1 / 42

CSETalk: A Spoken Dialogue System for CSE Course Information

CSETalk: A Spoken Dialogue System for CSE Course Information. Preethi Jyothi, Rohit Prabhavalkar, Thomas Lynch, Deepak Bal, Prateeti Mohapatra. Outline. Introduction Automatic Speech Recognition Language Understanding Galaxy Overview Backend Ravenclaw Architecture Speech Synthesis

Download Presentation

CSETalk: A Spoken Dialogue System for CSE Course Information

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. CSETalk: A Spoken Dialogue System for CSE Course Information Preethi Jyothi, Rohit Prabhavalkar, Thomas Lynch, Deepak Bal, Prateeti Mohapatra

  2. Outline • Introduction • Automatic Speech Recognition • Language Understanding • Galaxy Overview • Backend • Ravenclaw Architecture • Speech Synthesis • Evaluation Metrics • Challenges Faced • Conclusions and Future Work

  3. Introduction • Dialog systems seek to provide a natural conversational interaction between the user and the computer system • Spoken input from the human user -> meaning of the utterance -> results of the operation to the user • Two-way flow of information • User-to-system • System-to-user

  4. Introduction • Types of Dialogue Systems • System-initiative • User-initiative • Mixed-initiative

  5. Motivation • All the components of a speech and language processing systems are integrated • Building a SDS is a challenging task due to the interaction of the various components that are involved • Goal: To implement an end to end spoken dialog system for the Course Information system using the RavenClaw dialog management framework

  6. Other Dialogue Systems • Saplen system 1997 [R. López-Cózaret et al. Eurospeech1997]: Food ordering system • Let’s Go! Bus Information System [Raux et. al. Eurospeech 2000] • ITSpoke: Intelligent Tutoring SDS [Litman and Silliman HLT-NAACL 2004]

  7. Recognition SPHINX Synthesis THETA Overall Architecture Lang. Understand. PHOENIX/HELIOS Dialog Manag. RAVENCLAW Back-end (various) Lang. Generation ROSETTA

  8. CSETalk: Course Information System • Information-based • Mixed type of dialogue system • The user can seek information about course numbers, credits, instructors, call numbers, etc. • Typical query • What course is Prof. X offering? • What is the call number for CSE-333?

  9. ASR – Sphinx II • Using the Sphinx II decoding engine • Uses semi-continuous hidden markov models • Used an off the shelf acoustic model (male and female voice ) • Pronunciation model and language model were built using Sphinx Knowledge Base Tool • Used an n-gram language model, built from a corpus of sentences generated randomly from system grammar

  10. Helios – Confidence Annotation • Generates a confidence score for the utterance based on information from ASR, Parser and Dialog manager • ASR • Number of words, 'unconfident' words • Parser • Uncovered words, transitions between parsed fragments, unparsed fragments, etc. • Dialog manager • State of dialog, concepts expected at current state, number of turns at current state, etc.

  11. Helios – Confidence Annotation • Allows detection and recovering from misunderstandings • Unfortunately, due to lack of time, we were not able to explore it fully

  12. Phoenix - Parser • The parser parses the word sequence into a set of semantic frames • A frame is a set of named 'slots' • Frame: [CSETalk] Nets: [QueryCallNum] [QueryInstructor] ... • Nets are compiled into RTNs

  13. Phoenix - Parser RTN for [QueryCallNum]

  14. Phoenix - Parser

  15. Galaxy Communicator http://communicator.sourceforge.net/sites/MITRE/distributions/GalaxyCommunicator/docs/manual/index.html

  16. What is Galaxy Communicator • Per the Galaxy Website: • Distributed • Message based • Hub-and-spoke Infrastructure • Optimized for constructing spoken dialogue systems • Origins: • Based on MIT Galaxy System • Currently Mitre Corporation /Darpa Communicator (Now Concluded). • Available at Sourceforge http://communicator.sourceforge.net/sites/MITRE/distributions/GalaxyCommunicator/docs/manual/index.html

