Generating SD-Rules in the SPECIALIST Lexical Tools

Optimization for Suffix Derivation Rule Set

Chris J. Lu

1,2

, Destinee Tormey

, Lynn McCreedy

and Allen C. Browne

National Library of Medicine, Bethesda, MD, U.S.A.

Medical Science & Computing, LLC, Rockville, MD, U.S.A.

Keywords: Derivations, Suffix Derivations, SD-Rules, Natural Language Processing, the SPECIALIST Lexical Tools.

Abstract: Suffix derivations (SDs) are used with query expansion in concept mapping as an effective Natural Language

Processing (NLP) technique to improve recall without sacrificing precision. A systematic approach was

proposed to generate derivations in the SPECIALIST Lexical Tools in which SD candidate rules were used

to retrieve SD-pairs from the SPECIALIST Lexicon (Lu et al., 2012). Good SD candidate rules are gathered

as SD-Rules in Lexical Tools for generating SDs that are not known to the Lexicon. This paper describes a

methodology to select an optimized SD-Rule set that meets our requirement of 95% system precision with

best system performance from SD candidate rules. The results of the latest three releases of Lexical Tools

show: 1) system precision and recall of selected SD-Rules are above 95%. 2) a consistency between a

computational linguistic approach and traditional linguistic knowledge for selecting the best Parent-Child

rules. 3) a consistent approach yielding similar SD-Rule sets and system performance. Ultimately, it results

in better precision and recall for NLP applications using Lexical Tools derivational related flow components.

1 INTRODUCTION

NLP in medicine is used to retrieve information and

analyze linguistic patterns from electronic medical

records (EMRs), published biomedical research,

clinical trial reports, and other sources. Due to the

great deal of lexical variation in natural language,

managing this variability is an important key to

successful Medical Language Processing (Pacak et

al., 1980; Wolff, 1987). The National Library of

Medicine (NLM) distributes the NLP SPECIALIST

Lexicon and Lexical Tools as one of the Unified

Medical Language System (UMLS) Knowledge

Sources along with the Metathesaurus. These rich and

robust NLP resources have provided the

NLP/Medical Language Processing community with

an extensive NLP toolset since 1994 (McCray et al.,

1993). The SPECIALIST Lexical Tools is designed

to handle lexical variation, including derivational

variant generation.

Derivational variants are words related by a

derivational process, such as suffixation, pre-fixation,

and conversion, to create new words based on

existing words. They need not be synonymous. In

fact, derivation often entails thoroughgoing meaning

change. Derivations allow users to find closely

related terms that may differ in syntactic category, but

are nonetheless usefully related. They are used with

query expansion to increase recall for UMLS concept

mapping in MetaMap (Aronson and Lang, 2010) and

Sophia (Divita et al., 2014). For example, no CUI was

found by direct mapping if the source vocabulary is

“perforated ear drum”. By substituting the subterm

“perforated” for its derivational variant,

“perforation,” the UMLS Metathesaurus concept,

C0206504, is found. More information, such as its

preferred term (Tympanic Membrane Perforation)

and synonyms, can be retrieved for further NLP

analysis.

SD generation in the Lexical Tools derivational

flow component is based on a paradigm of SD-Facts

and SD-Rules (Lu et al., 2012). Over 100 SD

candidate rules have been collected by the Lexical

Systems Group (LSG). All terms in the Lexicon that

match a SD candidate rule are retrieved as SD

candidate pairs. About 30% of them are validated

automatically by a computer expert system while the

rest (~70%) are tagged by LSG linguists manually

(Lu et al., 2013). Valid SD-pairs are stored in the

Lexical Tools database to retrieve derivations that are

known to the Lexicon, that is, SD-Facts. The increase

of both the Lexicon and SD candidate rules

Lu, C., Tormey, D., McCreedy, L. and Browne, A.

Generating SD-Rules in the SPECIALIST Lexical Tools - Optimization for Sufﬁx Derivation Rule Set.

DOI: 10.5220/0005731303530358

In Proceedings of the 9th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2016) - Volume 5: HEALTHINF, pages 353-358

ISBN: 978-989-758-170-0

353

contributes to the growth of SD-Facts. The number of

SD-Facts has increased over 12.4 times from 4,110 to

50,814 since 2011 for better recall. The derivational

flow component’s algorithm is enhanced by this basis

on SD-Facts, while it was mainly based on SD-Rules

in 2011, for better precision. As a result, the Lexicon

and Lexical Tools provide one of the most

comprehensive and robust resources of derivations

for the NLP community.

