Machine learning-based coreference resolution of
concepts in clinical documents
Henry Ware,1Charles J Mullett,2Vasudevan Jagannathan,1Oussama El-Rawas1
Objective Coreference resolution of concepts, although
a very active area in the natural language processing
community, has not yet been widely applied to clinical
documents. Accordingly, the 2011 i2b2 competition
focusing on this area is a timely and useful challenge.
The objective of this research was to collate coreferent
chains of concepts from a corpus of clinical documents.
These concepts are in the categories of person,
problems, treatments, and tests.
Design A machine learning approach based on graphical
models was employed to cluster coreferent concepts.
Features selected were divided into domain independent
and domain specific sets. Training was done with the
i2b2 provided training set of 489 documents with 6949
chains. Testing was done on 322 documents.
Results The learning engine, using the un-weighted
average of three different measurement schemes,
resulted in an F measure of 0.8423 where no domain
specific features were included and 0.8483 where the
feature set included both domain independent and
domain specific features.
Conclusion Our machine learning approach is
a promising solution for recognizing coreferent concepts,
which in turn is useful for practical applications such as
the assembly of problem and medication lists from
The Health Information Technology for Economic
and Clinical Health (HITECH) act passed in
February 2009 as part of the American Reinvest-
ment and Recovery Act (ARRA) sets aside signifi-
cant funds as incentives for the adoption of
electronic medical records (EMR). In particular, the
act calls for providers to demonstrate ‘meaningful
use’ of a certified EMR to qualify for financial
incentives. Evidence of meaningful use has been
defined, in part, as the capturing of structured
elements in an EMR such as problem lists, medi-
cations, procedures, allergies, and quality measures.
Identifying coreferent concepts is an essential part
of capturing these elements.
The 2011 natural language processing (NLP)
challenge’s focus on coreference (different terms in
a document that refer to the same concept) gives it
a high practical relevance in the marketplace. In
particular, the relationship of problems and treat-
ments in transcribed medical documents can be
used to assemble complete, yet precise, problem
and medication lists. For instance, the terms:
‘congestive heart failure,’ ‘CHF,’ ‘systolic heart
failure,’ and ‘heart failure’ may be coreferentdthat
is, they may all describe the same condition in the
same patient. The objective of the challenge is to
assemble all such relationships into chains of core-
ferent problems, treatments, tests, or persons.
This year, i2b2 provided two sets of annotations
using two different guidelines for the challenge:
ODIE and i2b2. The ODIE guidelines were more
elaborate and detailed than the i2b2 guidelines. We
entered track 1c, with i2b2 annotations, to capitalize
on our experience with this annotation set from last
year’s competition. The training document set had
the following set of annotations: Test, Problem,
Procedure, Person, and Pronoun. The annotation
chains (coreferent chains) were of the following
categories: Test, Problem, Procedure, and Person.
The effort discussed in this report details our
solution to the coreference challenge using the i2b2
There has been significant recent research effort in
the NLP community to address the problem of
coreference resolution. The BART system1uses
a classifier that relies on a feature set that can be
tuned to different languages. Facets of the feature
set described in that system include: gender agree-
ment (he/she), number agreement (singular/plural),
animacy agreement (him/it, them/that), string
match (exact or partial match), distance between
concepts (physical separationdnumber of charac-
ters, words between mentions), and aliases (syno-
nyms). The authors also describe a semantic tree
compatibility, in which a frame of slot-value pairs
(that include the above features) is associated with
each concept and the frames are compared for
compatibility. Most approaches to coreference
resolution rely on supervised-learning techniques.
However, the method used by Raghunathan and
coworkers2uses a completely different approach.
They order the feature sets to resolve coreference
from most precise to least precise and apply them
successively to collate coreferent chains. A cluster-
ranking approach, where coreference resolution is
recast as a problem of finding the best preceding
cluster to link a particular mention, is discussed in
the paper by Rahman and Ng.3
Overview of procedure
Our approach in this effort builds upon the methods
described by Culotta et al.4It uses a learning engine
and a feature set that is fine-tuned to the clinical
domain. The core learning engine is implemented
using the Scala programming language.
