=Paper=
{{Paper
|id=Vol-2050/odls-paper9
|storemode=property
|title=Lexical Ambiguity in SNOMED CT
|pdfUrl=https://ceur-ws.org/Vol-2050/ODLS_paper_9.pdf
|volume=Vol-2050
|authors=Stefan Schulz,Catalina Martínez-Costa,Jose Antonio Miñarro-Giménez
|dblpUrl=https://dblp.org/rec/conf/jowo/SchulzMM17
}}
==Lexical Ambiguity in SNOMED CT==
https://ceur-ws.org/Vol-2050/ODLS_paper_9.pdf
Lexical ambiguity in SNOMED CT
Stefan Schulz1, Catalina Martínez-Costa, Jose Antonio Miñarro-Giménez
Institute of Medical Informatics, Statistics and Documentation,
Medical University of Graz, Austria
Abstract. Terminology systems that represent the language as used in human
communication have to deal with the problem of lexical ambiguity; i.e. the same
natural language term is assigned to two or more codes. A scrutiny of the large
international terminology standard SNOMED CT focused on concepts that are
linked by the same term and exhibit problems especially when using the
terminology in a Natural Language Processing context. We found 8,338
ambiguous terms from about 700k terms in SNOMED CT and provide
recommendations in order to improve its quality by curators.
Keywords. SNOMED CT, Biomedical terminologies, Biomedical ontologies,
Lexical ambiguity
1. Introduction
Lexical ambiguity is the capacity of a term to have multiple meanings. Humans
constantly produce ambiguous utterances, due to our capacity of intuitively inferring
the right sense guided by linguistic context and domain knowledge. For machine
processing, the processing of ambiguities is a known problem, and WSD (word sense
disambiguation) is a classical Natural Language Processing (NLP) task.
Reference ontologies and terminologies tend to avoid ambiguities in their choice of
preferred terms or labels. They are ideally self-explaining (e.g. “Transplantation of
liver (procedure)”), but often far away from clinicians’ jargon (“Liver transplant” or
“LT”). However, as soon as the terminology is enriched by (quasi)synonyms or close-
to-user entry terms, the ambiguity problem arises (e.g. “LT” for “Leishmania tropica”,
“Low testosterone” and others).
Besides classical cases of lexical ambiguity (e.g. “bank” for riverside vs. financial
institution), a notorious source of lexical ambiguity in medical language is the overuse
of short forms [1]. Among several types of word shortenings, acronyms are the most
common ones like CT, ECG, CA and LT. They are single tokens derived from the
initial (mostly capitalised) components of words in a phrase and/or syllables in a word.
In particular, short acronyms are known to have dozens of expansions. AcronymFinder
[2], provides 61 expansions of the acronym “CT” (“Clinical Trial”, “Cognitive Therapy,
1
Corresponding author: Stefan Schulz, Medical University of Graz, Auenbruggerplatz 2/V, 8036 Graz
(Austria), E-mail: stefan.schulz@medunigraz.at.
“Computed Tomography”, “Connective Tissue”,…) in the context of science and
medicine.
In the following we will scrutinise the problem of lexical ambiguity in the large
ontology-based clinical terminology standard SNOMED CT [3] and will provide
naming recommendations for SNOMED CT curators in a similar way as existing
naming conventions [4] for ontologies in OBO Foundry [5]. SNOMED CT’s January
2017 international release counts about 300k concepts and 700k terms. The main
purpose of this – still preliminary – work is to identify and to discuss the relevance of
lexical ambiguity in SNOMED CT.
Ambiguous terms in a terminology are those that map to more than one code.
Lexical ambiguity matters in terminologies particularly when they are used as
dictionaries in natural language processing systems, because the choice of the right
code becomes a matter of chance, unless contextual information is used. SNOMED CT
distinguishes between fully specified names (FSNs) and other terms, so-called
synonyms. All FSNs end with a hierarchy tag (i.e. semantic type) in parentheses
identifying the hierarchy into which the concept is placed, e.g. “B-cell lymphoma
(disorder)”. This guarantees a bijective function between the set of SNOMED CT
codes and the set of SNOMED CT FSNs. Like in all terminology systems, the set of
synonyms in SNOMED CT never fully covers the linguistic diversity of a domain. This
is the reason for advocating so-called interface terminologies as containers for
synonyms and close-to-user expressions in general, which should be constructed
bottom-up and linked to reference terminologies [6, 7].
