May 29 2007, Geneva

• Database
  • Define format of internal meeting database
  • Convert IM2 database to internal format
• Tool for entering calendar info
  • GUI
  • Tool to extract vocabulary from calendar
    • Regulus lexicon
    • Pronounciation
• Corpus
  • Initial development corpus (Eng)
  • Initial development corpus (Fre)
• Grammar
  • English
    • lexicon
    • grammar extensions
      • "my next meeting", "my next three meetings"
  • French
• Dialogue move format
  • Define representation
  • Define mapping LF -> dialogue move
• Dialogue manager
• Response grammar
• Integration of first version
• Help system (also for MedSLT)
  • Word confidence scores from Regserver
    • Talk with Nikos C to get code for NASA SR3 module
    • First version of Regserver with word confidence scores by mid-July
  • Annotated corpus
    • Maria to enhance MedSLT GUI to add "annotation mode"
      • Fill in transcription: initialise with rec result
      • Ask user if recognition acceptable/unacceptable
      • If unacceptable, mark each help result as very helpful/helpful/not helpful (radio buttons?)
    • Collect initial round of data from U Geneva people
    • Real data collection round in September
  • Maria to implement SVM-based methods
    • Help system packaged as separate server

  =========================================================================

  Example 1

  Input: "when is my next meeting"

  LF: [[whq, [[tense,present], [verb,be], [subj, [[spec,the_next], [possessive, [[pronoun, i]]], [event,meeting]]], [[time,when]]]]]

  Role-marked LF:

  [[sentence_type, whq]-unmarked, [verb,be]-unmarked, [pronoun, i]-next [event,meeting]-next [time, when]-unmarked ]

  ---

  (Translation step)

  Interlingua (language-independent)

  Unresolved form:

  [[sentence_type, whq]-unmarked, [event, meeting]-next [attendee, pronoun(i)]-next, [time, wh]-unmarked ]

  ---

  (Resolve names and personal pronouns)

  Resolved form 1:

  [[sentence_type, whq]-unmarked, [event, meeting]-next [attendee, manny_rayner]-next, [time, wh]-unmarked ]

  (Resolve 'next' etc)

  Resolved form 2:

  Resolved form:

  [[sentence_type, whq], [meeting, meeting_1], [time, wh] ]

  Query form:

  Time^(meeting(meeting_1, Day, Month, Year, StartTime, EndTime, Place), describe_time_interval([Day, Month, Year, StartTime, EndTime], Time))

  Answer (uses output grammar):

  "May 29 2007, from 0930 to 1200"

  =========================================================================

  Example 2

  Input: "when does it end"

  [[whq, [[tense,present], [subj,[[spec,pro],[head,[[pronoun,it]]]]], [event,end], [[time,when]]]]]

  Role-marked LF:

  [[sentence_type, whq]-unmarked, [event,end]-unmarked, [pronoun, it]-unmarked [time, when]-unmarked ]

  ---

  (Translation step)

  Interlingua (language-independent)

  Unresolved form:

  [[sentence_type, whq]-unmarked, [meeting, [[pronoun, it]]]-unmarked, [end_time, wh]-unmarked ]

  ---

  (Resolve names and personal pronouns)

  Resolved form 1:

  [[sentence_type, whq]-unmarked, [meeting, meeting_1]-unmarked, [end_time, wh]-unmarked ]

  (Resolve 'next' etc)

  Resolved form 2:

  Resolved form:

  [[sentence_type, whq]-unmarked, [meeting, meeting_1]-unmarked, [end_time, wh]-unmarked ]

  Query form:

  Time^(meeting(meeting_1, Day, Month, Year, StartTime, EndTime, Place), describe_time([Day, Month, Year, EndTime], Time))

  Answer (uses output grammar):

  "May 29 2007, 1200"

  ---

  July 9 2007, Cambridge

  Next goals:

1. Package up Prolog functionality as socket-based server (Manny) DONE 1a Use PrologBeans if possible to interface to Java/.NET Talk with Nikos 1b Think about how to allow multi-modal dialogue management
2. Start integrating with mobile platform (Nikos - requires 1)
3. Extend input-manager patterns to allow multiple components, e.g. "when does ... start" -> start_time (Manny) DONE
4. Cover aggregates/ordinals (Manny) 4a. Extend general grammar to allow e.g. "next three", "my next" 4b. Extend input manager to capture scope of ordinal constructions (requires 4a) 4c. Extend dialogue manager to handle ordinal constructions (requires 4b)
5. Extend development/training corpus 5a. First version (Pierrette) 5b. Clean up (Manny - requires 5a)
6. Extend lexicon 6a Cover new vocabulary in training corpus (Pierrette - requires 5) 6b future with "going to" (Manny)
7. Extend database 7a. Add more future data 7b. Reformat IM2 past data
8. Think about how to make resolution non-deterministic, to allow introduction of Machine Learning methods
9. Think about help system May 28 2008

• Compile list of N-best problems
• Features
  • Def/indef and similar
    • Disprefer indefinite in general
    • Collocations of def/indef with verb
    • Collocations of sing/plur with verb
  • Disprefering repetition
    • Keep more context
    • Avoid repetition of e.g. "where"