In community-based health services, vocabulary gap between health seekers and community generated knowledge has hindered data access. To bridge this gap, this paper presents a scheme to label question answer(QA) pairs by jointly utilizing local mining and global learning approaches. Local mining attempts to label individual QA pair by independently extracting medical concepts from the QA pair itself and mapping them to authenticated terminologies. However, it may suffer from information loss and lower precision, which are caused by the absence of key medical concepts and presence of irrelevant medical concepts. Global learning, on the other hand, works towards enhancing the local mining via collaboratively discovering missing key terminologies and keeping off the irrelevant terminologies by analyzing the social neighbors. Practically, this unsupervised scheme holds potential to large-scale data.
The rise of digital technologies has transformed the
patient-doctor relationships. Nowadays, when patients
struggle with their health concerns, the majority usually
explore the Internet to research the problem before and
after they see their doctors. For example, 70% of Canadians
turned to Internet to look up health-related information in
2009 [
To better serve the needs of health seekers,
communitybased health services have emerged as effective platforms
for health knowledge dissemination and exchange, such
as HealthTap1, HaoDF2 and WenZher[
In many cases, the community generated health content may not be directly usable due to the vocabulary gap, since participants with diverse backgrounds do not necessarily share the same vocabulary. Take HealthTap as an example. The same question may be described in substantially different ways by two individual health seekers. On the other hand, the answers provided by doctors may contain acronyms with multiple possible meanings, and nonstandardized terms. 1https://www.healthtap.com/ 2www.haodf.com
Noun Phrase
Extractor Medical Concept
Identifier Medical Concept
Normalizer Local Coding
Likeliness Pregnancy
Birth Control Pregnancy
Birth Control Unplanned Pregnancy
Contraception Corpus-aware
Vocabulary Unplanned Pregnancy Contraception What is the likeliness a pregnancy could occur on birth control? Unplanned Pregnancy Contraception Uses Contraceptive Sheath What is the likeliness a pregnancy could occur on birth control? Lexical Similarities
Terminology Sharing Network
Inter-expert Relationship Hierarchical Terminology Relationship
In this work, we define medical concepts as medical
domain-specific noun phrases, and medical terminologies as
authenticated phrases by well-known organizations that are
used to accurately describe the human body and associated
components, conditions and processes in a science-based
manner. Even though some health communities have
recently suggested doctors to annotate their answers with
medical concepts, we cannot ensure that they are medical
terminologies. Meanwhile, the tags adopted by doctors
often vary greatly [
It is worth mentioning that there already exist several
efforts dedicated to research on automatically mapping
medical records to terminologies [
To overcome these limitations, we propose a novel scheme that is able to code the QA pairs with corpus-aware terminologies. As illustrated in Figure 2, the proposed scheme consists of two mutually reinforced components, namely, local mining and global learning. 3.1
Local mining aims to locally code the QA pairs by extracting the medical concepts from individual instance and then mapping them to terminologies based on the external authenticated vocabularies. To accomplish this task, we establish a tri-stage framework, which includes noun phrase extraction, medical concept detection and medical concept normalization.
To extract all the noun phrases, we initially assign partof-speech tags to each word in the given QA pair by Stanford POS tagger3. We then extract tag sequences that match a fixed pattern of part-of-speech tags as noun phrases from the texts. This pattern is formulated as follows. (Adjective|N oun) (N oun ?(Adjective|N oun) N oun:
P reposition) (1) A sequence of tags matching this pattern ensures that the corresponding words make up a noun phrase. For example, the following complex sequence can be extracted as a noun phrase: “ineffective treatment of terminal lung cancer”.
Inspired by the efforts in [
As aforementioned, we cannot ensure that all medical
concepts are standardized terminologies. Take “birth control”
as an example. It is recognized as a medical concept by
our approach, but it is not an authenticated terminology.
Instead, we should map it into “contraception”. Therefore,
it is essential to normalize the detected medical concepts
according to an appropriate external standardized dictionary
and this normalization is the key to bridging the vocabulary
gap. In this work, we use SNOMED CT5 as our dictionary,
since it provides the core general terminologies for the
electronic health record and formal logic-based hierarchical
structure. The terminologies and their descriptions in
SNOMED CT are first indexed6. We then search each
medical concept against the indexed SNOMED CT. For
the medical concepts with multiple matched results, e.g.,
two results returned for “female”, we keep all the returned
terminology candidates for further selection. Enlightened
by Google distance [
Local mining, however, may suffer from various problems. The first problem is incompleteness. This is because some key medical concepts may not explicitly present in the QA pairs. The QA pair illustrated in Figure 2 shows an example of this situation, where the accurate terminology: “use contraceptive sheath” is absent from the QA pair. The second one is the lower precision. This is due to some irrelevant medical concepts explicitly embedded in the QA pairs, and are mistakenly detected and normalized by the local approach. For instance, given the question, “What are the risks getting pregnant and giving birth later in life ? ”, the terminology “finding of life event” as normalized from the irrelevant medical concept “life” is assigned as code. It is less informative to capture the main intent. 3.2
It is noteworthy that most previous efforts, including our local approach, attempted to map the QA pairs directly to the entries in external dictionaries without any pruning. This approach often presents problems since the external dictionaries usually cover relatively comprehensive terminologies and are far beyond the vocabulary scope of the given corpus. It may result in the deterioration in coding performance in terms of efficiency and effectiveness. The problem is caused by the over-widened scope of vocabularies, which may bring in unpredictable noises and make the precise terminology selection challenging. As a byproduct, a corpusaware terminology vocabulary is naturally constructed by our local mining approach, which can be used as terminology space for further learning.