2. Material and Methods
The international release of SNOMED CT is a hybrid between a reference
terminology and a user interface terminology, according to [6]. Ambiguous terms are
especially frequent when stripping the hierarchy tag from FSNs: “B-cell lymphoma” is
therefore a synonym both of the SNOMED CT concept “B-cell lymphoma (disorder)”
and “B-cell lymphoma (morphologic abnormality)”. Acronyms rarely appear as
synonyms in SNOMED CT, because naming conventions [8] require that acronyms be
followed by their expansion, such as in “PIN - Prostatic intraepithelial neoplasia”, with
a dash enclosed in white space characters as delimiting sequence. For retrieving
acronyms in text, however, the expanded form needs to be suppressed in the matching
procedure; in this example this means the match is done with “PIN”. These are cases
where lexical ambiguity becomes a serious issue, e.g. where the system has to choose
between “Prostatic intraepithelial neoplasia” and “Pressure-induced nystagmus” when
matching “PIN”. In the following we consider this a special case of lexical ambiguity.
Our analysis of the ambiguity of terms in SNOMED CT is based on all active
concepts and terms from the January 2017 release. Lexical ambiguity is investigated at
two different levels, viz. (i) full terms as obtained from the SNOMED CT description
table, and (ii) acronym extracts that correspond to our definition (see below).
To this end, two dictionaries D1 and D2 are built. D1 collects all SNOMED CT
concept IDs to which an ambiguous term was assigned. D2 does the same for acronyms,
by matching the abovementioned pattern and ignoring the expansion section. The
selection of acronyms was done according to a simple rule of thumb, which proved
highly selective in medical terminologies: only tokens between two and seven
characters, in which at least the second or third character is capitalised, are considered
acronyms.
Both D1 and D2 are then analysed according to the following criteria:
Combinations of SNOMED CT hierarchy tags, in order to better delineate
where ambiguities occur.
Cases where concepts that belong to ambiguous terms are semantically related
by direct non-taxonomic links like Associated morphology or Has active
Ingredient.
Cases where concepts that belong to ambiguous terms are semantically related
by direct taxonomic (is-a) links.
3. Results
Table 1 characterises either set. The existence of outliers is explained by the fact
that a few acronyms are not followed by their expansions. For example, the acronym
“O/E” (which means “on examination”) occurs in hundreds of terms like “O/E - toe” or
“O/E - eye”. This shows that SNOMED CT’s acronym – expansion pattern is not
specific. It is not very sensitive either, because there are occurrences of acronyms that
do not comply with the naming pattern at all. For example, “ENT” (Ear - nose - throat)
is never introduced according to the naming pattern. It occurs not only in terms like
“O/E - ENT” but also in isolation (just “ENT”) as synonym of “Ear, nose and throat
surgery”.
Table 1: Frequency and distribution of ambiguous readings of SNOMED CT terms.
Dictionary Count Cardinality Maximum
Mean Median
D1 (non-acronym terms) 7,439 2.02 2 6
D2 (acronyms) 899 5.54 2 1678
Regarding the five most frequent hierarchy tag patterns, Table 2 and 3 show very
different results comparing SNOMED CT full terms (D1) and acronyms extracted
according to the SNOMED CT acronym / definition pattern (D2).
Table 2: Leading patterns of concept tuples connected by the same SNOMED CT (non-acronym) term
Hierarchy tag combination Pattern Rate of non- Rate of
patterns count taxonomic links taxonomic links
| product | substance | 4,064 0.888 0.000
| disorder | morphologic abnormality | 1,047 0.707 0.000
| organism | organism | 221 0.000 0.452
| procedure | substance | 213 0.911 0.000
| procedure | procedure | 200 0.000 0.465
Other n-tuples (2 n 6) 1,694
Regarding D1, we see a high aggregation of ambiguous terms with two
combinations, viz. “| product | substance |” and “| disorder | morphologic abnormality |”.
These two distributions also exhibit a high degree of ontological connection, which is
also true for the combination “| procedure | substance |”. Taxonomic links between
concepts that share a term are quite frequent in all cases in which the ambiguity occurs
within the same hierarchy. This also applies to many for the less frequent patterns not
distinguished in Table 2.
In D2, the distribution between patterns is more balanced, and the degree of
connection between concepts that share the same acronym is lower.
