320 likes | 470 Views
Software Development Processes , Reuse and Knowledge Sources in Spoken Dialog System. 11761 Dialog Seminar Rohit Kumar Friday, Dec 1, 2006. Some basic ideas / Ground rules Practical Dialogue Hypothesis Domain Independence Hypothesis Rapid Prototyping of Dialogue Systems
E N D
Software Development Processes, Reuse and Knowledge Sourcesin Spoken Dialog System 11761 Dialog Seminar Rohit Kumar Friday, Dec 1, 2006
Some basic ideas / Ground rules • Practical Dialogue Hypothesis • Domain Independence Hypothesis • Rapid Prototyping of Dialogue Systems • Dialogue system frameworks • Re-use • SLDS development methodologies & Best practices • Iterative Method • (DISC) Dialogue Engineering Life Cycle • Model View Controller (MVC) approach • Rapid Dialogue Prototyping (RDP) Methodology • User involvement in development • Methodologies i have used : 2 cases • Some “potential” rules of thumb ? • Knowledge Sources view of Dialogue Systems • Overview • Template based • Tools and Methods required • Some ideas Also included: Full list of readings
Some basic ideas / Ground rules • Practical Dialogue Hypothesis • Domain Independence Hypothesis • Rapid Prototyping of Dialogue Systems • Dialogue system frameworks • Re-use • SLDS development methodologies & Best practices • Iterative Method • (DISC) Dialogue Engineering Life Cycle • Model View Controller (MVC) approach • Rapid Dialogue Prototyping (RDP) Methodology • User involvement in development • Methodologies i have used : 2 cases • Some “potential” rules of thumb ? • Knowledge Sources view of Dialogue Systems • Overview • Template based • Tools and Methods required • Some ideas Also included: Full list of papers surveyed
Spoken Dialog System development is costly and time consuming • Two optimistic hypothesis (Allen et. al., 2000) • Practical Dialogue Hypothesis “The conversational competence required for practical dialogues, while still complex is significantly simpler to achieve than general human conversational competence” • Domain Independence Hypothesis “Within the genre of practical dialogue the bulk of complexity in the language interpretation and dialogue management is independent of the task being performed” • Agreeable ?! (I say “Yes”) Some basic ideas / Ground rules
Some basic ideas / Ground rules • Practical Dialogue Hypothesis • Domain Independence Hypothesis • Rapid Prototyping of Dialogue Systems • Dialogue system frameworks • Re-use • SLDS development methodologies & Best practices • Iterative Method • (DISC) Dialogue Engineering Life Cycle • Model View Controller (MVC) approach • Rapid Dialogue Prototyping (RDP) Methodology • User involvement in development • Methodologies i have used : 2 cases • Some “potential” rules of thumb ? • Knowledge Sources view of Dialogue Systems • Overview • Template based • Tools and Methods required • Some ideas Also included: Full list of papers surveyed
Dialogue System Frameworks • Can be considered a consequence or a partial proof of the domain independence hypothesis • Various Frameworks • TRIPS • Adriane / Ravenclaw / Olympus (CMU) • Communicator (SRI, CMU, U. Colorado, …) • Speech Builder (MIT) • CSLU Rapid Application Developer • Industrial Frameworks: Voice Platforms (IBM, Microsoft, Nuance…) • Tutorial Dialog: TuTalk • Many more… Rapid Prototyping of Dialogue Systems
(Denecke, 2002) • Able to develop very simple dialog systems with 8 – 10 hours of effort using the Adriane framework • Domain independent Interaction Patterns implemented • Question Undo, Correction, State • Development comprises of 7 steps • Backend Application • Databases • Ontology • Dialogue Goals • Database Conversion Rules • Parsing Grammars • Generation Templates Rapid Prototyping of Dialogue Systems
Rapid Prototyping of Dialogue Systems Observations Effort calculated as per the knowledge source. Rapid prototyping can be considered a method in which knowledge is added to a framework to customize it. All design exhibit lack of coverage in Language resources. Direct consequence of not doing any data collection ahead of time.
