Skip to content

Latest commit

 

History

History
829 lines (563 loc) · 23.6 KB

Design.md

File metadata and controls

829 lines (563 loc) · 23.6 KB
Raw

MetaMap Lite

System Description

Processing stages

MetaMapLite segments text using sentence or blank line markers using OpenNLP's sentence segmenter or MetaMapLite's blank line seqmenter. Tokenizes sentences using tokenizer based of Original MetaMap's tokenization regime. Part-of-speech information is added to tokenized text using OpenNLP's part-of-speech tagger. Mapping is done by dividing sentence or line based chunks into sublists and then each sublist is normalized and then looked up in a dictionary. Any match found in the dictionary that is subsumed by a longer match is discarded.

Description

Segments text using sentence or blank lines using OpenNLP's sentence segmenter or MetaMapLite's blank line seqmenter. Tokenizes sentences using tokenizer based of Original MetaMap's tokenization regime. Part-of-speech information is added to tokenized text using OpenNLP's part-of-speech tagger. Mapping is done using dictionary lookup discarding any match subsumed by longest match found in the dictionary.

input text ->
  sentence/line segmentation ->
    tokenization ->
      part-of-speech tagging ->
        token window generation ->
          term normalization ->
            concept dictionary lookup ->
              negation detection ->
			  result presentation
  1. add support for exclusion of some cui/term combination (see MetaMap's special terms file) [done?]
  2. support entity lookup using two separate dataset (UMLS and custom, etc.)

Below is an example of sentence level named entity recognition in which a token list for the sentence "Papillary Thyroid Carcinoma is a Unique Clinical Entity." is processed.

Given the example:

    "Papillary Thyroid Carcinoma is a Unique Clinical Entity."

The token-based sublists are produced as follows:

    "Papillary Thyroid Carcinoma is a Unique Clinical Entity"
    "Papillary Thyroid Carcinoma is a Unique Clinical"
    "Papillary Thyroid Carcinoma is a Unique"
    "Papillary Thyroid Carcinoma is a"
    "Papillary Thyroid Carcinoma is"
    "Papillary Thyroid Carcinoma"   --> match
                                "is a Unique Clinical Entity"
                                "is a Unique Clinical"
                                "is a Unique"
                                "is a"
	                            "is"
                                   "a Unique Clinical Entity"
								   "a Unique Clinical"
								   "a Unique"
                                   "a"
							         "Unique Clinical Entity"
								     "Unique Clinical"
								     "Unique" --> match
								            "Clinical Entity"
								            "Clinical" --> match
                                            "Entity" --> match

Four entities are found by MetaMapLite:

   "Papillary Thyroid Carcinoma"
   "Unique"
   "Clinical"
   "Entity"

Term Normalization

MetaMapLite normalizes the terms in the current token window using a Java implementation of of MetaMap's Prolog predicate mwi_utilities:normalize_meta_string/2 (used by filter_mrconso) with some slight modifications (including the removal of some operations.)

The following operations are performed of the term before dictionary lookup:

  • removal of (left []) parentheticals;
  • syntactic uninversion;
  • conversion to lowercase;
  • stripping of possessives.

Dictionary Lookup

The underlying dictionary lookup uses a inverted file implemenations where the dictionary is divided in to several partitions where each partition contains only terms which have the same termlength. The implementation uses the Java NIO class java.nio.MappedByteBuffer to access the systems virtual memory facilities to improve I/O performance.

Negation Detection

Wendy Chapman's ConText is used for negation detection using the current sentence and the list of found entities.

An Alternative Implementation of Negation Detection

Dealing with Conjunctions

Given sentence:

"Tylenol was effective in patients without fever but effective in patients with headache."

Separate utterances delimited by conjunction "but", "cause for", etc.

"Tylenol was effective in patients without fever"  (negate "fever", trigger: "without")
"but"
"effective in patients with headache"              (do not negate "headache")

Given sentence:

"The patient denies chest pain but has no shortness of breath."

Similarly, the independent clause delimited by the coordinating conjunction may each contain a negation phrase.

 "The patient denies chest pain"  (negate "chest pain", trigger: "denies")
 "but"
 "has no shortness of breath."    (negate "shortness of breath", trigger: "no")

Dictionaries

MetaMapLite currently uses three dictionaries originally created for MetaMap:

cuiconcept
cuisourceinfo
cuist

cuisourceinfo

This is the table used for mapping strings in the input text to concepts, this is the primary dictionary used by MetaMapLite.

