Multiword Candidate Generation Processes:
UMLS_CUI with top endWords in the Distilled Medline N-gram Set

• N-grams from the latest Distilled Medline N-gram Set that are:
  • core-term, lowercase
  • Filter: exclude terms from Lexicon
  • Matcher: pass terms that are direct match in UMLS-Str (field 15) because such string have CUI
  • Filter: exclude single Words
  • Matchers: pass terms that match the top (33+) endWords from Lexicon
    • top endWords are: syndrome, protein, disease, proteins, cell, etc..
  • Pre-Preocess: 06.NGramUtil: Steps 20-21 (core.lc)
  • Pre-Preocess: Matcher EndWord: Steps 1 (10.MatcherEndWord)
    • flds 1 EndWord.1.analysis.stats > EndWord.1.analysis.stats.1
    • Get the top N end words (endWords.top.data.${YEAR})
  • Process: Matcher CUI: Steps 30-33 (09.MatcherCui)
  • Proocess: Matcher CUI: Steps 34 (09.MatcherCui)
  • Use Step 35/36 to rearrange the order in candidate list by grouping singulars and plurals together

  • Generated files:

    Distilled MEDLINE nGram Set	Candidate Files	Status	Notes
    2016	35.disNGram.Core.endword.out.gsp.2016	Done	Top 33 endWords
    2017	36.disNGram.Core.endword.out.rmYesTagNo.gsp.2017	Done	Top 43 endWords
    2018	36.disNGram.Core.endword.out.rmYesTagNo.gsp.2018	Done	Top 51 endWords
    2019	36.disNGram.Core.endword.out.rmYesTagNo.gsp.2019	Done	Top 57 endWords
    2020	36.disNGram.Core.endword.out.rmYesTagNo.gsp.2020	Done	Top 80 endWords
    2021	36.disNGram.Core.endword.out.rmYesTagNo.gsp.2021	Processing	Top 85 endWords

  • TBD: In the future, use all high frequency n-gram without endWord Matcher.
    => That is to use 33.disNGram.Core.multiword.out. However, this seems does not have high precision. Maybe use it with SpVars model or Deep learning model.