A Strategy for Automatic Corpus Generation

1
A Strategy for Automatic Corpus Generation

• Alison Alvarez
• Language Technologies Institute
• Carnegie Mellon University

2
Research Background

• If you encountered a new language, what question
  would you ask first?
• Answer Does it have a bilingual aligned corpus?
• Our solution for minority languages add one
  bilingual consultant and the Elicitation Tool

3
The Elicitation Tool
4
How We Represent Sentences

• Main Components
• Vector of Head Containing Phrases
• Tree Structure
• Surface Text

5
Feature Structures

• What is a Feature?
• What is a Feature Structure?
• One for the pronoun she
• What are they used for?
• Corpus Navigation
• Sentence Generation
• Feature Detection (Next Slide)

6
Feature Detection

• Compare sentences with highly similar feature
  vectors
• Note the differences in translation
• Big problem Combinatorics

7
The Generation Process Resources

• Mini-Copora of Phrasal Components
• Includes all possible feature combinations
• Lists Subcategorizations
• A Phrase that will be put in the sentence
• Uninflected for things outside of phrase
• Example A VP that could be inflected by a 3rd
  person Singular subject
• Inflected for things inside the phrase
• Example Most plural nouns
• A predicate field for the head
• A Comment Field
• Can itself be generated

8
The Generation Process Resources

• A set of tree structures with substitution sites
• Each site has a list of selectional restrictions
• A transitive sentence
• S(NPSubj VP(VS NPObj))
• VS verb sequence
• Also includes an subcategorization description
• The subcategorization will be used so that the
  right trees get put with the right verbs, etc

9
The Generation Process Resources

• The Lexicon
• The most of the lexical description will come
  from the feature vector
• Will simply list selectional restrictions
• Example for to paint
• Subject Human, Object Inanimate
• Can vary based on tree structure if needed
• Example for to be
• Subject Gender Object Gender, Subject Number
  Object Number, Subject Animacy Object
  Animacy

10
The Generation Process Resources

• A collection of word inflections
• the original uninflected word, not always in
  infinitive form
• its inflections
• And their corresponding circumstances
• Organized by part of speech
• Examples
• Have Subject 3rd Singular has
• Is Subj 1st sg am, Subj pl are, Subj 2nd
  are
• Mouse pl mice
• Most nouns will already be inflected

11
The Generation Process Algorithm

• Start with the Mini-Corpus for Verb Sequences
• For each verb, select the appropriate trees
• Replace the VS tag with the uninflected verb if
  it is not restricted
• For each partially filled in tree
• Fill in the appropriate slots with words that can
  unify
• Eliminate all that dont meet the selectional
  restrictions of its mother

12
The Generation Process Algorithm

• Once all of the Phrases are filled in
• Create a new sentence structure
• Create full phrasal vector
• Concatenate phrases together for surface text
• Inflect everything that needs to be inflected
• Match up head words with words from head word
  collection
• Only checks for verb sequence heads in current
  implementation
• Check for matching circumstances
• Replace old word with inflected word

13
Examples

• She paints it
• She is a mother
• She is an idea
• Paint the house
• I think that she likes Bill
• I know to run
• The implementation

14
Current Problems/Future Plans

• How to do long distance attachment
• Fluency/Lexical Matching She went up a ladder
  and up an aardvark
• Dont put in so many words
• Write a lot of rules
• Make sure every combination is covered despite
  selectional restrictions
• Expand the Implementation