Each record contains the UMLS concept identifier, UMLS string identifier, a sequence number, the source derived string, the source abbreviation, and the source term type:

cui|sui|seqno|string|src|tty

Sample records are below:

C0000005|S0007492|1|(131)I-Macroaggregated Albumin|MSH|PEN
C0000005|S0007491|2|(131)I-MAA|MSH|EN
C0000039|S0007564|1|1,2-Dipalmitoylphosphatidylcholine|MSH|MH
C0000039|S0007564|2|1,2-Dipalmitoylphosphatidylcholine|NDFRT|PT
C0000039|S0007564|3|1,2-Dipalmitoylphosphatidylcholine|MTH|PN
C0000039|S1357296|4|1,2 Dipalmitoylphosphatidylcholine|MSH|PM
C0000039|S0007560|5|1,2-Dihexadecyl-sn-Glycerophosphocholine|NDFRT|SY
C0000039|S0007560|6|1,2-Dihexadecyl-sn-Glycerophosphocholine|MSH|EN
C0000039|S1357276|7|1,2 Dihexadecyl sn Glycerophosphocholine|MSH|PM
C0000039|S0007563|8|1,2-Dipalmitoyl-Glycerophosphocholine|NDFRT|SY

cuiconcept

This table maps concept identifiers to concept preferred names.

format:

cui|preferred name

Some example records are below:

C0000005|(131)I-Macroaggregated Albumin
C0000039|1,2-Dipalmitoylphosphatidylcholine
C0000052|1,4-alpha-Glucan Branching Enzyme
C0000074|1-Alkyl-2-Acylphosphatidates
C0000084|1-Carboxyglutamic Acid
C0000096|1-Methyl-3-isobutylxanthine
C0000097|1-Methyl-4-phenyl-1,2,3,6-tetrahydropyridine
C0000098|1-Methyl-4-phenylpyridinium
C0000102|1-Naphthylamine
C0000103|1-Naphthylisothiocyanate

cuist

This table maps concept identifiers to semantic types.

Each record contains the UMLS concept identifier and a semantic type abbreviation:

cui|semantic type abbreviation

Usually, there are several records for each UMLS concept:

C0000005|aapp
C0000005|irda
C0000005|phsu
C0000039|orch
C0000039|phsu
C0000052|aapp
C0000052|enzy
C0000074|orch
C0000084|aapp
C0000084|bacs

meshtcrelaxed

This table maps terms to MeSH treecodes for terms that are present in MeSH. Each record contains the term and its associated treecode. Term that do not have a treecode are mapped to "x.x.x.x".

term|treecode

Sample records are below:

(131)I-Macroaggregated Albumin|x.x.x.x
(131)I-MAA|x.x.x.x
1,2-Dipalmitoylphosphatidylcholine|D10.570.755.375.760.400.800.224
1,2 Dipalmitoylphosphatidylcholine|D10.570.755.375.760.400.800.224
1,2-Dihexadecyl-sn-Glycerophosphocholine|D10.570.755.375.760.400.800.224
1,2 Dihexadecyl sn Glycerophosphocholine|D10.570.755.375.760.400.800.224
1,2-Dipalmitoyl-Glycerophosphocholine|D10.570.755.375.760.400.800.224
1,2 Dipalmitoyl Glycerophosphocholine|D10.570.755.375.760.400.800.224

vars

This table maps a term to their associated lexical variants and lexical variants to their associated terms.

original term|source category|target term|target category|flow history|mutate info

Sample records are below:

A10|all|A10|noun|G|128|1|n|0|3|
A10|noun|A-10|noun|G|128|1|n+s|0|3|
A11|all|A11|noun|G|128|1|n|0|3|
A1ATD|all|A1ATD|noun|G|128|1|n|0|3|
A1ATD|noun|A1ATD's|noun|G|128|1|n+i|1|3|
A1ATD|noun|A1ATDs|noun|G|128|1|n+i|1|3|
A1ATD|noun|alpha 1 antitrypsin deficiencies|noun|G|128|1|n+a+s+i|3|1|
A1ATD|noun|alpha 1 antitrypsin deficiency|noun|G|128|1|n+a+s|2|1|
A1ATD|noun|alpha 1-antitrypsin deficiencies|noun|G|128|1|n+a+s+i|3|1|
A1ATD|noun|alpha 1-antitrypsin deficiency|noun|G|128|1|n+a+s|2|1|

