NLP Notes

System transforms each T-H pair into logic form representaion with semantic relations

System automatically generates:
- NLP axioms serving as linguistic rewriting rules
- lexical chain axioms that connect concepts in T and H

COGEX logic prover(modified version of OTTER) is used to prove entailments using:
- semantic relations,
- WordNet lexical chains
- NLP axioms

The hypothesis is negated, and if it contradicts anything in T or anything inferred from T, return TRUE.

System Description:
- convert T and H into logic forms (Moldovan and Rus, 2001), which includes:

• pos tagging
• parse tree generation
• WSD (removed because of the inaccuracy)
  
• semantic relation detection

Input for COGEX to initiate the search for inferences:
- a list of cluases called the "set of support"

• negated form of H
• predicated that make up T

- a second list called "usable list", axioms generated either automatically or by hand.
They provide external world knowledge:

• knowledge of syntactic equivalence between logic form predicates, and
  
• lexical knowledge in the form of lexical chains.

World Knowledge Axioms:
- a small common-sense knowledge base of 310 world knowledge axioms:

• 80 manually designed (based on development data)
• 230 from previous projects

NLP Axioms:
- linguistic rewriting rules that help break down complex logic structures and express syntactic equivalence.
- automatically generated by the system through logic form and parse tree analysis.
- generated to break down complex nominals and coordinating conjunctions into their components so that other axioms can be applied to the components individually to generate a larger set of inferences.
- other axioms help:

• establish equivalence between prepositions
• establish equivalence between different POS
• equate words that have multiple noun forms
• equate substantives within appositions

WordNet Lexical Chains:
- WordNet provides links between synsets
- Each synset has a set of corresponding predicates for each word in the synonym set.
- predicate name is formed by synonym word form, pos, sense id
- predicates corresponding to noun synsets usually have a single argument
- predicates corresponding to verb synset three: event, subject, object

- A lexical chain is a chain of relations between two synsets.
- System prouces axioms using the predicates corresponding to the synsets in the relation.
EX:
buy_VB_1 (e1, x1, x2) -> pay_VB_1 (e1, x1, x3) (neden x3 ama x2 degil?)

- The following three classes of relations are used:

• pure WordNet relations
• generated from WordNet derivational morphology
• generated from WordNet glosses
  

Logic Prover:
- sort the "support list" in the order of weights
- drop the axioms with small weights
- seach the "usable list" for new inferences that can be made
- any inferences produced are assigned an appropriate weight depending on the axioms it is derived from
- continue until support list is empty
- if a refutation is found, return TRUE
- else relax the predicate arguments.
- if a refutation is found, return TRUE
- else drop predicates from the negated hypothesis until a refutation is found.

- earlier a refutation found, higher the confidence weight

accuracy: between 0.5 and 0.7
cws: between 0.4 and 0.6

The Paper: www.cs.biu.ac.il/~glikmao/rte05/fowler_et_al.pdf