Good Stuff Happens in 1:1 Meetings: Why you need them and how to do them well
Automated Identification of Similar Health Questions
1. Automated Identification of Similar Health Questions
Geoffrey W. Rutledge MD, PhD
Chief Medical Information Officer
HealthTap.com
Introduction Discussion
People with health questions are increasingly The three similarity criteria tested are The problem of identifying semantically
looking for physician answers to their health 1. Lexical identity after removal of all non- similar health questions is complicated by the
questions online. Given the repetitive nature alphanumeric characters variability of consumer health language and
of common questions, there is a high value in 2. Sum of semantic weights of all matching the difficulties that consumers have in
identifying previously answered questions health concepts spelling medical terms. A comprehensive
that are semantically similar (or identical) to 3. Sum of weights of only the moderate or ontology and synonym set of consumer
each new question, so that an answer can be high weight matching health concepts health terms enabled the accurate detection
given without delay and without waiting for a of a large fraction of semantically similar
new answer from a physician. consumer health questions that were entered
1
(2)
in an online health site.
Background 0.9
True
positive
rate
(3)
0.8
Previous methods to evaluate question pairs 0.7
The automated identification of similar
were based on sentence similarity [1,2] and 0.6
0.5
consumer health questions is challenging
are not suitable for consumer health 0.4
(1)
because of the common occurrence of
questions, which contain many consumer- 0.3
0.2
complex, colloquial, and often misspelled
health variations and frequent misspellings of 0.1
medical terms in consumer health questions.
0
medical concepts. We developed a method 0
0.1
0.2
0.3
0.4
We collected online health questions and
to identify questions with “high semantic False
positive
rate
their paired "nearest search result" matching
similarity” from a corpus of consumer health questions to evaluate 3 question similarity
questions and answers, in which the metrics. The best performing metric was
Examples of medical concepts:
questions and answers are character limited moderate weights: antibiotics, heart disease,
based on the sum of semantic weights for all
to 150 and 400 characters respectively. sharp pain matching health concepts from a
high weights: penicillin, congestive heart failure, comprehensive ontology of consumer health
Method squeezing chest pain terms and common misspellings, with a
We compare the text of new questions to the measured sensitivity of 0.61 and specificity of
closest matching question from the Q&A 0.99.
corpus. For a set of 1,000 questions and their Results
closest match, we evaluated the sensitivity We compared the three similarity criteria [1] The Evaluation of Sentence Similarity Measures, I.-
and specificity of alternative similarity criteria Y. Song, J. Eder, and T.M. Nguyen (Eds.): DaWaK
against an expert assessment of question
2008, LNCS 5182, pp. 305–316, 2008.
for the assertion of “high semantic similarity.” pair similarity. The sensitivities and [2] Finding Similar Questions in Large Question and
We first identified the most similar question specificities for the three criteria are (1) 0.47, Answer Archives. Jiwoon Jeon, W. Bruce Croft and
within the Q&A corpus using a search engine 1 (2) 0.61, 0.99 (3) 0.63, 0.97, as plotted on Joon Ho Lee. CIKM’05, October 31–November 5, 2005
augmented with a semantic-weight driven the chart of False positive versus True
ontology of consumer health concepts, which positive rates (ROC). The criterion with the
includes a rich set of synonyms of consumer best performance was Sum of semantic
health terms, and frequent misspellings of weights of all matching concepts. We are hiring
consumer health terms. geoff@healthtap.com