To reduce development costs: Re-Use • Can be considered to be the driving force being the Domain Independence Hypothesis (Hanna et. al. 2005) • While Domain Independence Hypothesis is acceptable, the challenge is to develop such an domain independent system • Part of the challenge is difference in terms of complexity of applications which in turn requires different degrees and forms of domain independence • Object oriented methodology to support extensibility and re-use of (dialog) components • All discourse object contain capabilities • On request, domain spotter polls and gets scores from each discourse object indicating object’s ability for successfully handle the request • Supports maintainability and re-use • Idea of hierarchies of increasing degrees of specialism Rapid Prototyping of Dialogue Systems
Rapid Prototyping of Dialogue Systems • Dynamic Linking of discourse objects • Keeps frames sizes small • Facilitates delegation and delimitation of agent responsibility • Open ended structure which can grow and adapt depending on objectives set by user on run time
Some basic ideas / Ground rules • Practical Dialogue Hypothesis • Domain Independence Hypothesis • Rapid Prototyping of Dialogue Systems • Dialogue system frameworks • Re-use • SLDS development methodologies & Best practices • (DISC) Dialogue Engineering Life Cycle • Iterative Method • Model View Controller (MVC) approach • Rapid Dialogue Prototyping (RDP) Methodology • User involvement in development • Methodologies i have used : 2 cases • Some “potential” rules of thumb ? • Knowledge Sources view of Dialogue Systems • Overview • Template based • Tools and Methods required • Some ideas Also included: Full list of papers surveyed
Dialogue Engineering Life Cycle • Based on DISC project on best practices for development and evaluation of SLDSs • Includes: Life Cycle model, Grid (set of issues and options), Set of Evaluation Criteria • SLDS Development Methodologies
Grid comprises of 6 aspects of SLDS • For each aspect, the grid lists various issues along with options (based on the available state of the art at that time) along with pros and cons • Choices to be made by the developers • Essentially and Iterative Life Cycle with 2 legs • Development Phases • Grid Issues are addressed in these phases • Evaluation • Evaluation criteria derived from choices made for the various Grid issues are applied here • This life cycle does not talk much about the phase after integration i.e. maintenance. • Also scope for re-use is not discussed • SLDS Development Methodologies
SLDS Development Methodologies • Evaluation • Analysis: Sufficiency and clarity of documents • Design: Design goals and constraints are sound, non-contradictory and feasible • Simulation: WOZ with questionnaires • Construction: Glass box and black box tests • Integration: Collected user system interaction is analysed
Iterative Method • Work iteratively from 2 angles • Conceptual Design • Framework customization • Solid Lines • Creative Progression • Dotted Line • How levels of the method correlate • E.g. Interpretation Module • Conceptual design starts with Parsing theory and Requirement Specifications • Framework customization starts with Parser and builds on corpus created from WOZ or other approaches • SLDS Development Methodologies
Iterative Method: Recommended Pre-requisites • Classification of possible dialogues and identification of main use-cases • Specification of requirements for the system behavior in each class of dialogues • Considering Dialog Management in Iterative Development • Design • Modularization • Knowledge Representation • Interfaces • Customization • Tools • Framework templates • Code Patterns • Identified Design choices • Dialogue History, User Request Handling, Sub-Dialogue Control • SLDS Development Methodologies
Model View Controller Approach: Very Industrial • Derived from well established GUI development practice • Possible to upload the user interface without modifying the application data or control logic • SLDS Development Methodologies
Rapid Dialogue Prototyping Methodology (RDPM) • Focuses on Dialogue components only • 5 steps • Producing the Task model • Draw relations between attributes and task to be performed • Deriving the initial dialogue model • Generic Dialogue Nodes (GDN) • Simple (Static): User inputs • List (Dynamic): Selects from menu • Internal: Start/resets, etc. • Local / Global dialogue flow management strategy • Using a WoZ experiment to instantiate the initial dialogue model • Using an Internal field test to refine the dialogue model • Using an external field test to evaluate the final dialogue model • SLDS Development Methodologies
(Atwell et. al., 2000) • Experience with user involvement in development • Particular case for CALL applications • Involve users at every stage • Two distinct classes of users • Prospective End-Users • Meta Level Experts • Interesting findings reported in the paper within the application domain which was possible due to user involvements at all stages • SLDS Development Methodologies
