Spreading activation in semantic memory using UMLS Metathesaurus

1
Spreading activation in semantic memory using
UMLS Metathesaurus
2
Overview

• Background
• in spite of many theoretical works there are no
  practical neurolinguistic algorithms available
• Objective
• a practical application of a neurolinguistic
  approach to text categorization/clustering is
  possible
• Specific aims
• comparison between neurolinguistic approach and
  text categorization/clustering methods that use
  prior knowledge (literature review)
• development of an algorithm that reproduces
  experts associative/semantic neural path of
  activations (cognitive experiments)
• investigation of properties of the algorithm and
  its application to medical text processing
  (examples, visualization, performance measures)

3
Semantic Memory

• Semantic memory is a personal and general
  knowledge of the world built by a process of
  abstraction (Baddeley, 1976).

4
Priming

• Priming is an improvement in performance in a
  perceptual or cognitive task, relative to an
  appropriate baseline, produced by context or
  prior experience (McNamara, 2005).

5
Spreading activation

• Spreading activation model is considered as the
  canonical model of semantic priming.
• There are three main assumptions in the spreading
  activation model (McNamara, 2005)
• Retrieving an item from memory amounts to
  activating its internal representation.
• Activation spreads from a concept to related
  concepts.
• Residual activation accumulating at concepts
  facilitates their subsequent retrieval.

6
Summary

• Neurocognitive linguistics so far
• computational models of neurocognitive phenomena
• practical applications are not possible (too slow
  for big data, not design to solve practical
  problems)
• Prior knowledge natural language processing so
  far
• practical applications (information retrieval,
  categorization, clustering, disambiguation)
• limited use of neurocognitive findings (no
  inhibition)
• Bridging the gap
• practical application (clustering,
  categorization)
• neurocognitive inspiration (add inhibition)

7
UMLS

• UMLS version 2007AC has
• 92 English sources (e.g. SNOMED CT, MeSH,
  ICD-9-CM, ICD-10)
• 54,245 ambiguous phrases
• 3,723,408 unique English phrases
• 1,516,299 concepts
• Concepts have
• 16,918,281 unique structural (semantic) relations
• 13,226,382 unique co-occurrence (associative)
  relations (e.g. PubMed medical subject headings
  co-occurrence)
• attributes, contexts, definitions, semantic
  types, ...

8
Example

• Four documents
• Patient has viral skin disease.
• Patient has staphylococcal skin infection.
  Staphylococcal infection is drug resistant.
• Patient has viral pneumonia.
• Patient has staphylococcal pneumonia.
  Staphylococcal infection is not drug resistant.
• How many categories/clusters?

9
(No Transcript)
10
(No Transcript)
11
(No Transcript)
12
How to find right model for inhibition?

• For parameter in parameters do
• For sample in bootstraps do
• compute performance based on
  document_concept_matrix
• For iteration in iterations do
• compute semantic_memory based on UMLS
• compute inhibited_semantic_memory based on
  parameter
• compute document_concept_matrix based on
  inhibited_semantic_memory
• compute performance based on
  document_concept_matrix
• End for
• End for
• End for

13
Cluster Users

• Why you decided to use the cluster for your
  research?
• Need more computational power
• How the cluster helped you towards your research?
• Faster to get the results
• What specific issues did you have to overcome to
  use the cluster effectively?
• None. I was already forking.
• Any data about your code performance in a
  serial/parallel environment and comparisons.
• Every N bootstrap samples are forked
• What else would you like to see in the cluster
  that could help your research?
• More CPUs on one node

14
(No Transcript)