If no derivations are found by SD-Facts, SD-

Rules are used to generate SDs in Lexical Tools. SD-

Rules are good rules selected from the SD candidate

rules and stored in a Trie mechanism (Aho et al.,

1983). Four heuristic algorithms are implemented in

Lexical Tools to eliminate nonrealistic derivational

variants and increase precision (Lu et al., 2012).

SD candidate rules can be derived by computer

algorithms from a non-derivational lexicon (Gaussier,

1999), based on synonym relations (Grabar and

Zweigenbaum, 1999), structured terminology

(Grabar and Zweigenbaum, 2000), etc. Derivations

from Lexical Tools were used as a gold standard for

comparison in the above studies. There are hundreds

of SD candidate rules. Due to limited resources, the

LSG evaluated the most common English suffixes for

SD candidate rules. Before these candidate rules can

become SD-Rules in the Lexical Tools, they undergo

evaluation to ensure they provide the required system

performance. This paper describes a methodology of:

1) finding the SD-Rules - the optimized subset from

collected SD candidate rules, and 2) finding the best

SD rule(s) in a Parent-Child (P-C) family.

2 OBJECTIVE AND APPROACH

Our goal is to find a set of SD-Rules for generating

derivational variants that are not unknown to the

Lexicon using Lexical Tools. This set of SD-Rules is

a subset of SD candidate rules selected by filtering out

bad rules. This subset must meet our objectives of

cumulative system precision above 95% and best

system performance (see section 2.3). In addition,

rules in a P-C family with the best system

performance are chosen for the optimal set. The

methodology is described below.

2.1 New SD Candidate Rules

The LSG has enhanced SD generation in the Lexicon

and Lexical Tools by adding new SD candidate rules

and Lexical records. There are 13 and 11 new SD

candidate rules introduced in the 2015 and 2016

releases, as shown in Tables 1 and 2. A SD rule

includes suffixes and parts of speech for a candidate

SD pair. For example, the suffix “es” in certain nouns

can be replaced with the suffix “ic” to create a related

adjective (adj). Thus, replacing “es” with “ic” in

“diabetes” creates “diabetic”, expressible as SD-pair

diabetes|noun|diabetic|adj. This SD rule (ES-15-03 in

Table 1) is coded in the following format in Lexical

Tools: es$|noun|ic$|adj, where “$” means the end of

the word. SD rules can be applied to generate SD-

pairs in both directions; “diabetic” generates

“diabetes” from this same rule. Invalid SD-pairs

(exceptions to this SD candidate rule) such as,

comes|noun|comic|adj, are also retrieved from

Lexicon. These candidate SD-pairs are validated by

computer programs and linguists during the

derivation generation process (Lu et al., 2012).

Table 1: New SD Candidate Rules in Lexicon, 2015.

ID New SD Candi. Rule Rank

CG-15-01 se$|verb|zation$|noun 2

CG-15-02 sation$|noun|ze$|verb 3

CG-15-03 ility$|noun|le$|adj 9

CG-15-04 $|adj|ally$|adv 15

CG-15-05 ce$|noun|t$|adj 18

CG-15-06 cy$|noun|t$|adj 19

CG-15-07 e$|verb|ion$|noun 20

CG-15-08 c$|adj|s$|noun 43

ES-15-01 e$|verb|ing$|noun 45

ES-15-02 al$|adj|us$|noun 61

ES-15-03 es$|noun|ic$|adj 67

ES-15-04 $|noun|ize$|verb 78

ES-15-05 es$|noun|ic$|noun 101

Table 2: New SD Candidate Rules in Lexicon, 2016.