Dictionary and Postings File Organization

The dictionary is constructed as several partitions, each partition consists of three files for each set of terms of a specific length from a specific column of the originating table: a dictionary characteristics file, a term-dictionary file, and a postings extents file. These files refer to a global postings file used by all of the partitions. Each term-dictionary contains records of term with the same length with the number of postings for the term and the address

Within the postings extents file (indexname-column-termlength-partition-offsets) each record contains address of posting with the length of the posting.

Each partition has four files, the file cuisourceinfo-3-30-term-dictionary-stats.txt contains term and record lengths along with the number of records in the partition:

termlength|30
reclength|46
datalength|16
recordnum|76361

The file cuisourceinfo-3-30-term-dictionary is a binary file of records, all the same length with each record containing a term, the number of postings for that term and address of the posting extents for that term:

|            term              | # of postings | address |
+------------------------------+---------------+---------+
|Dipalmitoylphosphatidylcholine|       4       | FFF4556 |

The file cuisourceinfo-3-30-partition-offsets is a binary file that contains the start and end offsets (the extent) of each posting:

| address | start |  len  |
+---------+-------+-------+
| FFF4556 |   58  |   57  |
|   ...   |  176  |   66  |
|   ...   |  279  |   59  |
|   ...   |    0  |   58  |

Defined in irutils.MultiKeyIndex.Extent (irutils/MultiKeyIndex.java):

class Extent {
  /** address of posting */
  long start;
  /** length of posting */
  long end;
}

The postings for all of the partitions reside in the file postings.

address | data 
--------+-------------------------------------------------------------------
      0 | C0000039|S0033298|4|Dipalmitoylphosphatidylcholine|SNMI|PT
     58 | C0000039|S0033298|7|Dipalmitoylphosphatidylcholine|LNC|CN
    176 | C0000039|S0033298|6|Dipalmitoylphosphatidylcholine|SNOMEDCT_US|OAP
    279 | C0000039|S0033298|5|Dipalmitoylphosphatidylcholine|NDFRT|SY

The postings file is shared among all of the partitions.

Dictionary and Postings File Construction

for each partition:

  • write temporary partition tables grouped by table column and term length.
  • write postings storing extent(offset,length) keyed by checksum of postings content and storing checksum list keyed by term.
  • write dictionary and extents after postings are written.
  • term to checksum map is discarded after partition is saved, checksum to extent map is preserved.

Required files

  1. Java class and supporting jars
  2. POM and ANT files
  3. Open NLP sentence model files
  4. Inverted Multi-key index files
  5. Local libraries which are not available through Maven.

Organization

 metamaplite -+- src 
              +- config
              +- data -+- models
              |        +- ivf -+- strict -+- indices -+- cuisourceinfo
              |                                       +- cuiconcept
			  |                                       +- cuist
              +- logs
              +- scripts
              +- target

Possible implementations

IRUtils multi-key index

Multi-key Inverted File index is now the default implementation.

current tables:

cuiconcept
cuisourceinfo
cuist

Generating Tables and Indexes

The CreateIndexes class generates tables cuiconcept, cuisourceinfo, cuist, meshtcrelaxed, and vars from MRCONSO.RRF, MRSTY.RRF, and MRSTY and then produces corresponding indexes for tables.

Usage:

 java -cp target/metamaplite-1.0-SNAPSHOT.jar \
  gov.nih.nlm.nls.metamap.dfbuilder.CreateIndexes <mrconsofile> <mrstyfile> <mrsatfile> <ivfdir>

The Previous Method for Generating Tables Only

cuiconcept

format:

cui|preferred name

example records:

C0000005|(131)I-Macroaggregated Albumin
C0000039|1,2-Dipalmitoylphosphatidylcholine
C0000052|1,4-alpha-Glucan Branching Enzyme
C0000074|1-Alkyl-2-Acylphosphatidates
C0000084|1-Carboxyglutamic Acid

program:

MRCONSO.RFF -> ExtractMrconsoPreferredNames -> cuiconcept

example of running program:

