Mid-April Project Plan for Transcriber R&D

May	Syntax Modules code review	Context adds change manager
June	Scanner and Diagrammer repairs	Context loads sentence charts
July	Mark up the loaded example inputs for phonetics, linking, prosody
August	Q/A: Transcribe paragraphs into output files plus domain lexikons
September	Demo: Transcribe paragraphs in training domain(s) to output files
Q4, Q1, ..	Refine Transcriber, while training it for new domain(s) of discourse

Typical Modules in a Text Transcriber (demo will not integrate these)

• Scanner - co-process to load example inputs to/from file folders
  
• Lexical - extraction of tokens from English input text
  
•  Morphology – assignment of lexemes to each token
• Diagrammer – filter lexemes by assembling parse trees
•  Semantics – filter trees by using phrase meanings
•  Discourse – anaphora; context-specific vocabulary
• Modeler - expands Context with charts of example sentences
  
• Pragmatics - JSPs drive Translation; I/O, interactive usage
  
• This adds LTI-custom code to Q/A the output markup
  
• This also embeds in Context a Domain Specific Lexicon
  
• Later, DSL shadows bulk lexical data (e.g, Webster's)?
• Translation – Java classes for application-specific output
  

Design of Demo Pragmatics Modules (Modeler JSPs customized for LTI):

• Lextend - uses a DSL in Context to find or build phonetic spelling
• lets Q/A operator define new record for unknown MEANING
• each normal record gets TERM spelling, model, POS
  
• we expand to add a Lessac-specific phonetic spelling
• Later, each DSL may merge into a CoreLex, and shadow it
• Validation - lets Q/A operator review and adjust sentence chart
• operator may call Exporter to hear the sentence played
• default prosody was generated by Translator rules
• operator may optionally edit the prosody and repeat
• approves best, with Q/A notes for off-line repair/review
• Demo Driver - calls Importer, then loops over Markup + Validation
• if no example input given, imports per a default listing
• uses DSL (or Lextend) to get all non-standard lexical data
• if called (by it or Validation) in mid P-graph, it just resumes
  
• after final P-graph sentence, it auto-loops to import next

Design of LTI-specific Translation Modules (Java Classes - mostly your IP):

• Importer - reads example paragraph into contextual P-graph
• initial format to test is just word stream built in Notepad
• working memory adopts Modeler's under TMAPI 1.0
  
• later, formats may come Scanner and/or Diagrammer
• Markup - expands a P-graph sentence to add markup for it
• DSL at this level always has a phonetic word spelling
• Linking rules are employed to join all words together
• Prosody is applied via P-graph data and rules TDBL
• Exporter - given any P-graph and sentence, it plays the markup
• comes mostly from audio team - I just dump a markup file
• can use any remote web app  (in  PERL?) to emit sound