Non Monotonic Reasoning for Disambiguation and Parsing

1
Non Monotonic Reasoningfor Disambiguation and
Parsing
Arthi Murugesan
2
Outline

• what you're doing overall
• tell us your plan
• what you've accomplished so for
• tell us your next few steps
• remind us of where you're going to get to in the
  end

3
Objective To deal with ambiguity in Parsing
using non-monotonic reasoning

• What does parsing ambiguity share with
  non-monotonic reasoning ?
• All conditions are not predefined
• Assume a default but can take it back
• bug insect , verb

4
Defining the problem

• Ambiguity
• Distorts a one to one relationship between a
  sentence and a meaning
• Types of ambiguity
• Lexical
• Structural
• Referential

5
Lexical ambiguity

• HPSG Parser Head-driven Phrase Structure
  Grammar
• Lexical Entry (Noun Version)
• Senses within POS animal, code error, listening
  device

6
Structural Ambiguity

• I saw an (astronomer) (with a telescope)
• I (saw) a mouse in the field (with a telescope)
• Ambiguity in attachment of phrases.

7
Other Ambiguities

• Referential Ambiguity
• Grounding in real world (Metonym)
• Pronouns
• Colloquial way of speech
• With stutters
• Changing sentences midway

8
What has been done

• The frog ate the bug.

det
NP
The-frog-ate-the-bug
VP
S
Ate-the-bug
ate
The-frog-NP
The-bug-NP
frog
The
bug
the
9
Existing Work

• HPSG Parsing mechanism has been used for
  predetermined unambiguous lexical entries.
• Phrases or Output assertions in the assumption
  base

10
Future Step

• To automate the selection of LEXICAL entry
• Probability ( POS tagger )
• Context Semantics
• Constraints
• Forms Valid sentences?
• Maybe earlier heuristics?
• Worst case - Alternate worlds
• Likely concept

11
Requirements

• Part of speech
• Probability ( POS tagger)
• Integration with a tool Penn tree bank or
  wordnet
• Net Input Prioritized list of lexical entries
  of a word
• If no sentence is formed
• A mechanism of taking back inferences
• A mechanism of introducing the lexical entry on
  the next greater priority - HARD

12
Requirements

• Context and Semantics
• Semantics of a sentence can be checked against
  existing knowledge
• Semantics of a parsed sentence should be
  generated (HUGE! Planning to underplay)
• Lambda expressions
• System's knowledge should in a format
  interoperable with parsers output knowledge.

13
Plan

• Shallow implementation
• Prioritized lists
• Semantic generation from parse tree
• Concentrate on
• Varying strengths of rules
• Understanding the process by which a proposition
  or rule is taken back
• Explicitly creating worlds controlling transfer
  of information between worlds

14
Expected Output Againand Theory of Working

• Example word bug
• Preference of Word senses according to WordNet
  (http//wordnet.princeton.edu/perl/webwn)
• Noun
• S (n) bug (general term for any insect or
  similar creeping or crawling invertebrate)
• S (n) bug, glitch (a fault or defect in a system
  or machine)
• S (n) bug (a small hidden microphone for
  listening secretly)
• Verb
• S (v) tease, badger, pester, bug, beleaguer
  (annoy persistently) "The children teased the boy
  because of his stammer"
• S (v) wiretap, tap, intercept, bug (tap a
  telephone or telegraph wire to get information)
  "The FBI was tapping the phone line of the
  suspected spy" "Is this hotel room bugged?"
• I saw a bug crawling on the carpet (or) The frog
  ate the bug -
• Using the first sense (Insect) No contradiction
• I bug my dad for money
• Using the first 3 senses (NOUN) will result in no
  valid sentence. So these lexical senses are to be
  retracted
• The first verb sense forms a valid sentence (
  tease, badger, pester or annoy)
• The spy used the bug to eavesdrop.
• The first sense (Insect) is used
• Forms a contradiction in meaning or Semantics
  insect cannot eavesdrop
• (Probably hack this contradiction for now)

15
Existing work

• http//citeseer.ist.psu.edu/anthony00reasoning.htm
  l
• Reasoning with Output from Parsing Using World
  Knowledge (2000)