Machine learning approach
We used the ‘Factorie’ toolkit5to support the
learning task. The toolkit is used to implement
1M*Modal, Inc., Morgantown,
West Virginia, USA
2Department of Pediatrics, West
Morgantown, West Virginia,
Dr Vasudevan Jagannathan,
M*Modal, Inc., 235 High Street,
Suite 214, Morgantown, WV
Received 13 December 2011
Accepted 15 April 2012
Published Online First
12 May 2012
J Am Med Inform Assoc 2012;19:883e887. doi:10.1136/amiajnl-2011-000774883
Research and applications
factor graphs. In the factor graph, the mentions are represented
as nodes. Mentions which are coreferent in a given configuration
are connected by edges, which we call pairwise-affinity factors.
As the system considers different possible configurations, it
constructs a factor graph to represent each configuration.
For example, figure 1 shows an incorrect configuration with
four mentions divided into three chains. The mentions ‘gnr’ and
‘gram negative rods’ are chained with each other, but the
mention ‘gnr bacteremia’ is incorrectly omitted from the chain.
The system will consider adding ‘gnr bacteremia’ to the
correct chain as in figure 2.
It will also consider adding ‘gnr bacteremia’ to the chain
containing the mention ‘hypoxic’ as in figure 3. The features for
the pairwise-affinity factors are of two types. Some of the
features attempt to capture the relationship of the two
mentions. This uses a fairly standard hand-constructed list
including: distance metrics, gender agreement (he/she), laterality
agreement (left/right), number agreement (singular/plural),
overlap, synonyms in SNOMED, hypernyms (broader concepts)
in SNOMED, string equality, etc. For the second set of pair-wise
features, we take the cross-product of certain mention-wise
features. Mention-wise features include words, bi-grams, four
character prefixes, and enclosing section type. We also used
chain-wise factors, one per chain, to capture information about
the chain as a whole. As an example, a chain which included
both ‘Mr.’ and ‘she’ would be noted as having a gender incon-
sistency. The graph was trained using a maximum entropy
model with adaptive regularization of weight vectors (AROW)
Sampling was from the plausible permutations
generated for the mentions.
In the next section, we discuss feature selection in greater
Regardless of whether the features discussed above are used in
a pair-wise fashion or as an aspect of a single mention or a whole
chain, we can classify the feature selection as being domain
independent or domain dependent.
Domain independent features
These include four and five character prefixes, words, bi-grams,
string match, gender match, and number match. We also
considered headword (root/stem matches) and animacy match
approaches but lacked the development time to implement
these prior to the contest deadlines. Although we had access to
parts-of-speech tagging based on the cTAKES system,7we did
not use them. We anticipate trialing these as time permits in the
Clinical domain specific features
Here we implemented a variety of features, as follows:
< Laterality compatibleddetermines if the concept refers to the
left side or the right side: for example, left knee meniscus tear
versus right hip fracture.
< Sitedidentifies whether the body location site is compatible.
Clearly meniscus tear (site: knee) and hip fracture (site: hip)
are not. (Note: This feature was implemented in the end but
there was not enough time to train with this feature before
the competition test date).
< Section typedclinical documents sections are generally well
structured and a mention’s location within the document
conveys significant information. For instance, a section
labeled, ‘past medical history,’ clearly conveys information
about the patient’s past and any mention there may not be
coreferant to similar mentions in other sections.
< Aliasesdgeneral implementation of coreference resolution
tends to use WordNet for gathering aliases. However, for the
clinical domain, such aliases are better assembled using
SNOMED or other clinical vocabularies. We used the
Apelon TermWorks engine to search for aliases for concepts.
< Parents of mentionsdwe also used the SNOMED hierarchy
to determine the parents and grandparents of problem
< Acute or chronicdchronic conditions are generally coreferent
as they refer to patient conditions that persist over time. For
Four mentions configured as three chains.
Four mentions configured correctly as two chains.
Four mentions in alternate configuration.
884 J Am Med Inform Assoc 2012;19:883e887. doi:10.1136/amiajnl-2011-000774
Research and applications
example, hypertension and diabetes are chronic conditions.
Acute conditions on the other hand may or may not be
coreferent in the same document. So we introduced a feature
to recognize acute conditions. Using a data mining infra-
structure, we determined that terms such as the following co-
occur with acute conditions: acute, attack, stroke, accident,
infarction, exacerbation, meningitis, trauma, hemorrhage,
rupture, pneumonia, infiltrate, epilepticus, etc.