Methodologies I have used • Case 1: ConQuest dialogue system development • Case 2: CycleTalk dialogue system development • Sorry no time to make the slides for the above. But lets discuss • SLDS Development Methodologies
Some basic ideas / Ground rules • Practical Dialogue Hypothesis • Domain Independence Hypothesis • Rapid Prototyping of Dialogue Systems • Dialogue system frameworks • Re-use • SLDS development methodologies & Best practices • (DISC) Dialogue Engineering Life Cycle • Iterative Method • Model View Controller (MVC) approach • Rapid Dialogue Prototyping (RDP) Methodology • User involvement in development • Methodologies i have used : 2 cases • Some “potential” rules of thumb ? • Knowledge Sources view of Dialogue Systems • Overview • Template based • Tools and Methods required • Some ideas Also included: Full list of papers surveyed
Some basic ideas / Ground rules • Practical Dialogue Hypothesis • Domain Independence Hypothesis • Rapid Prototyping of Dialogue Systems • Dialogue system frameworks • Re-use • SLDS development methodologies & Best practices • Iterative Method • (DISC) Dialogue Engineering Life Cycle • Model View Controller (MVC) approach • Rapid Dialogue Prototyping (RDP) Methodology • User involvement in development • Methodologies i have used : 2 cases • Some “potential” rules of thumb ? • Knowledge Sources view of Dialogue Systems • Overview • Tools and Methods required • Some ideas Also included: Full list of papers surveyed
Dialogue system development can be seen as a process of putting knowledge into an existing framework in order to customize the framework • Tools and methods must • Provide visual languages for specification of structured dialogue data • Represent all relevant data and the dependences between them • Support formalism for defining constraints on the models and informing the user of violation of these models • Generate code that can be interpreted be generic dialogue system • Support the reuse for formerly developed domain models • Knowledge Sources View
Development Team & Requirements 16 15 10 11 14 User Logs 9 12 7 13 8 Dialog Task Specification Randomly created Corpus 6 5 2 3 4 Templates Grammar 1 Voice Domain Knowledge Lexicon Language Model Acoustic Model
Dialogue system development can be seen as a process of putting knowledge into an existing framework in order to customize the framework • Tools and methods must • Provide visual languages for specification of structured dialogue data • Represent all relevant data and the dependences between them • Support formalism for defining constraints on the models and informing the user of violation of these models • Generate code that can be interpreted be generic dialogue system • Support the reuse for formerly developed domain models • Knowledge Sources View
Some basic ideas / Ground rules • Practical Dialogue Hypothesis • Domain Independence Hypothesis • Rapid Prototyping of Dialogue Systems • Dialogue system frameworks • Re-use • SLDS development methodologies & Best practices • Iterative Method • (DISC) Dialogue Engineering Life Cycle • Model View Controller (MVC) approach • Rapid Dialogue Prototyping (RDP) Methodology • User involvement in development • Methodologies i have used : 2 cases • Some “potential” rules of thumb ? • Knowledge Sources view of Dialogue Systems • Overview • Tools and Methods required • Some ideas • Template based Also included: Full list of papers surveyed
Somewhere in between the spectrum of domain independence and domain dependence • Based on extensive in domain re-use • Changes in knowledge sources with every re-use • Issues • Incorporating knowledge changes • Propagation of knowledge changes while maintaining constraints • Hence view of the template system development would be • Which knowledge source is to be changed ? Some Ideas: Template based development
Readings • An architecture for Generic Dialogue Shell, J. Allen, D. Byron, M. Dzikovska, G. Ferguson, L. Galescu, A. Stent • Rapid Prototyping for Spoken Dialogue Systems, Matthias Denecke • Developing Extensible and Reusable Spoken Dialogue Components: An examination of the Queen’s Communicator, Philip Hanna, Ian O’Neil, Xingkun Liu, Michael McTear • Iterative implementation of Dialogue System Modules, Lars Degerstedt Arne Jonsson • The Dialogue Engineering Life Cycle, Laila Dybkjaer, Niels Ole Bernsen • DISC Website, www.disc2.dk • Rapid Dialogue Prototyping Methodology, Trung H. Bui, Martin Rajman, Miroslav Melichar • User Guided System Development in Interactive Spoken Language Education, E. Atwell et. al. • OpenSpeech Dialog: More powerful applications at lower cost, SpeechWorks white paper • A survey of knowledge sources in dialogue systems, Annika Flycht-Erikssn • Universal Dialogue Specification for Conversational Systems, Anke Kolzer