DataPDF Available

Bringing2017Zhang_Slides

Authors:
Bringing Semantic Structures to User Intent
Detection in Online Medical Queries
In#collaboration#with#Wei#Fan#(Tencent#Medical#AI#Lab),#
Nan#Du,#Yaliang#Li (Baidu#Research#Big#Data#Lab),#
Chun-Ta#Lu,#and#Philip#S.#Yu#(University#of#Illinois#at#Chicago,#Tsinghua#
University)
1
Chenwei Zhang
University of Illinois at Chicago
Dec.12, 2017
IEEE BigData 2017
Chenwei Zhang
Online medical forums
The world will face a shortfall of nearly 13 million
healthcare professionals by 20351;
80 percent of Internet users, or about 93 million Americans,
have searched for a health-related topic online2;
processes over 6 billion search queries every day,
while 60 million of them are healthcare-related text
queries3;
has 120 million registered users and more than 22
million unique daily visitors4.
Voice Assistants
2
Online medical queries generated by
users
Thanks for your question. There are many
different possible explanations for what
makes a person dizzy. So it is important to
speak with your doctor. Often, there are
problems related to …
Dr. Alex Dwight
Said
Hi i want know what wrong with me. When
I sit in chair or ride a bike or sleep and my
back is at bed push I get dizzy. I checked
my heart is ok i have asthma.
P0020292222
Said
Hey Siri why I am
feeling headache
all the time after I
take Ibuprofen
1. http://www.nbcnews.com/id/3077086/t/more-people-search-health-online 2. http://www.who.int/mediacentre/news/releases/2013/health-
workforce-shortage/en/ 3. http://science.china.com.cn/2016-1124content 9180719.htm 4. http://club.xywy.com/
Chenwei Zhang
For information-seeking purposes
“I got a temperature of 103 degrees. Should I take
Tylenol?”
“Can I get a kidney infection from a tooth infection?”
“How long after gastritis surgery can I have spicy
food?”
Knowing the intention can potentially achieve a
transformative effect on improving costumers satisfaction
for healthcare solution providers
E.g. looking for some medicine -> recommend the
nearest pharmacy
E.g looking for treatment -> find the nearest clinic
specialized in certain field
3
Online medical queries generated by
users
Query Intention: the information-seeking
behavior of a user based on what the user
described
Chenwei Zhang
Challenge 1: Intention Modeling
Define a discriminative representation for
intentions
Text Classification (News)
Trump has to live with the consequences of his
Israel decision Politics
Manu Ginobili sank the Boston Celtics with a
game-winning dagger Sports
Intention Detection
“How long after gastritis surgery can I have
spicy food?” < Surgery, Diet >
4
The existence of some topic-specific words has a
dominating contribution to its classification.
The resulting intention is encoded in a query through
1) the mention of different concepts: Surgery, Diet;
2) semantic transition between concepts Surgery -> Diet.
Chenwei Zhang
Concept Graph:
Node: Concept Mention
Directed Edge: Concept Transition
Observation: multiple semantic transitions in a single
question may conjugate with each other by
mentioning the same concept.
5
Graph-based Structured Intention
Cause
Treatment
DiseaseSymptom
Cause
Disease
Medicine
Symptom
Cause
Treatment
DiseaseSymptom
Cause
Treatment
Disease
Symptom
Cause
Disease
Symptom
Disease
Medicine
Symptom
Instruction
Medicine
Side Effect
Symptom
Surgery
Sequela
Disease Treatment
Disease
Symptom
Symptom -> Medicine -> Instruction
My 3 year old is sick with a temperature of 100 degrees she
can't keep anything down including liquids. What kind of
medicine should I give my child, and how much?
Examine
Instruction
Diet Cause
Treatment
Surgery
Recover
Risk
Sequela
Syndrome
Diagnosis
Fee
Department
Disease
Medicine
Side Effect
Symptom
Chenwei Zhang 6
The Concept Transition Inference Problem
Examine
Instruction
Diet Cause
Treatment
Surgery
Recover
Risk
Sequela
Syndrome
Diagnosis
Fee
Department
Disease
Medicine
Side Effect
Symptom
My 3 year old is sick with a temperature
of 100 degrees she can't keep anything
down including liquids. What kind of
medicine should I give my child, and
how much?
Concept Transition Inference
Examine
Instruction
Diet Cause
Treatment
Surgery
Recover
Risk
Sequela
Syndrome
Diagnosis
Fee
Department
Disease
Medicine
Side Effect
Symptom
!" #$ %& ' (
"
)*#$ %
+*& '
)*(
Chenwei Zhang
Challenge 2: Diverse Expressions
7
Intentions are expressed implicitly and diversely
Explicit/Implicit Concept Mentions
Concept: Medicine
Explicit Mention: Tylenol, Ibuprofen, or xxx
caplet/capsule/drop/syrup
Implicit Mention: remedy, which
medication/medicine/drug
Concept: Symptom
nose plugged, blocked nose and sinus congestion
Diverse Semantic Transitions Expressions
I have (got) a fever, should Itake Tylenol?
Which medicine should I take if I’m running a fever?
I’ve come down with a fever, should I take Aspirin?
Is it okay to use ibuprofen when I'm running a temperature?
My temperature is 103, can I use Advil?
