Mikrokosmos Text Generation
Brief Introduction to Text Generation as Choice
-
TMRs - show sample sentence tmr
-
ontology
-
events
-
Objects
-
Relations
-
Case Role Relations
-
Textual/Discourse Relations
-
Attitudes
-
Modalities
-
Aspect/Temporal
-
scripts
-
Non-literal meaning
-
Output
-
Output Display Tool
Lexicons
-
It is instructive to look at the difference between analysis and generation
lexicons:
-
Lexicon Content
-
Analysis Lexicon
-
input: words
-
indexed on words
-
different meanings - sense disambiguation
-
semantic constraints important (example)
-
one ``correct'' answer
-
planning unimportant
-
preferences unimportant (i.e. collocational info not used because we already
know that they go together since they are in input)
-
Generation Lexicons
-
Generation Lexicon Organization
-
Learn
-
input concepts
-
indexed on concepts
-
meaning already known
-
semantic constraints less important
-
many possible ``correct'' answers
-
planning important
-
preferences important (i.e. collocational info used)
-
Analysis vs. Generation Organization
-
semantics to match on
-
complex mappings
-
matcher combinations
-
possible mismatches
-
slots missing in lex entry -> undergeneration penalty, plan separately
-
extra slots in lex entry -> overgeneration penalty
-
slot filler discepancies -> penalties
-
Example of subcat entries for MERGE to show penalties
-
Options set up by lexicon entries for events, relations and properties.
Examples.
-
Decision-Aids
-
lexicon-based decision-aids
-
stylistic
-
collocational
-
grammaticall constraints
-
combination-based decision-aids
-
pick one or two of the following and explain, then say we are working on
many others.
-
global stylists (i.e. limit nesting of PPs)
-
discourse preferences on main verb selection
-
semantic content determinations (short chunks as verb subcategorizations,
longer chunks as subordinate clauses)
-
``goal'' driven: prefer ordering that maximizes coreference possibilities
(pronoun, definite descriptions)
-
structure-preserving (backup)
-
...
-
discussion on place of focus?
-
Options set up by discourse lexicon entries and sentence boundary rules.
-
Decision-Aids.
-
Structure-preserving
-
makes sense at this level since it is mostly concerned with information
packaging. If the SL grouped it together, there's a good chance the TL
can too.
-
some other heuristic
Coreference Chain Processing
-
briefly go over HG methodology
-
advantages for Generation
-
Allows ``exhaustive'' look at local combinations.
-
No need to make early decisions.
-
Allows interacting constraints, constraints at any level, while still utilizing
modular, declarative knowledge
-
Guarantees optimal answers (as measured by preferences)
-
No separate microtheory needed to generate sentence-level structure.
-
Speed
Knowledge Source Acquisition
-
general background
-
comparison to Elhadad and Diogenes
Future Work
-
Expand heuristics for preferences
-
knowledge source updates
-
Surface Generator interfaces - intermediate results
-
.Enhance Analyzer Outputs
-
...