  17. Galaxy Overview • Modular with Galaxy as the communications controller. • Everything is sent using sockets and frames. • Servers/modules can be located on the network or internet. • Galaxy transforms the information from text strings to sockets and back again. • Easy to interface with C++ or any other language with a sockets library. • Handles storage and forwarding for each independent module. • Monitors the status of the other modules. • Logs information pertaining to the other servers specified. (Very flexible.) • No need to negotiate communications parameters with the other modules. • Simple to add different package to handle some aspect of the system. • Write an interface from the new system to Galaxy. • For MySQL, changed the original backend server call to load and call MySQL. • The same strategy would work for other modules. • Downside: Galaxy is a single point of failure. If Galaxy goes down, everything is down and no one is notified.

  18. The Details • What its not: • Not an end-to-end dialogue system. • No run-time semantic standards • No Configuration-time semantic standards • But compatible with many standards like W3C Voice Browsers group's proposed specifications. • Knowledge Requirements: • You need to know C programming. • Some background in distributed processing. • RPC, CORBA, Java RM • Reasonable command of English - to read the documentation • Platforms: Sparc Solaris, Intel Linux and Win32 • Compiler: GNU gcc and make.

  19. The Exchange { query course_query courseId seven thirty three credits callNum instructor fosler-lussier ) Select * From courses Where courseId =“730” and instructor like “fosler-lussier”

  20. The Exchange { query refine_results results : 1 { { courseId "730" credits "3" callNum "04581-7" callNumM "04581-7" dow "T R" starttime "1130" endtime "1248" room "DL 0305" instructor "Eric Fosler-Lussier" title "Survey of Artificial Intelligence II: Advanced Topics U G 3" description "A survey of advanced concepts, techniques, and applications of artificial intelligence, including knowledge-based systems, learning, natural language understanding, and vision." ) } } aStr = Gal_GetObject(f, ":inframe"); inframe = Gal_StringValue(aStr); // Our Stuff next line GetCourseInfoC( inframe, &outframe ); aStr = Gal_StringObject(outframe); Gal_SetProp(f, ":outframe", aStr);

  21. Backend Server MySQL is the database. Possibly a different choice will be made next time. Maybe something simple like SQLite. Original Roomline calls a perl script’s for data retrieval. Many of the components also use perl scripts for various functions.

  22. Galaxy Summary • Makes Life Easy • It Glues Everything Together (Seamlessly?) • Do Not Start From Scratch! You’ll be sorry. • Modify an Existing System. • Great Piece of Software to Allow Experimenting with various components. • However, too many single points of failure!

  23. RavenClaw Architecture • Domain-specific dialog control logic specified • Most of the effort goes into creating this specification Dialog Task Specification (Domain Specific) Dialog Engine (Domain Independent) • Runs the Dialog Task Specification to specify the dialog control at runtime • Provides a large number of domain-independent conversational strategies

  24. Dialog Task Specification • Hierarchical plan for the dialog • Tree of dialog agents • Non-terminals - Dialog Agencies • Terminals - Fundamental Dialog Agents • Inform • Request • Expect and Execute

  25. Dialog Agents • Concepts – associated with agents • Concepts have pre-defined types • Set of value/confidence pairs • Structure of an agent • Execute routine – dependent on the agent type • Preconditions • Success/Failure criteria • Trigger conditions/ Trigger commands

  26. RavenClaw Dialog Engine • Functions using two data structures • Dialog Stack • Expectation Agenda • Functions in two phases • Execution Phase – Dialog agents executed from Dialog Stack • Input Phase – Uses Expectation Agenda to map user inputs to concept values

  27. Dialog Stack Expectation Agenda Dialog Task Specification Dialog Task Specification [captures domain-specific dialog control logic] Dialog Engine [domain-independent reusable component] Execution Phase

  28. Dialog Stack Expectation Agenda Dialog Task Specification Dialog Task Specification [captures domain-specific dialog control logic] Dialog Engine [domain-independent reusable component] Execution Phase CSETalk