$ java -cp $CLASSPATH \
gov.nih.nlm.nls.metamap.dfbuilder.ExtractMrconsoPreferredNames \
data/ivf/2015AA/mrconso.eng \
data/ivf/2015AA/strict/tables/cuiconcept.txt

cuisourceinfo

format:

cui|sui|seqno|string|src|tty

example records:

C0000005|S0007492|1|(131)I-Macroaggregated Albumin|MSH|PEN
C0000005|S0007491|2|(131)I-MAA|MSH|EN
C0000039|S0007564|1|1,2-Dipalmitoylphosphatidylcholine|MSH|MH
C0000039|S0007564|2|1,2-Dipalmitoylphosphatidylcholine|NDFRT|PT
C0000039|S0007564|3|1,2-Dipalmitoylphosphatidylcholine|MTH|PN

program: gov.nih.nlm.nls.metamap.dfbuilder.ExtractMrconsoSources

MRCONSO.RRF -> ExtractMrconsoSources -> cuisourceinfo

example of running program:

$ java -cp $CLASSPATH \
gov.nih.nlm.nls.mmtx.dfbuilder.ExtractMrconsoSources \
data/ivf/2015AA/mrconso.eng \
data/ivf/2015AA/cuisourceinfo.txt

cuist

format:

cui|semantic type abbreviation

example records:

C0000005|aapp
C0000005|irda
C0000005|phsu
C0000039|orch
C0000039|phsu

program: gov.nih.nlm.nls.metamap.dfbuilder.ExtractMrstySemanticTypes

MRSTY.RRF -> ExtractMrstySemanticTypes -> cuist

example of running program:

$ java -cp $CLASSPATH \
gov.nih.nlm.nls.metamap.dfbuilder.ExtractMrstySemanticTypes \
data/ivf/2015AA/mrsty.rrf \
data/ivf/2015AA/strict/tables/cuist.txt

meshtcrelaxed

(131)I-Macroaggregated Albumin|x.x.x.x (131)I-MAA|x.x.x.x 1,2-Dipalmitoylphosphatidylcholine|D10.570.755.375.760.400.800.224 1,2 Dipalmitoylphosphatidylcholine|D10.570.755.375.760.400.800.224 1,2-Dihexadecyl-sn-Glycerophosphocholine|D10.570.755.375.760.400.800.224 1,2 Dihexadecyl sn Glycerophosphocholine|D10.570.755.375.760.400.800.224 1,2-Dipalmitoyl-Glycerophosphocholine|D10.570.755.375.760.400.800.224 1,2 Dipalmitoyl Glycerophosphocholine|D10.570.755.375.760.400.800.224 Dipalmitoylphosphatidylcholine|D10.570.755.375.760.400.800.224 Dipalmitoylglycerophosphocholine|D10.570.755.375.760.400.800.224

program: gov.nih.nlm.nls.metamap.dfbuilder.ExtractTreecodes

MRCONSO.RRF + MRSAT -> ExtractTreecodes -> mesh_tc_relaxed.txt

example of running program:

$ java -cp $CLASSPATH \
gov.nih.nlm.nls.metamap.dfbuilder.ExtractTreecodes \
data/ivf/2015AA/mrconso.eng
data/ivf/2015AA/mrsat.rrf \
data/ivf/2015AA/strict/mesh_tc_relaxed.txt

vars

A10|all|A10|noun|G|128|1|n|0|3| A10|noun|A-10|noun|G|128|1|n+s|0|3| A11|all|A11|noun|G|128|1|n|0|3| A1ATD|all|A1ATD|noun|G|128|1|n|0|3| A1ATD|noun|A1ATD's|noun|G|128|1|n+i|1|3| A1ATD|noun|A1ATDs|noun|G|128|1|n+i|1|3| A1ATD|noun|alpha 1 antitrypsin deficiencies|noun|G|128|1|n+a+s+i|3|1| A1ATD|noun|alpha 1 antitrypsin deficiency|noun|G|128|1|n+a+s|2|1| A1ATD|noun|alpha 1-antitrypsin deficiencies|noun|G|128|1|n+a+s+i|3|1| A1ATD|noun|alpha 1-antitrypsin deficiency|noun|G|128|1|n+a+s|2|1|

program: gov.nih.nlm.nls.metamap.dfbuilder.GenerateVariants

MRCONSO.RRF + LVG2020 -> GenerateVariants -> vars.txt

NOTE: The GenerateVariants program requires the Lexical Variant Generator library and database to work. LVG is available from the Lexical Tools Page: https://lexsrv3.nlm.nih.gov/LexSysGroup/Projects/lvg/current/web/index.html

example of running program:

$ export LVG_DIR=/opt/lvg2020
$ java -cp $CLASSPATH \
gov.nih.nlm.nls.metamap.dfbuilder.ExtractTreecodes \
data/ivf/2015AA/mrconso.eng
data/ivf/2015AA/strict/vars.txt

Trie based

Trie based NER using a lexically enhanced trie dictionary.

