Please enable JavaScript.
Coggle requires JavaScript to display documents.
Social Talk Capabilities for Dialogue Systems (NLU (Domain classification…
Social Talk Capabilities for Dialogue Systems
NLU
part-of-speech analysis, named entity recognition, anaphora resolution and dependency parsing
Domain classification (p.84)
probabilistic approach
OoD classification
basis for recognition process - topics
Dialogue-act recognition (p.75)
based on syntactic and semantic relations
gathering linguistic info -> input for an information extraction component (delivers the relations embedded in the actual utterance) -> relations + additional features (a small dialogue context and mood of the sentence) as features for the machine-learning based recognition
Topic detection
Dialogue management
graph-based dialogue management
based on conversation threads
enables multi-threading support
embedded dialogue threads - are inserted into a mother thread as a completed unit
interwoven dialogue threads - two or more threads are alternated over a period of time
dialogue actions on lowest level; dialogue sequences = groups of dialogue actions; dialogues threads encapsulate an arbitrary number of dialogue sequences
Graphs: used in many applications, easy to develop, but lack flexibility (transition of the nodes beyond the predefined order is not possible), solution - thread-based implementation
component that decides what the system should do next + triggering all necessary supplemental steps such as embedded reasoning
finite-state dialogue management (p.63)
Multi-Threaded Dialogue Management (p.64)
characteristic for human conversation
Dialogue models
Error handling
misunderstandings
all erroneous interpretation results for the incoming utterances
Recovery: use confidence scores to detect misunderstandings + several strategies
non-understandings
out-of-domain utterances = out-of-application utterances - utterances targeting content outside of the knowledge domain of a system
strategies
new strategies (around 25)
common strategies: trying to repair the error (inappropriate reaction), ignoring the input or confessing the understanding problem (do not generate a conversational feeling)
Topic detection
Domain classification
all cases in which no interpretation could be assigned to an incoming utterance at all
Small talk = Social talk, social conversation / interaction
characteristics
there are specific rules -> not unrestricted and uncontrollable behaviour
comparatively unrestricted in content
people negotiate the social relationship between them
sensitive to violation of rules
dependent on culture
important in a SDS
The social-science theory of "face" forms the basis for the taxonomy used to order the dialogue acts - see Goffman 1967
Face - perception of the self by the interactions participating in a conversation
Faces need to be negotiated
2 main classes in the taxonomy
"request-face acts" - express a request for the support of the talker's face
"support-face acts" - utterances that strengthen the listener's face
Dialogue systems
classifications
modality
multi-modal systems - integrate more sensory information (gestures, facial/body expressions)
text-based systems - widespread in www
SDS - traditionally, originate from telephone software (call centres, hotlines)
mode of initiative
system-driven - the system always controls the flow of the dialogue. Very robust, no unpredicted conversation states occur
mixed-initiative
user-initiative - e.g. a simple question-answering system
Dialogue
a conversation between two or more par- ticipants or “agents”, making use of at least one change of speaker
Dialogue description
Speech acts
descriptions of the intentions encapsulated in utterances and therefore focus on the functional level of dialogue
commonly used to annotate dialogue data and develop dialogue models for DS
in further research also known under the names “dialogue acts”, “conversation acts”, “dialogue moves”, and many more
text analytics
Discourse Representation Theory, Rhetorical Structure Theory, Linguistic Discourse Model and the Penn Discourse Treebank
describe the semantic content of a discourse and the content-oriented relations between its elements, and do not focus on the intentional and functional aspect of discourse
more problematic that speech acts
Structure: dialogue acts -> dialogue sequences -> conversation phases
dialogue acts describe single dialogue actions
Examples of dialogue sequences: adjacency pair (e.g. question - answer, offer - accept / decline); the necessary succession of several actions to fulfil a sub-task
Conversation phases - just a few: opening phase, closing phase, and one core phase, e.g., for one special task
Dialogue threads - containers for one or several dialogue sequences which are grouped according to a specific functional goal
Computational dialogue models
Functions
describe the structure of dialogues in DS
are used to calculate how the system should act next
ways to formalise dialogue models in DS - p 41
Frame based
Plan based
Graph based
Probabilistic models
Architecture
Main components: NLU, dialogue management, output generation