ID New SD Candi. Rule Rank

CG-16-01

se$|verb|sis$|noun 27

CG-16-02

sia$|noun|tic$|adj 40

CG-16-03 on$|noun|ve$|adj 48

CG-16-04 e$|noun|ic$|adj 49

CG-16-05 $|adj|ism$|noun 51

CG-16-06 ation$|noun|ed$|adj 67

CG-16-07

$|noun|ship$|noun 70

CG-16-08 e$|adj|ion$|noun 88

CG-16-09 $|noun|age$|noun

ES-16-01 esis$|noun|ic$|adj

ES-16-02 al$|adj|ine$|noun 98

New rules suggested by NLP experts, including

Lexicon and Lexical Tools users and LSG linguists,

are marked as ES# (expert-suggested). In addition to

ES candidate rules, the LSG developed a system to

derive SD candidates from known SD-pairs

automatically. First, the LSG collected all known SD-

pairs from nominalizations (nomD) in the Lexicon

HEALTHINF 2016 - 9th International Conference on Health Informatics

354

and the original SD-pairs (orgD) that existed in the

deprecated Lexical Tools release of 2011. These

orgDs are not necessary in the current SD-Facts,

because some of them were removed after the new

derivation generation system was deployed in 2012

(Lu et al., 2012). Next, computer programs found SD

candidate rules by removing all common leading

characters (the word stem) of the SD-pairs. For

example, nomD, celebrate|verb|celebration|noun, is

collected from the Lexical record

celebrate|E0015729, as shown in Figure 1. The SD

rule, e$|verb|ion$|noun, is derived by removing the

stem “celebrat” from “celebrate” and “celebration”.

In the 2015 release, 1,017 SD rules were derived from

23,384 nomDs and sorted by frequency. The top 17

rules with the highest frequency cover above 80% of

all nomDs. Six new rules (non-existent in the

previous SD rule set) were selected. Similarly, 1,421

SD rules were derived from 4,110 orgDs. Two new

rules from the top 6 with the highest frequency were

selected. These 8 (6 from nomDs and 2 from orgDs)

new computer-generated rules were added to the 2015

SD candidate rules, marked as CG# in Table 1. For

Figure 1: Lexical Record of “celebrate”.

the 2016 release, 9 computer-generated SD rules (4

from nomDs and 5 from orgDs) from the following

top ranking have similarly been selected, as shown in

Table 2.

2.2 Establish the Baseline

We will use 2015 as an example to illustrate these

processes. First, the 13 new SD rules (5 expert-

suggested and 8 computer-generated) are added to the

107 SD candidate rules of the prior release (2014).

Second, these rules are normalized and ordered

alphabetically. The result is 120 unique SD candidate

rules in a standard format. Third, all rules related in a

P-C family in the new SD candidate set are identified.

Table 3 identifies 14 Parent-Rules (P#) and 19

associated Child-Rules (C#) for 2015 release. Fourth,

all 19 Child-Rules are removed to avoid duplication,

because the best rules from all generations for all

these 14 P-C families will be chosen through the

optimization processes (see section 2.4). This set of

normalized SD candidate rules, composed of 101 (=

120 – 19) unique rules (with no Child-Rules), is used

as the baseline for further processes.

2.3 System Performance

Our goal is to choose a set of SD-Rules that perform

as well as SD-Facts in Lexical Tools. Retrieved

candidate SD-pairs (retrieved instances) and valid

SD-pairs in SD-Facts (relevant instances) are used to

calculate precision and recall. First, we retrieve all

Table 3: Parent-Child SD Rule Families, 2015.

PID Parent-Rule CID Child-Rule

P1 $|adj|ally$|adv

C1 ic$|adj|ically$|adv

P2 $|adj|ity$|noun

C2 ic$|adj|icity$|noun

P3 $|noun|al$|adj

C3 ic$|noun|ical$|adj

P4 $|verb|ion$|noun

C4 ss$|verb|ssion$|noun

P5 a$|noun|an$|adj

C5 ia$|noun|ian$|adj

P6 a$|noun|an$|noun

C6 ia$|noun|ian$|noun

P7 a$|noun|ar$|adj

C7 ula$|noun|ular$|adj

P8 ation$|noun|e$|verb

C8 ization$|noun|ize$|verb

P9 c$|adj|s$|noun

C9 ic$|adj|is$|noun

P10 ce$|noun|t$|adj

C10

C11

C12

ance$|noun|ant$|adj

iance$|noun|iant$|adj

ence$|noun|ent$|adj

P11 cy$|noun|t$|adj

C13

C14

ency$|noun|ent$|adj

iency$|noun|ient$|adj

P12 e$|verb|ion$|noun

C15

C16

ate$|verb|ation$|noun

se$|verb|sion$|noun

P13 ility$|noun|le$|adj

C17

ability$|noun|able$|adj

P14 sis$|noun|tic$|adj

C18

C19

esis$|noun|etic$|adj

osis$|noun|otic$|adj

Generating SD-Rules in the SPECIALIST Lexical Tools - Optimization for Sufﬁx Derivation Rule Set

355

SD-pairs that match SD rules in the evaluation set.