Lucene index based

Tables represented by Lucene indices.

Optimizations

Cache references to entities in Map to avoid unnecessary lookups, this is in current implementation.

Input Formats

Current

  • FreeText
  • Single Line Input (SLI)
  • Single Line Delimited Input (SLDI)
  • Medline
  • PubMedXML
  • PubMedRIS
  • ChemDNER
  • ChemDNERSLDI
  • NCBICorpus

Future

  • JSON
  • EDN

output formats

  • MMI
  • BRAT (Stanoff Format)
  • JSON

Pipeline

Pipeline Definition

YAML versus Java Properties versus XML

  • XMLis too verbose
  • Java Properties are name=value pairs which are too restrictive.
  • JSON is more versatile but not human readable.
  • YAML

Example YAML configuration file

tokenizesentence:
   description: tokenize sentence
   class: gov.nih.nlm.nls.metamap.lite.SentenceAnnotator.tokenizeSentence
   input parameters: bioc.BioCSentence
   return output: bioc.BioCSentence

genentityset:
   description: gind entities
   class: gov.nih.nlm.nls.metamap.lite.EntityLookup1.addEntities
   input parameters: bioc.BioCSentence
   return output: bioc.BioCSentence

applycontext:
   description: hedging
   class: gov.nih.nlm.nls.metamap.lite.context.ContextWrapper
   input parameters: bioc.BioCSentence
   return output: bioc.BioCSentence

displayentityset:
   description: display entities
   class: gov.nih.nlm.nls.metamap.lite.EntityLookup1.displayEntitySet
   input parameters: bioc.BioCSentence
   return output: void

simple.sentence.pipeline:
   - tokenizesentence
   - genentityset
   - displayentityset

Pipeline Protocols

object based

Pipeline Wrapper Modules

Plugins

Should use OSGi or JPF (Java Plugin Framework)

chunkers

full parsers

dictionary lookup

normalization

Filtering for Other

The term "Other" is ignored in MetaMapLite if it's the first token in candidate token list. (It is a stop_phrase in MetaMap)

Java Interfaces

Current

ChunkerMethod EntityLookup NegationDetector VariantLookup Phrase

Other Future

Tokenization SentenceExtractor (SentenceBreaker?) FullParser

MRCONSO file

/nfsvol/nls/specialist/module/metawordindex/data.Base.2014AA/mrconso.suppressed

5365358

Evaluation

Similarity Measures

Configuration

Sample properties file

# Lucene indexes
metamaplite.cuiconceptindex: /nfsvol/nlsaux15/lucenedb/strict/cuiconcept
metamaplite.firstwordsofonewideindex: /nfsvol/nlsaux15/lucenedb/strict/first_words_of_one_WIDE
metamaplite.cuisourceinfoindex: /nfsvol/nlsaux15/lucenedb/strict/cui_sourceinfo
metamaplite.cuisemantictypeindex: /nfsvol/nlsaux15/lucenedb/strict/cui_st
metamaplite.varsindex: /nfsvol/nlsaux15/lucenedb/strict/vars
#
# OpenNLP model files
opennlp.en-sent.bin.path:/usr/local/pub/nlp/opennlp/models/en-sent.bin
opennlp.en-token.bin.path:/usr/local/pub/nlp/opennlp/models/en-token.bin
opennlp.en-pos-maxent.bin.path:/usr/local/pub/nlp/opennlp/models/en-pos-maxent.bin

Entities

Essentially a span within body of text.