< Surgerydas for chronic/acute disambiguation for problems,
one can make a similar distinction between treatments that
are surgeries versus those that are medications versus those
that are device based.
< Sign or symptomdthis feature attempts to categorize
whether a problem concept is a sign or a symptom.
< Diagnostic proceduredthis feature categorizes a treatment or
a test as to whether it is a diagnostic procedure based on the
SNOMED list of diagnostic procedures.
< Screening proceduredthis feature categorizes a treatment or
a test as to whether it is a screening procedure based on the
SNOMED list of screening procedures.
< Body part mentiondthis feature analyzes a mention based
on the SNOMED body parts list.
< Related termsdwe used the Microsoft Bing search engine to
search the web for terms related to the mentions. To carry
out this search, we used Bing to search the ‘eMedicine’ (now
called Medscape) and ‘webMD’ websites and took the
intersection of terms (words) from the two searches to
determine related terms. Example are shown below:
< meniscal tear eMedicine: [, drez’s, disk, sports, good,
expendable, ga, that’s, surgery, center, medical, refer, tear,
howard, meniscal, medicine, miller, knee, drez, informa-
tion, connection, cartilage.]
< meniscal tear webMD: [, discomfort, replacement,
vertical, lateral, via, organizedwisdom, treatment, hip,
more,, jointreplacement, type, surgery, prior, doctor, disc,
medical, webmd:, risks,, tear, depends, knee, joint, tear.]
< Intersection of eMedicine and webMD: [, medical,
meniscus, tear, knee, treatments, tears]
Although this approach identified ‘knee’ as a related term of
‘meniscal tear,’ we found that using SNOMED vocabulary
afforded better results in general. However, the use of web
searches does appear interesting and promising and merits
further investigation. Of note, Google did not allow program-
matic use of their search engine, while Microsoft Bing provided
useful software sources to help with such searches.
< Temporal featuresdthis mostly identified whether a term is
current or past. If, for example, a concept occurred in the
‘past medical history’ section, it would be tagged as in the
past. We considered, but did not implement, future tense
assessment and tagging.
Technologies and support systems used
The whole NLP framework was developed using the Scala
programming language which supports both object-oriented
computing and functional programming. Implementations of
the language are available for Java and .NET platform. Our NLP
platform was built using the Java environment. We also relied on
the Apelon terminology environment to determine relevant
Evaluation metrics used
The evaluation metrics were supplied by the i2b2 contest
organizers. Four evaluation metrics have been specified. For an
excellent comparison of the metrics chosen by the i2b2 contest
see Recasens and Hovy.8Cai and Strube also have discussions
surrounding some of these metrics.9The various measures are:
< B3dmeasures10the number of mentions in the response set
(R) that are in common with the gold key set (K). The
precision and recall are computed as shown below:
where Rmiis the response chain (system response) for the ith
mention and Kmiis the gold standard. These are then summed
over the entire set.
B3dmeasures are overly sensitive to a large number of
singleton mentions, a fact that we verified by assigning every
mention as its own coreference chain. This resulted in a B3value
of 0.955, higher than any result we obtained in the actual runs.
< MUCdThis is a link-based scoring scheme, where a link is
the coreference relationship between two mentions.11The
measure fundamentally evaluates how many links are in
common between the sets R and K. Recall errors are equated
to missing links and precision errors linked to superfluous
links. The MUC measure ignores singletons and can be fooled
by assigning all mentions to be one big chain.
< CEAFdcomputes a similarity metric between key and
response.12In essence, to score a coreference task it attempts
to find the best one to one mapping between the ground
truth and the system output. This measure also is susceptible
to providing optimistic results in the presence of a large
number of singleton mentions.
< BLANCdBiLateral Assessment of Noun-phrase Coreferenced
developed by Recasens and Hovy,8is an attempt to address the
limitations of the above-mentioned measures. The metric
capitalizes on the fact that every mention falls into two
categories: one that is part of a coreference chain or one that is
part of a non-coreference chain. Precision and recall metrics are
calculated independently for both categories and the result is
then averaged. Singleton mentions will only contribute 50% of
the measure and hence will not overwhelm the metric when
large numbers of such mentions exist.