Second, we sort these unique SD candidate rules by

rank. That is, by the descending order of precision (=

relevant, retrieved No./retrieved No.), frequency

(retrieved No.), and then the alphabetical order of SD

rules. Third, we calculate the cumulative system

precision (SP) and system recall (SR = relevant,

retrieved No./relevant No.) for all rules. Fourth, we

find the cutoff rule with the least SP above 95%. All

SD candidate rules with a higher rank than the cutoff

rule are selected as good rules for the SD-Rule set.

The system performance of the SD-Rule set is

measured by both (sum of) cumulative SP and SR at

the cutoff rule.

2.4 Parent-Child Rules Optimization

In Table 3, C9, ic$|adj|is$|noun, is the first generation

Child-Rule of P9 by adding one character, ‘i’, to P9,

c$|adj|s$|noun. P9 is a root Parent-Rule because it

does not have a possible Parent-Rule (no common

leading characters). Note that all computer-generated

SD candidates are root Parent-Rules. Parent-Rules

might have multiple Child-Rules, such as Parent-Rule

P14, which has two Child-Rules, C18 and C19, by

adding “e” and “o”. Further, multiple generations of

Child-Rules could be involved. For example, the root

Parent-Rule, P10, has 2 second generation Child-

Rules C10 and C12 (by adding the two characters

“an” and “en”, respectively) and 1 third generation

Child-Rule C11 (by adding the three characters

“ian”). In such a case, C10 is both a Child-Rule (of

P10) and a Parent-Rule (of C11).

The root Parent-Rule is used to find the best

rule(s) in its P-C family. Theoretically, a Parent-Rule

should have better recall than its Child-Rule because

SD-pairs derived from a Child-Rule are a subset of

those derived from the associated Parent-Rule. A

Parent-Rule, however, could have worse precision

than its Child-Rule because it could include more

invalid SD-pairs. Local precision (valid SD-pairs

over retrieved SD candidate pairs from the rule and

Lexicon) and frequency (instances of the retrieved SD

candidate pairs from the rule and Lexicon) are two

criteria we used for finding the qualified Child-Rules

in a P-C family for further evaluation. A sophisticated

heuristic algorithm has been developed. First, the root

Parent-Rule of the P-C family replaces the original

Parent-Rules (P1 – P14). For example, the root

Parent-Rule, $|noun|n$|adj, replaces P5 (by removing

character ‘a’ from P5). Second, all possible Child-

Rules from all generations are derived from this root

Parent-Rule. Third, qualified Child-Rules are

selected from all generations if they 1) have higher

precision than their Parent-Rules; and 2) meet

frequency criteria (> 25% of their root Parent-Rule

and > 40% of their immediate Parent-Rule). Finally,

the system performance of all qualified P-C rules of a

P-C family are calculated. The one with the best

system performance is chosen. Linguistic knowledge

guides the P-C choice when the system performance

of different generations of P-C rules are the same.

Finally, a Parent-Rule is chosen over a Child-Rule for

better recall when the above conditions are equal.

This process is repeated for finding the best P-C rules

from all 14 P-C families to obtain the optimized SD-

Rule set.

3 RESULTS AND CONCLUSION

Table 4 summarizes the results of generating the

optimized SD-Rule set from 2014-2016. The system

performance (~1.90), SP (~95%) and SR (~95%) of

the SD-Rule sets stay consistent as the number of

candidate SD rules and retrieved SD-pairs increases

about 10% each year. Figure 2 shows a similar pattern

of the system precision-recall vs. SD candidate rules

in the optimized rule set for 2014-2016. The

consistency over three releases indicates this

approach is stable. The intersections of precision and

recall curves are at the cutoff SD rule with precision

and recall both around 95%. This equilibrium

between precision and recall confirms that setting the

precision at 95% as the minimum cutoff requirement

is the right choice for representing system

performance.

The 2014 and 2015 releases have the same cutoff

SD rule, ar$|adj|e$|noun. This rule drops to rank 83

(just below the cutoff rank of 82) in 2016. Rules

above the cutoff SD-Rule are considered good rules

and vice versa. From a linguistic standpoint, a good

SD rule should remain good. We observed the good

rules remain above the cutoff SD rule over the past

three years. The complete log of optimization

processes and SD-Rule sets for the 2014, 2015 and

2016 releases are available at the SPECIALIST

Lexical Tools web site (National Library of

Medicine, Lexical Tools 2014, 2015 and 2016).