Table 3: Leading patterns of concept tuples linked by the same acronym extracted from SNOMED CT
terms
Hierarchy tag combination Pattern Rate of non- Rate of
Patterns count taxonomic links taxonomic links
| disorder | disorder | 66 0.015 0.167
| disorder | procedure | 59 0.034 0.000
| procedure | procedure | 38 0.000 0.263
| procedure | substance | 33 0.333 0.000
| disorder | substance | 28 0.000 0.000
Other n-tuples (2 n 1678) 675
4. Discussion and recommendation
The way acronyms are introduced in SNOMED CT is neither specific nor sensitive.
The pattern recommended by SNOMED CT to characterise acronym/definition pairs
(e.g. “DNA - Did not attend”) is also found in as acronym/specialisation pairs (e.g.
“DNA - appointment mix-up”), which explains extreme cardinality outliers. Besides,
the overall number of acronyms in SNOMED CT is not high, compared to the size of
the terminology.
More than half of the term-level ambiguities are explained by concept pairs that
are also ontologically connected. It concerns the combination of product concepts with
substance concepts via the relation Has active ingredient, which relates, “Folinic acid
(product)” with “Folinic acid (substance)”. This is quite similar with disorder and
morphology concepts connected via Associated morphology, relating, e.g. “Solar
keratosis (disorder)” with “Solar keratosis (morphologic abnormality)”, as well as
substance concepts connected, e.g., via Component, such as “Curcumin stain
(procedure)” and “Curcumin stain (substance)”.
These frequent types ambiguities occur in a rather systematic way. Especially in
the case of | disorder | morphology |, these parings can be considered as dot categories
[9], i.e. complex categories that classify tightly connected concepts. Dot categories are
often not really discerned by language and common sense. A commonly cited example
for this is “book” as being both an information object and a physical object depending
on the context (e.g. “this thick book” is an incomprehensible
“ book”). Dot categories are well defined and easy to handle and
comply, in their majority, with the SNOMED CT concept model. More problematic are
lexical ambiguities in which the two competing concepts represent children and parents
in the taxonomy, which is most likely to be found in the procedure and the organism
branch. For instance, “Blepharotomy (procedure)” is a child of “Incision of eyelid
(procedure)” and has as synonym “Incision of eyelid”.
Nevertheless, the phenomenon of ambiguous terms in SNOMED CT has to be seen
in the context of the size of the terminology. We found 7,439 ambiguous SNOMED CT
terms and 899 ambiguous acronyms, which represents just 8,338 ambiguous terms.
Especially the diversity of acronyms found in SNOMED seems small compared to their
real occurrence in medical texts.
For the SNOMED CT curators we give the following recommendations:
Create awareness that the lexical coverage of SNOMED CT with synonyms is
limited.
Concentrate on SNOMED CT as reference ontology, leaving the maintenance
of collections of close-to-user terms (user interface terminologies) to user
groups and release centres, according to the ASSESS-CT recommendations.
Eliminate synonyms from parent concepts if there is already the same term in
a child concept.
Reconsider naming conventions.
If possible, complete missing relations between those concepts that are linked
by truly polysemous terms.
References
[1] Grange B, Bloom DA. Acronyms, abbreviations and initialisms. BJU Int. 2000 Jul;86(1):1-6.
[2] Acronym Finder. http://www.acronymfinder.com/
[3] SNOMED International (January 2017). http://www.snomed.org/snomed-ct
[4] Schober D, Smith B, Lewis SE, Kusnierczyk W, Lomax J, Mungall C, et al. Survey-based naming
conventions for use in OBO Foundry ontology development. BMC Bioinformatics. 2009 Apr
27;10:125
[5] Smith B, Ashburner M, Rosse C et al.: The OBO Foundry: coordinated evolution of ontologies to support
biomedical data integration. Nat Biotechnol 2007, 25: 1251–1255
[6] ASSESS-CT. Assessing SNOMED CT for Large Scale eHealth Deployments in the EU.
http://assess-ct.eu
[7] Schulz S, Rodrigues JM, Rector A, Chute CG. Interface Terminologies, Reference Terminologies and
Aggregation Terminologies: A Strategy for Better Integration. Accepted for MEDINFO 2017
[8] SNOMED CT Editorial Guide (January 2017). https://confluence.ihtsdotools.org/display/DOCEG
[9] Arapinis A,Vieu L (2015). A plea for complex categories in ontologies. Applied Ontology, 10(3-4), 285-
296.