The BLANC measure, although computed, was not used by
the i2b2 organizers and is not included in our results.
RESULTS AND DISCUSSION
Table 1 shows the results from training using only the domain
independent features set. On our computing hardware, the
training phase required approximately 18 h to run on the 489
documents in the contest training set.
Table 2 shows the results from training using only the domain
independent and domain dependent features set that extensively
used the SNOMED vocabularies as discussed in the previous
sections. The training phase took approximately 40 h to run.
In general, we did less well on the MUC metric, and scored
particularly poorly in the ‘tests’ category by MUC analysis,
perhaps because the training documents had fewer instances of
‘tests’ than the other categories, offering less material for our
learning engine to build upon.
Table 3 shows the result of testing/evaluation carried out on 322
documents. The test required only 10 min to run. It used the
J Am Med Inform Assoc 2012;19:883e887. doi:10.1136/amiajnl-2011-000774885
Research and applications
model created from training on the domain independent
Table 4 shows the result for the test run, based on the model
created from using domain independent and domain dependent
features. This test run was completed in about 10 min.
Table 5 compares the results across all four runs.
The striking finding is that the addition of the domain depen-
dent features, which extensively used SNOMED concepts, did
not provide as much benefit in the scoring. Our assumption
when approaching the task was exactly the opposite, a notion
supported by the literature.13Multiple factors appear at play in
our results. The training samples were fairly extensive (for some
of the categories) and the testing samples were drawn from the
same corpus. Routine aliases were already captured in the
training set and therefore the benefit of the SNOMED vocabu-
lary was less powerful. In addition, our domain dependent
features were not targeted at pronoun resolutions which formed
the greater part of the recognition task for the challenge.
Our machine learning technique performed poorly at recog-
nizing tests, most likely because of the fewer number of tests in
the training documents to engage the learning process. We also
performed relatively weakly at pronouns recognition, primarily
due to lack of attention to this facet of the challenge by us, the
investigators. Disambiguating pronouns are of less practical
significance to a commercial entity such as ours, than collating
a precise collection of problems or medications.
Tables 1, 2 and 5 also show that the machine learning algo-
rithm did not over-fit. In fact, the results from the test run were
quite comparable to the training run.
We looked at the errors made by the learning engine and placed
them in the following general classes:
1. Synonymy failuredfailure to recognize two terms repre-
senting the same concept. Examples here include: ‘cardiac dz’
and ‘coronary artery disease,’ ‘Transesophageal Echocardio-
gram’ and ‘The TEE,’ ‘SBP’ and ‘his blood pressure,’
‘Oxycodone’ and ‘Oxycontin.’ Failure to recognize synonyms
typically leads to a number of false negatives.
2. Temporality failuredfailure to recognize the temporality of
a concept. Example: ‘surveillance cultures’ versus ‘Plan to
have surveillance cultures’ in the future. Although the
concept is clearly the same, it is not referring to the same
Results for domain independent feature training
Category Statistical measuresRecall F measurePrecision
Average total F measure
Results for domain independent and dependent feature
CategoryStatistical measures Recall F measurePrecision
Average total F measure
Results for domain independent testing run
Category Statistical measuresRecallF measurePrecision
Average total F measure
Results for domain independent and dependent testing run
CategoryStatistical measuresRecallF measurePrecision
Average total F measure
886J Am Med Inform Assoc 2012;19:883e887. doi:10.1136/amiajnl-2011-000774
Research and applications
3. Contextual cues missed. In one example: Patient Name: Mrs.
YYY. Later the patient is referred to as ‘the patient’ and with
various pronouns. In another example, ‘chronic back pain’ is
falsely linked to ‘chronic pain’ and ‘pain’ where the later two
references to pain refer to other types of pain understandable
within the context of the note.
Our efforts to use SNOMED vocabulary were clearly targeted
at addressing errors that arose from failure to recognize synon-
ymous terms. However, that effort has not yet borne fruit.