System performance is used to choose the best

rules from all generations in those P-C families with

Parent-Rules and Child-Rules coexisting in the

candidate set. As we observed, the best rules from a

P-C family chosen by this computational approach

concurs with the conventional linguistic standpoint.

For example, the 3

generation Child-Rule,

genesis$|noun|genic$|adj, has a higher system

performance than its root-parent rule, ES-16-01,

HEALTHINF 2016 - 9th International Conference on Health Informatics

356

Table 4: Summary of the Optimized SD-Rule Set, 2014-2016.

Release 2014 2015 2016

No. of SD Candidate Rules 107 120 132

No. of Baseline Rules 96 101 111

No. of Good Rules 73 76 82

No. of Retrieved SD-Pairs 48,579 53,885 58,391

No. of Relevant SD-Pairs 42,552 46,950 50,814

System Precision (SP) 95.30% 95.22% 95.00%

System Recall (SR) 95.01% 95.70% 95.26%

System Performance 1.9031 1.9093 1.9026

Cutoff Rule ar$|adj|e$|noun ar$|adj|e$|noun $|noun|ist$|noun

Figure 2: System Precision-Recall vs. SD Candidate Rules in the Optimized Set for Lexical Tools Releases 2014-2016.

esis$|noun|ic$|adj. This result concurs with the

linguistic standpoint (Dorland’s, 2003). In some

cases, Parent-Rules and Child-Rules have the same

system performance. For example, nce$|noun|nt$|adj

and its Child-Rules (ance$|noun|ant$|adj and

ence$|noun|ent$|adj) have the same system

performance (highest in this P-C family).

Linguistically, these two Child-Rules have the same

definition and Latin suffix source. In such cases, this

model chooses the Parent-Rule over the Child-Rule to

keep better coverage while remaining linguistically

valid

Derivational growth in the Lexicon is based on the

increase of the Lexicon as well as SD candidate rules.

Expert-suggested and computer-generated rules have

been introduced into the system. Eight and nine new

computer-generated SD candidate rules have been

added to the 2015 and 2016 releases, respectively.

100% (8/8) and 78% (7/9) of them are deemed good

rules in the SD-Rule set of the Lexical Tools’ 2015

and 2016 releases. We plan to add more computer-

generated rules for better coverage in future releases.

This methodology is valuable for evaluating future

SD candidate rules and is crucial when using SD-

Rules in Lexical Tools.

Finding the optimized set from different SD

candidate rule sets with different possible P-C rules,

makes this tedious process even more complicated.

For example, 11 of 13 new SD candidate rules are

evaluated as good rules in the 2015 release. As shown

in Table 4, the total number of candidate SD rules in

2015 is 120 (= 107 + 13). Only 5 of the 11 good rules

are not parent-child related rules. Four of them have

a 1-to-1 P-C relationship. For example, rule CG-15-

03 in Table 1 is the parent rule of C-17 in Table 3

from 2014. These 4 rules replace old rules and do not

affect the number of good rules. The other 2 good

rules have 1-to-2 P-C relationships and thus reduce

the number of good rules by one for each rule. For

example, rule CG-15-07 in Table 1 is the parent rule

of C15 and C16 in Table 3 from 2014. Thus, the total

number of good rules in 2015 is 76 (= 73 + 1x5 + 0x4

– 1x2) instead of 84 (= 73 + 11). Similarly, the

number of candidate rules, baseline rules and good

rules in 2016 can be derived by considering P-C

relationship.

This approach is a generic method to find the

optimized set from a candidate set to reach the best

system performance. It provides a maintainable and

scalable system for generating SD-Rules in Lexical

Tools to suggest derivations unknown to the Lexicon.

Ultimately, it produces better precision and recall for

NLP applications that use Lexical Tools derivational

flow components. As the Lexical Tools’ analysis of

English suffix relationships thus becomes ever

broader and deeper, so is its usefulness to NLP

strengthened. In the future, we plan to assess the

impact of our work by examining whether NLP

applications show improvement using derivational

features. Lexical Tools is distributed annually by

Generating SD-Rules in the SPECIALIST Lexical Tools - Optimization for Sufﬁx Derivation Rule Set

357

NLM via an open source license agreement.

ACKNOWLEDGEMENTS

This research was supported by the Intramural

Research Program of the NIH, National Library of

Medicine. The authors would like to thank Mr. Guy

Divita, Mr. Howard Lu, and Dr. Fiona M. Callaghan

for their valuable discussions.