Literally different but they have the same concept transition
Symptom -> Medicine
Chenwei Zhang
Challenge 3: Domain Coverage
Identical sentences are rarely observed or not observed at all.
<Symptom, Medicine > < Symptom, Medicine >
8
Labeled queries only cover a small domain of all
possible topic-specific words that users might provide.
103 degree
sweating
headache
fatigue
Tylenol 103 degree
Tylenol
Aspirin
Advil
Ibuprofen
High human cost in human labeling
I got a temperature of 103 degrees. Should I take Tylenol?
I got ach in my stomach. Should I take aspirin?
LABEL
Chenwei Zhang
Semantic-Syntax Representation
Concept / Transition Encoder
Graph-based Collective Inference
9
Proposed model
Query
Word POS Tag
Embedding Embedding
RNN RNN
Concept Encoder Transition Encoder
Concept
Transitions
Concepts
Active Concept Graph
Chenwei Zhang
Help dealing with ambiguous words and diversified
expressions
Disambiguation:
Diverse Expressions: similar word embedding/POS
patterns
10
Semantic-Syntax Representation
fever
5. https://www.merriam-webster.com/dictionary/fever
Word embedding
Part-of-speech
embedding
noun
indicate
Symptom
a rise of body temperature above
the normal5
verb
indicate Disease
any of various diseases of which
fever is a prominent symptom5
Query
Word POS Tag
Embedding Embedding
RNN RNN
Concept Encoder Transition Encoder
Concept
Transitions
Concepts
Active Concept Graph
Chenwei Zhang 11
Concept / Transition Encoder
My 3 year old is sick with a temperature
of 100 degrees she can't keep anything
down including liquids. What kind of
medicine should I give my child, and
how much?
My 3 year old is sick with a
temperature of 100 degrees she can't
keep anything down including
liquids. What kind of medicine should
I give my child, and how much?
< Symptom, Medicine > <
Medicine,Instruction >
Query
Word POS Tag
Embedding Embedding
RNN RNN
Concept Encoder Transition Encoder
Concept
Transitions
Concepts
Active Concept Graph
%
+*
'
)*
!
Chenwei Zhang
Exploit the correlations of the inferred concept
transitions and corresponding concepts.
12
Graph-based Co-inference
Query
Word POS Tag
Embedding Embedding
RNN RNN
Concept Encoder Transition Encoder
Concept
Transitions
Concepts
Active Concept Graph
Concept Transitions
Concept Mentions
Symptom
Medicine
Instruction
Transition
Encoder
Instruction
Medicine
Symptom
,-./& .
0/1
.
0/
'*
Instruction
Medicine
Symptom
Concept
Encoder
A mutual transfer loss:
2 3
0/& .
0/
# 45-3
0/& .
0/6.)+45-.
0/& 3
0/61
3
0/
6.
6
4789 #, ./& .
0/
:;
:2 3
0/& .
0/
Minimize 4789 via gradient
based optimizers
Chenwei Zhang
200 thousand registered doctors
Over 80 million registered users
64 million medical queries as the corpus
Covering over 10 thousand diseases
10 thousand labeled medical queries(70% training, xvals)
The vocabulary contains 382,216 words.
13
Medical Data Set
13.8
±6.1
average length
of question
3.6
±0.8
2.47
±0.7
average number
of concepts
average number of
concept transitions
Chenwei Zhang
LR: Logistic Regression model + POS tagging features and word
representations.
NNID-JM: the neural network intention detection model with joint modeling.
CI: the concept inference model which only infers mention of concepts from
queries with the concept encoder.
CTI:the concept transition inference model without co-inference.
coCTI: the concept transition inference model with co-inference.
coCTI-MTL: the proposed model with co-inference and a mutual transfer
loss
14
Baselines
Query
Word POS Tag
Embedding Embedding
RNN RNN
Concept Encoder Transition Encoder
Concept
Transitions
Concepts
Active Concept Graph
CI CTI
Query
Word POS Tag
Embedding Embedding
RNN RNN
Concept Encoder Transition Encoder
Concept
Transitions
Concepts
Active Concept Graph
coCTI
Query
Word POS Tag
Embedding Embedding
RNN RNN
Concept Encoder Transition Encoder
Concept
Transitions
Concepts
Active Concept Graph
coCTI-MTL
Chenwei Zhang 15
Performance
0 0.2 0.4 0.6 0.8 1
False Positive Rate
0
0.2
0.4
0.6
0.8
1
True Positive Rate
LR (area = 0.7450)
NNID-JM (area = 0.7981)
CTI (area = 0.8020)
coCTI (area = 0.8483)
coCTI-MTL (area = 0.8731)
AUC-ROC
Micro/Macro-AUC Coverage Error / LRAP
Chenwei Zhang
Problem Studied:
Model and discover user intentions with structured concept
transitions in medical text queries
Proposed Model:
A graph-based inference approach
Semantic-Syntax Representation
Concept/Transition Encoder
Graph-based Collective Inference with Mutual Transfer
16
Summary
Chenwei Zhang,+Nan+Du,+Wei+Fan,+Yaliang Li,+Chun-Ta+Lu+and+Philip+S.+Yu
czhang99@uic.edu,+nandu@baidu.com,+davidwfan@tencent.com,+
yaliangli@baidu.com,+psyu@uic.edu,+
17
Thanks!
Bringing Semantic Structures to User Intent
Detection in Online Medical Queries
IEEE BigData 2017

File (1)

Content uploaded by Chenwei Zhang
Author content
ResearchGate has not been able to resolve any citations for this publication.
ResearchGate has not been able to resolve any references for this publication.