Temporal features were not implemented, which might have
eliminated some of the errors. Recognizing when a concept is
coreferent and when it is not remains a challenge for NLP
The coreference challenge has focused attention on an area that
has sometimes been ignored in clinical document analysis. Our
machine learning approach is a promising solution to the task of
automating the assembly of problem and medication lists from
clinical documents. Our results also suggest that having a high-
quality set of annotated training documents is the keydand
domain independent features are sufficient for obtaining
reasonable results. In real world deployment, however, we
consider that it will be critical to employ well formulated
domain specific features to provide for a more robust engine that
will work across different document types and sources. Of
course, problem and medication lists generated through this
method will need to be manually reviewed and corrected, but
the techniques explored here will capably render an excellent
first draft for a human validator.
Acknowledgments We would like to thank our M*Modal management for
supporting i2b2 and our participation in this challenge.
Funding The 2011 i2b2/VA challenge and the workshop are funded in part by grant
number 2U54LM008748 on Informatics for Integrating Biology to the Bedside from the
National Library of Medicine. This challenge and workshop are also supported by the
resources and facilities of the VA Salt Lake City Health Care System with funding
support from the Consortium for Healthcare Informatics Research (CHIR), VA HSR HIR
08-374 and the VA Informatics and Computing Infrastructure (VINCI), VA HSR HIR
08-204, and the National Institutes of Health, National Library of Medicine under grant
Competing interests None.
Provenance and peer review Not commissioned; externally peer reviewed.
Data sharing statement The i2b2 organizers are making the data used for the
challenge available to research institutions.
Broscheit S, Poesio M, Ponzetto SP, et al. BART: a multilingual anaphora resolution
system. Proceedings of the 5th International Workshop on Semantic Evaluation. PA,
USA: Association for Computational Linguistics Stroudsburg, 2010:104e7.
Raghunathan K, Lee H, Rangarajan S, et al. A multi-pass sieve for coreference
resolution. Conference on Empirical Methods in Natural Language Processing. PA,
USA: Association for Computational Linguistics Stroudsburg, 2010:492e501.
Rahman A, Ng V. Supervised models for coreference resolution. Conference on
Empirical Methods in Natural Language Processing. PA, USA: Association for
Computational Linguistics Stroudsburg, 2009:968e77.
Culotta A, Wick M, Hall R, et al. First-order probablistic models for coreference
resolution. Proceedings of NAACL HLT. PA, USA: Association for Computational
Linguistics Stroudsburg, 2007:81e8.
McCallum A, Schultz K, Singh S. FACTORIE: probabilistic programming via
imperatively defined factor graphs. In: Bengio Y, Schuurmans D, Lafferty J, et al, eds.
Advances in Neural Information Processing Systems. Cambridge, MA, USA: MIT
Crammer K, Kulesza A, Dredze M. Adaptive regularization of weight vectors. In:
Bengio Y, Schuurmans D, Lafferty J, et al, eds. Advances in Neural Information
Processing Systems. Cambridge, MA, USA: MIT Press, 2009;22:414e22.
Savova GK, Masanz JJ, Ogren PV, et al. Mayo clinical Text Analysis and Knowledge
Extraction System (cTakes): architecture, component evaluation and applications.
J Am Med Inform Assoc 2010;17:507e13.
Recasens M, Hovy E. BLANC: Implementing the Rand index for coreference
evaluation. Nat Lang Eng 2011;17:1e26.
Cai J, Strube M. Evaluation Metrics For End-to-End Coreference Resolution Systems.
Stroudsburg, PA, USA: Association for Computational Linguistics, SIGDIAL,
Bagga A, Baldwin B. Algorithms for scoring coreference chains. The First
International Conference on Language Resources and Evaluation Workshop on
Linguistics Coreference. Stroudsburg, PA, USA: Association for Computational
Vilain M, Burger J, Aberdeen J, et al. A model-theoretic coreference scoring
scheme. Proceedings of the 6th Conference on Message Understanding.
Stroudsburg, PA, USA: Association for Computational Linguistics, 1995:45e52.
Luo X. On coreference resolution performance metrics. Proceedings of the
conference on Human Language Technology and Empirical Methods in Natural
Language Processing. Stroudsburg, PA, USA: Association for Computational
Tetrault JR. A corpus-based evaluation of centring and pronoun resolution. Comput
Results comparison across the four runs
PAGE fraction trail=4.75
J Am Med Inform Assoc 2012;19:883e887. doi:10.1136/amiajnl-2011-000774887
Research and applications