REFERENCES

Aho, A.V., Ullman, J.D., and Hopcroft, J.E., 1983. Data

Structure and Algorithms. Addison Wesley, pages 163-

169.

Aronson, A.R. and Lang, F.M., 2010. An Overview of

MetaMap: Historical Perspective and Recent

Advances. JAMIA, Vol. 17, pages 229-236.

Divita, G., Zeng, Q.T., Gundlapalli, A.V., Duvall, S.,

Nebeker, J., and Samore, M.H., 2014. Sophia: An

Expedient UMLS Concept Extraction Annotator. In

proceeding of AMIA 2014 Annual Symposium, pages

467-476, Washington, DC, USA, November 15-19.

Dorland, W.A., 2003. Dorland’s Illustrated Medical

Dictionary, 30

edition, W. B. Saunders Company.

Philadelphia, Pa, page 763.

Fung, K.W., McDonald, C., and Srinivasan, S., 2010. The

UMLS-CORE Project: A Study of the Problem List

Terminologies Used in Large Healthcare Institutions.

JAMIA, Vol. 17, pages 675-680.

Gaussier, E., 1999. Unsupervised learning of derivational

morphology from inflectional lexicons. In: Kehler A

and Stolcke A, eds, ACL workshop on Unsupervised

Methods in Natural Language Learning, College Park,

MD. June.

Grabar, N. and Zweigenbaum, P., 2000. A General Method

for Sifting Linguistic Knowledge from Structured

Terminologies. In proceeding of AMIA 2000 Annual

Symposium, page. 310-314, Los Angeles, CA, USA,

November 4-8.

Grabar, N. and Zweigenbaum, P., 1999. Language

Independent Automatic Acquisition of Morphological

Knowledge from Synonym Pairs. In Proceeding of

AMIA 1999 Annual Symposium, page. 77-81,

Washington, DC, USA, November 6-10.

Lu, C.J., McCreedy, L., Tormey, D., and Browne, A.C.,

2012. A Systematic Approach for Automatically

Generating Derivational Variants in Lexical Tools

Based on the SPECIALIST Lexicon. IEEE IT

Professional Magazine, May/June, pages 36-42.

Lu, C.J., Tormey, D., McCreedy, L., and Browne, A.C.,

2013. Implementing Comprehensive Derivational

Features in Lexical Tools Using a Systematical

Approach. In proceeding of AMIA 2013 Annual

Symposium, Wash., DC, USA, Nov. 16-20, page 904.

McCray, A.T., Aronson, A.R., Browne, A.C., Rindflesch,

T.C., Razi, A., and Srinivasan, S., 1993. UMLS

Knowledge for Biomedical Language Processing. Bull.

Medical Library Assoc., vol. 81, no. 2, page 184-194.

National Library of Medicine, Lexical Tools 2014,

Optimizing 2014 SD-Rule Set – Add SD-Rules from

Other Suggestions. Available from: <http://lsg3.

nlm.nih.gov/LexSysGroup/Projects/lvg/2014/docs/desi

gnDoc/UDF/derivations/SD-Rules-Opti/ex-add-

suggest.html>. (24 October 2013).

National Library of Medicine, Lexical Tools 2015,

Optimizing 2015 SD-Rule Set - Optimum Log.

Available from: <http://lsg3.nlm.nih.gov/LexSysGroup

/Projects/lvg/2015/docs/designDoc/UDF/derivations/S

D-Rules-Opti/Ex-2015/optiLog.html>. (9 September

2014).

National Library of Medicine, Lexical Tools 2016,

Optimizing 2016 SD-Rule Set - Optimum Log.

Available from: <http://lsg3.nlm.nih.gov/LexSysGroup

/Projects/lvg/2016/docs/designDoc/UDF/derivations/S

D-Rules-Opti/Ex-2016/optiLog.html>. (17 September

2015).

Pacak, M.G., Norton, L.M., and Dunham, G.S., 1980.

Morphosemantic Analysis of - ITIS Forms in Medical

Language. J. Methods of Information in Medicine, vol.

19, no. 2, page 99-105.

Wolff. S., Automatic Coding of Medical Vocabulary. 1987.

Medical Information Processing – Computer

Management of Narrative Data. Addison Wesley,

Reading Mass, page 145-162.

HEALTHINF 2016 - 9th International Conference on Health Informatics

358