity, are
associated with a reallocation of the speaker role. A turn ends
either because the current speaker assigns the speaker role to the
addressee, or because he offers the speaker role without putting any
pressure on the addressee to take the turn, or because the addressee
interrupts the speaker and 'grabs' the speaker role. Turn Assign
and Turn Release are thus two of the possible turn-final functions.
A turn may also include smaller units with boundaries where a
reallocation of the speaker role might have occurred, but where in fact it does not
occur because the speaker indicates that he wants to keep the
turn. Such a smaller unit then has Turn Keep function as the
unit-final Turn Management function.
A turn may also have a turn-initial function, indicating whether the
speaker of this turn obtained the speaker role by 'grabbing' it
(Turn Grab), by taking it when it was available (Turn Take) or by
accepting the addressee's assignment of the speaker role to him
(Turn Accept).
The units of Turn Management can thus have both a turn-unit-initial
and a turn-unit-final function, which is captured by giving them a
pair of functions, an initial and a final one.
- Time management functions:
- Stalling: S needs a little bit of time to formulate an utterance
- Pausing: S needs some time
to do something (either in preparation of continuing the dialogue, or because
something else came up which is more urgent for him to attend to)
- Contact management functions
- Contact Check : S wants to establish whether A is ready to receive messages from and
to send messages to S
- Contact Indication: S wants to make it known to A that S is ready to send messages to
and receive messages from A
- Own communication management functions:
- Error signaling: S wants A to know that S has made a mistake
in speaking
- Retraction: S wants to withdraw something that he said within the same turn
- Self-correction: S wants to correct an error that he made within the same turn
- Partner communication management functions:
- Completion: S wants to help A to complete an utterance
that A is struggling to complete
- Correct-misspeaking:
S wants to correct (part of) an utterance by A, believing that A made a speaking error
- Dialogue structure management functions:
- Social obligations management functions:
- Salutation
- Initial greeting: S wants A to be aware of S's presence; S is aware of A's presence;
S believes that S and A are in a position to exchange messages
- Return greeting: S wants A to be aware of A's presence; S is aware of A's presence;
S believes that S and A are in a position to exchange messages;
S is pressured to respond to a self-introduction by A
- Self-introduction
- Apologizing
- Apology: S wants A to know that S regrets having made an error in perceiving, understanding, evaluating, or executing
an utterance by A; or not having paid attention to, perceived well, or misunderstood an uuterance from A,
or being unable to evaluate or execute an utterance from A
- Apology-downplay: S wants to mitigate A's feelings of regret
- Gratitude
- Thanking: S wants A to know that S is grateful for what A has done in the current dialogue
- Thanking-downplay: S wants to mitigate A's feelings of gratitude
- Valediction
Examples
Sources:
'DIAMOND' = from DIAMOND project corpus of dialogues (in Dutch);
'IMIX' = from IMIX project corpus of dialogues (in Dutch);
'AMI' = from AMI project corpus of dialogues (in English)
'SCHISMA' = from SCHISMA project corpus of dialogues (in Dutch);
'OVIS" = from OVIS project corpus of dialogues (in Dutch).
General-Purpose Communicative Functions
Feedback Elicitation acts:
Partner Communication Management acts
Own Communication Management acts
Time management acts
- Stalling: Let me see
- Pausing: Just a moment; Een ogenblik alstublieft (OVIS)
Discourse structure management acts:
Different in release 3 from previous release (release 2, October 2006):
Terminological changes:
- "WH-Question" and "YN-Question" have been replaced by "Set Question" and "Propositional Question". respectively, bringing the terminology in line with that proposed
in the EU-project LIRICS and by the ISO expert group on standardization in linguistic resources. The terms "WH-Question" and "YN-Question" are dispreferred
because they suggest that these communicative functions depend on the use of WH-words, which is not true (and if it were true it would apply only to
English).
- "Check" has been replaced by "Check Question", again to be in line with ISO- and LIRICS terminologies.
- "Turn Give" has been replaced by "Turn Assign", to differentiate more clearly from "Turn Release".
Deleted:
- H-Question and Indirect H-Question. The term "H-Question" (or "How-Question") is rather unfortunate because of its suggested a one-to-one relation
to the use of particular words (notably the word how).
A more appropriate but rather cumbersome term would be "Non-exhaustive set question" - which would describe accurately what was intended by H-Question.
However, instead of making this terminological change, the definition of Set Question has been broadened so as to encompass both exhaustive and non-exhaustive
questions. (For example: "Where are the offices of some of your colleagues?" (non-exhaustive Set Question) versus
"Where are the offices of (all of) your colleagues?" (exhaustive Set Question).)
- The two types of Inform with a rhetorical function called Explain and Clarify. Annotators often found it very hard to distinguish between the two and between
either of them and Elaborate. The Informs with a rhetorical function form an open class anyway, which can be populated by functions as the need arises in a
particular annotation (or other) activity, so we have reduced the class in its base form to only contain Elaborate, Exemplify and Justify.
- Topic Closure (in the Topic Management dimension, which has been merged with the Dialogue Structuring dimension into the dimension called `Discourse
Structure Management' - see below). It was found that explicitly closing a topic is always done by means of dialogue acts with a general-purpose function
(such as Inform, Propositional Question, or Suggestion).
- Dialogue Act Announcement and Dialogue Act Invitation (in the Dialogue Sructuring dimension). As in the previous case, it was found that these functions
were always or nearly always performed by means of dialogue acts with a general-purpose function.
- Closing (in the Dialogue Structuring dimension), since it was found that explicitly closing a dialogue in spoken dialogue is always performed by a farewell
greeting in combination with some physical activity such as turning away or putting the phone down, which has the effect of breaking contact.
(In human-computer dialogue the situation may be different; a "quit" or "exit" may be defined for the purpose of closing a dialogue. However, it may still
be better in such cases to regard the act as one of breaking contact. If desired, a Closing function can of course be added back in.)
- Opening Indication and Opening Request (in the Dialogue Sructuring dimension) were merged into a single function called Opening, in order to avoid
confusions with the general-purpose functions Inform and Request, applied in the dimension of Dialogue Sructuring.
Added:
- Address Suggestion, for accepting or declining a suggestion.
- Positive Auto-Feedback, Negative Auto-Feedback, Positive Allo-Feedback, and Positive Allo-Feedback, as feedback functions which are unspecified for the
level of processing that they refer to. These have been added since it is in practice often impossible to say of a feedback act like Okay to which
level it refers.
Merged:
- The dimensions called `Topic Management' and `Dialogue Structuring' have been merged into the new dimension `Discourse Structure Management', since it was
found that these two dimensions were not at all independent (explicitly changing the topic, for instance, is a way to structure the dialogue).
Different in release 3, version 2 (February, 2008) from release 3.1:
In the definitions of commissive and directive functions a few typos have been corrected and minor improvements made.
In previous releases, the taxonomy included the functions Accept Request, Decline Request, Accept Offer, Decline Offer,
Accept Suggestion, and Decline Suggestion, which in release 3.1 were replaced by the more general functions Address
Request/Offer/Suggestion, in order to accomodate the occurrence of conditional acceptance of requests, offers and
suggestions. We found that the more specific Accept/Decline functions are also useful, however, so they have been
reintroduced as special cases of the Address functions.
Release 3.1: November 15, 2007
Last modified, as Release 3.2: February 8, 2008
<harry.bunt@uvt.nl>