Let Q = {q1; q2; :::; qN } and T = {t1; t2; :::; tM }
respectively denote a repository of QA pairs and their associated
4http://www.ldc.upenn.edu/
5http://www.ihtsdo.org/snomed-ct/
6http://viw2.vetmed.vt.edu/sct/menu.cfm
locally mined terminologies. The target of global learning
is to learn appropriate terminologies from the global
terminology space T to annotate each q in Q. In this work,
the global learning task is regarded as a multi-label learning
problem[
M { arg min ∑
F i=1 Ω(fi) + L(fi) +
M } ∑ Rij∥fi − fj∥2 ; (2) j=1 where M refers to the number of classes, i.e., the number of medical terminologies to be assigned. Vector fi is the ith column of F, representing the relevance scores of each QA pair to the i-th terminology. Ω(f) and L(f) denotes the regularizer on the hypergraph and empirical loss, respectively. In addition, Rij is the inter-terminology relationship between terminology i and terminology j. They are mined by exploiting the external well-structured ontology, which are able to alleviate the granularity mismatch problems and reduce the irrelevant sibling terminologies. By differentiating the above equation with respect to F, we can obtain a closed-form solution.
The philosophy to formulate these three objectives is as
follows. The first objective aims to guarantee that the
relevance probability function is continuous and smooth in
semantic space. This means that the relevance probabilities
of semantically similar QA pairs should be close to each
other. The second objective is ensured by the empirical
loss function, which forces the relevance probabilities to
approach the initial roughly estimated relevance scores. These
two implicit constraints are widely adopted in
rerankingoriented approaches [
When it comes to hypergraph construction, the N QA
pairs from Q are regarded as vertices and they are connected
by three types of hyperedges. The first type takes each
vertex as a centroid and forms a hyperedge by circling
around its k-nearest neighbors based on QA pair content
similarities. This procedure was first adopted in [
We crawled more than 109 thousand QA pairs from ```````````
Metric Apporach
LocalMining
and locally+globally recommended terminologies. Answers are not displayed due to limited space.
QA pairs Is it safe to color my hair
during pregnancy ? If I get an infection caused by gum disease, can that be transferred to my fetus ?
Locally Mined Terminologies
Local Mining + Global Learning hair structure, dyed hair, feeling safe,
hair structure, patient currently pregnant, patient currently pregnant, first trimester pregnancy... infectious disease, gingival disease, entire fetus, inflammation, periodontal disease...
coal tar allergy, hair color change,
disorder of endocrine system... infectious disease, prematurity of fetus, gingival disease, periodontal disease low birth weight infant...
HealthTap, which involve 5; 958 unique doctors. For ground truth construction, we invited three professionals with master degrees majored in medicine programme. labelers were trained with a short tutorial and a set of demonstrating examples. A majority voting scheme among the three labelers can partially alleviate the subjectivity problem. The annotators were required to label only top five recommended terminologies for each QA pair, and they were labeled either as “positive” or “negative”. 100 QA pairs were labeled as testing set.
We adopted two metrics that are able to characterize
precisions from different aspects. The first is average S@K
over all testing QA pairs, which measures the probability
of finding a relevant terminology among the top K
recommended ones. To be specific, for each testing QA pair, S@K
is assigned to 1 if a relevant terminology is positioned in the
top K and 0 otherwise. The second one is average P @K that
stands for the proportion of recommended terminologies
that are relevant[
jCj where C is a set of the top K terminologies, and R is the manually labeled positive ones.
We can see that the local mining approach achieves the worst performance. This is reasonable, because irrelevant concepts may be mapped to terminologies because of their presence in the QA pairs. pair samples with locally minded terminologies and locally+globally recommended ones. Intuitively, the terminologies are more comprehensive and reliable after enhancement with global learning.
This paper presented a medical terminology assignment scheme to bridge the vocabulary gap between health seekers and community generated knowledge.
A strong unified framework of local mining and global learning is proposed to tackle this research issue, instead of the conventional isolated utilization.
It proposes the concept entropy impurity approach to comparatively detect and normalize the medical concepts locally, which naturally construct a corpus-aware terminology vocabulary with the help of external knowledge. In addition, it builds a novel global learning model to enhance the local coding results. This model seamlessly integrates various heterogeneous cues.
In the future, we will investigate how to flexibly organize the unstructured medical content into user needs-aware ontology by the recommended medical terminologies.
This work was supported by NUS-Tsinghua Extreme Search project under the grant number: R-252-300-001-490.