Annotations are associated with the span.

typical annotations Concept Id with associated semantic types and source references

Generating datafiles

Generating cuisourceinfoindex (cui_sourceinfo)

Use extract_mrconso_sources.perl:

http://indlx1.nlm.nih.gov:8000/cgi-bin/cgit.cgi/public_mm/tree/bin/extract_mrconso_sources.perl

Generating cuisemantictypeindex (cui_st)

Timing

NCBI Disease Training Corpus (592 documents)

MetaMapLite

indlx1

First run (indlx1):

real	29m12.477s
user	39m51.089s
sys		1m7.615s

Second run (indlx1):

real	26m10.628s
user	39m24.213s
sys		0m37.145s

ii-server6

First run (ii-server6):

real	44m47.399s
user	52m13.420s
sys		1m23.783s

Secord run (ii-server6):

real	39m53.229s
user	47m53.044s
sys		2m38.590s

MetaMap

First run (ii-server6)

start time: Wed Mar 11 08:33:52 EDT 2015
end time: Wed Mar 11 09:20:22 EDT 2015

47 minutes

Second run (ii-server6):

start time: Wed Mar 11 09:23:50 EDT 2015
end time: Wed Mar 11 10:05:18 EDT 2015
2149.859u 230.497s 41:28.33 95.6%	0+0k 0+7760io 0pf+0w

42 minutes

Segmentation

myProperties.setProperty("metamaplite.segmentation.method", methodname);

1 Times for songs collection using various segmentation methods

  program        | method     | time1           | time2
 ---------------------------------------------------------------
 test.MMLtest    | SENTENCES  | real	0m50.637s  | real	0m49.971
                 |            | user	1m12.791s  | user	1m13.957s
                 |            | sys 	0m2.068s   | sys	0m2.163s
                 |            |				       |
                 | BLANKLINES | real	0m52.322s  | real	0m49.817s
                 |            | user	1m15.008s  | user	1m16.559s
                 |            | sys 	0m2.859s   | sys	0m4.176s
                 |            |				       |
                 | LINES      | real	0m51.343s  | real	0m48.730s
                 |            | user	1m19.078s  | user	1m8.178s
                 |            | sys 	0m2.200s   | sys	0m2.313s
                 |            |				       |
 test.MMLtestAlt | SENTENCES  | real	0m58.387s  | real	0m55.934s
                 |            | user	1m22.942s  | user	1m18.993s
                 |            | sys 	0m2.314s   | sys	0m3.644s
                 |            |		      		   |
                 | BLANKLINES | real	0m59.280s  | real	0m56.476s
                 |            | user	1m27.496s  | user	1m20.455s
                 |            | sys	    0m4.482s   | sys	0m2.460s
                 |            |			    	   |
                 | LINES      | real	0m58.441s  | real	0m56.544s
                 |            | user	1m23.249s  | user	1m17.672s
                 |            | sys	    0m3.539s   | sys	0m3.341s

Scoring

MetaMap Scoring Components

Described in paper "MetaMap Evaluation", Alan R. Aronson, https://ii.nlm.nih.gov/Publications/Papers/mm.evaluation.pdf

score = 1000*((centrality + variation + coverage + cohesiveness)/6)

Variation

An estimation of how much the variant in the Metathesaurus string differ from the corresponding words in the phrase.

(defn lexical-variation-sum
   "Sum of lexical variation of phrase to match the metastring"
   [phrase metastring]
   (map (fn [phrase-token metastring-token]
           (lexical-variation phrase-token metastring-token))
        (tokenize phrase) (tokenize metastring)))

Coverage

An indication of how much of the phrase string and the Metathesaurus string are involved in the match.

Cohesiveness

Involvement

MetaMapLite Scoring Components

Currently based on MetaMap's Scoring Components

score = 1000*((centrality + variation + (2.0*(coverage + cohesiveness)))/6.0)

MMI Ranking

Describe in paper "MMI Ranking Function", Alan R. Aronson, https://ii.nlm.nih.gov/Publications/Papers/ranking.pdf

Maven Issues

These are related to the use of Antonio Jimeno Yepes' WSD library.

package                            |  new repository location
-----------------------------------+-----------------------------------------------------
monq:monq:jar:1.1.1                | bioinformatics.ua.pt/maven/content/groups/public
org.w3c.xml:thirdparty:jar:1.2.0   | ?
com.sleepycat.db:db:jar:4.1        | ?

Feature Requests

  1. Support entity lookup using two separate dataset (UMLS and custom, etc.)

Known bugs

  1. Semantic Type filter does not always filter out all excluded types
  2. Abbreviation detector doesn't always work