  29. Dialog Stack Expectation Agenda Dialog Task Specification Dialog Task Specification [captures domain-specific dialog control logic] Dialog Engine [domain-independent reusable component] Execution Phase S: Welcome to CSE Talk ... Welcome CSETalk

  30. Dialog Stack Expectation Agenda Dialog Task Specification Dialog Task Specification [captures domain-specific dialog control logic] Dialog Engine [domain-independent reusable component] Execution Phase S: Welcome to CSE Talk ... CSETalk

  31. Dialog Stack Expectation Agenda Dialog Task Specification Dialog Task Specification [captures domain-specific dialog control logic] Dialog Engine [domain-independent reusable component] Execution Phase S: Welcome to CSE Talk ... Task CSETalk

  32. Dialog Stack Expectation Agenda Dialog Task Specification Dialog Task Specification [captures domain-specific dialog control logic] Dialog Engine [domain-independent reusable component] Execution Phase S: Welcome to CSE Talk ... Course_Id: [courseNumber] S: How can I help you? How May I Help You Task CSETalk

  33. Dialog Stack Expectation Agenda Dialog Task Specification Dialog Task Specification [captures domain-specific dialog control logic] Dialog Engine [domain-independent reusable component] Execution Phase S: Welcome to CSE Talk ... Course_Id: [courseNumber] Course_Id: [courseNumber] Course_Inst: [professor] … S: How can I help you? How May I Help You Task CSETalk

  34. Dialog Stack Expectation Agenda Dialog Task Specification Dialog Task Specification [captures domain-specific dialog control logic] Dialog Engine [domain-independent reusable component] Input Phase S: Welcome to CSE Talk ... Course_Id: [courseNumber] Course_Id: [courseNumber] Course_Inst: [professor] … S: How can I help you? How May I Help You Task U: Who’s teaching seven thirty? [courseNumber](seven thirty) CSETalk

  35. Language Generation - Rosetta • Template-Based • Attribute-Value pairs  Text for synthesis • Input from Dialog Manager, Output to TTS • Input Frame { act inform object welcome } • Output Welcome to CSETalk Automated Course Information….

  36. Speech Synthesis Within Galaxy, Kalliope manages the speech synthesis Is the link between Festival and Galaxy Hub Submits synthesis requests to Festival, holds a synthesis queue, etc Formatted to work with multiple synthesis systems (Festival, Swift, Theta)

  37. Speech Synthesis • Actual speech synthesized by Festival • Open Source speech synthesis software • Developed at University of Edinburgh and worked on extensively at CMU • Currently includes many types of voices • Diphone (what we use) • HMM based (referred to as HTS) • Unit Selection

  38. Speech Synthesis • CMU’s Festvox project hopes to make the building of new voices more systematic and better documented • Extensive instructions and tools on the web to help with the creation of new voices • Limited domain could be useful for our project • Uses unit selection methods

  39. Performance Evaluation • Many facets to measuring performance • Efficiency • Processing time • How much time is wasted with corrections, etc. • Quality • How many times did the system misinterpret • How many times did the user have to correct the system • User Satisfaction • Does the user feel happy with their interaction • Have user fill out survey • Task Success • Did the user leave with the information that they came for

  40. Performance Evaluation • One framework that tries to incorporate all of these facets is PARADISE (PARAdigm for DIalogue System Evaluation) • PARADISE proposes to compute a performance measure as a function of both task success and dialogue costs • If design changes were made, PARADISE could be used to evaluate the effectiveness of the changes.

  41. Challenges Faced • NO DOCUMENTATION!!  • Hard to localize faults since ASR/DM/NLG are tightly coupled • Festival runs best on UNIX whereas the rest of the system runs on Windows • Limited Time

  42. Conclusions and Future Work • Using more classes in the language model • Using other corpuses to obtain the acoustic model and the language model information • Creating or using better TTS voice • Actually evaluating performance as discussed

More Related