large-scale discourse analysis of counseling conversations · •collaboration with a nonprofit...
TRANSCRIPT
Large-scale Discourse Analysis of Counseling Conversations
Kevin Clark, Tim Althoff, Jure Leskovec
Mental Health by the Numbers
• 43.8 million adults (18.5%) in the U.S. experience mental illness each year
2
Statistics provided by the National Institute of Mental Health: http://www.nimh.nih.gov/
• Suicide is leading cause of death for people aged 15-24
• Nearly 50% of youths (aged 12-18) with mental illness didn’t receive treatment in the previous year
Counseling
• Treatments like psychotherapy and counselingcan help!• Lots of great research on how to counsel
effectively • Typically small scale and qualitative
• Technology-mediated counseling has greatly broadened access to counseling resources• Also allows for large scale quantitative studies
3
This Talk
• What makes a good counselor?• How do you help someone feel better?
• Various techniques from NLP to discover effective conversation strategies
• Largest quantitative study of crisis counseling to date
4
The Data
5
• Collaboration with a nonprofit supporting teenagers in crisis through text messaging
• Texters matched with extensively trained volunteer counselors
• Counseling conversation via SMS• Texter gets follow-up survey
• So data includes conversation outcomes
Dataset Statistics
6
• 80,855 conversations, 3.2 million messages• 15,555 (19.2%) of conversations have survey responses• 408 counselors
• 130 counselors with over 15 conversations with survey responses
6
Split by success rate
Less successful (~50% SR) More successful (~75% SR)
Counseling “Strategies”
1. Adapt to the conversation2. Be creative in responses3. Work towards making progress4. Facilitate perspective change
7
Adaptability
• Are counselors aware of how conversations are going? How do they react?
• Compute distance between counselor language in positive/negative conversation• Represent language with TF-IDF vector of word occurrences• Cosine similarity for distance
• Observe how this changes over time
8
Adaptability
9
Time
Difference in languagebetween positive and negative conversations
Not adapting
Adapting
1. Adapt to the conversation
10
0-20 20-40 40-60 60-80 80-1003ortLon of conversDtLon (% of PessDges)
0.012
0.013
0.014
0.015
0.016
0.017
0.018
0.019
DLs
tDnc
e be
twee
n po
sLtLv
e D
nd n
egDt
Lve
conv
ersD
tLons 0ore successful counselors
Less successful counselors
Not adapting
Adapting
Creativity & Generic Responses
11
• Do counselors use generic or “templated” responses?• “How does that make you feel?” vs.
“Thanks for sharing that with me. That sounds really challenging. How do you feel about X, Kevin?”
• Measure “creativity”• Compute the number of close neighbors to each response• Cosine distance in TF-IDF space is below a threshold
Message with few neighborsMessage with many neighbors
2. Be creative in responses
12
0ore successful Less successfulCounselor qualLty
8
10
12
14
16
18
20
22
24# 6LPLlar counselor reactLons ConversatLon qualLty
PosLtLve1egatLve
Finding: More successful counselors use more creative responses than less successful counselors
Creative
Generic
How do more and less successful counselors talk differently?
• More successful counselors …• writing longer messages• use more check questions• “it sounds like…”
• use more hedges (lessen the impact of an utterance)• “maybe”, “fairly”
• avoid “why” questions
• Many more examples in the paper
13
Conversation Progress
14
• Is there a general higher-level structure to counseling conversations?
• How do counselors navigate this structure?
• Use techniques from unsupervised conversation modeling to learn ordered sequence of conversation stages
Conversation Model
15
• Assign each message in each conversation a stage using a variant of Hidden Markov Models • Force stages to be in increasing order
Conversation as sequence of text messages m1
m2
m3
m4
m5
m6
m7
Model assigns a stage to each messagem1 stage 1m2 stage 2m3 stage 2m4 stage 3m5 stage 4m6 stage 4m7 stage 5
Conversation Stages
16
Stage Interpretation Texter top words Counselor top words1 Introduction hi, hello, name, listen, hey hi, name, hello, hey, brings
2 Problem introduction
dating, moved, date, liked, ended
gosh, terrible, hurtful, painful, ago
3 Problem exploration
knows, worry, burden, teacher, group
react, cares, considered, supportive, wants
4 Problem solving write, writing, music, reading, play
hobbies, writing, activities, distract, music
5 Wrap up goodnight, bye, thank, thanks, appreciate
goodnight, 247, anytime
Conversation Stages
17
Stage Interpretation Texter top words Counselor top words1 Introduction hi, hello, name, listen, hey hi, name, hello, hey, brings
2 Problem introduction
dating, moved, date, liked, ended
gosh, terrible, hurtful, painful, ago
3 Problem exploration
knows, worry, burden, teacher, group
react, cares, considered, supportive, wants
4 Problem solving write, writing, music, reading, play
hobbies, writing, activities, distract, music
5 Wrap up goodnight, bye, thank, thanks, appreciate
goodnight, 247, anytime
1 2 3 4 5Stage
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.35
0.40
Sta
ge d
urat
Lon
(Ser
cent
of c
onve
rsat
Lon)
0ore successful counselorsLess successful counselors
3. Work towards making progress
18
1 2 3 4 5Stage
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.35
0.40
Sta
ge d
urat
Lon
(Ser
cent
of c
onve
rsat
Lon)
0ore successful counselorsLess successful counselors
3. Work towards making progress
19
• More successful counselors are quicker to know the problem and spend longer on the problem solving stage
Perspective Change
• Prior research on depression finds• Focusing on others instead of yourself can help (Campbell and
Pennebaker, 2003)• Having a positive view of the future can help (Pyszczynski et al.,1987)
• We quantify perspective change by tracking the frequency of LIWC markers (Tausczik and Pennebaker, 2010)• “I, me, myself, …” vs “he, she, they, …”• Past vs Present vs Future
20
Perspective Change: Self-Focus
21
• Texters who talk less about themselves and more about others tend to have successful conversations
0-20 20-40 40-60 60-80 80-100Portion of conversation (% of Pessages)
0.70
0.75
0.80
0.85
6el
f rel
ativ
e fre
quen
cyPositive conversations1egative conversations
Perspective Change: Future
22
0-20 20-40 40-60 60-80 80-1003ortion of Fonversation (% of Pessages)
0.060.070.080.090.100.110.120.130.140.15
)utu
re re
lativ
e fre
quen
Fy 3ositive Fonversations1egative Fonversations
0-20 20-40 40-60 60-80 80-100Portion of conversation (% of Pessages)
0.60
0.62
0.64
0.66
0.68
0.70
0.72
0.74
0.76
Pre
sent
rela
tive
frequ
ency Positive conversations
1egative conversations
0-20 20-40 40-60 60-80 80-1003ortion of conversation (% of Pessages)
0.120.140.160.180.200.220.240.260.280.30
3as
t rel
ativ
e fre
quen
cy
3ositive conversations1egative conversations
• Texters who talk less about the present and more about the future tend to have successful conversations
Past Present Future
• Simple hypothesis: The texter will talk more about something (e.g., the future) if the counselor talks about it first
• Linguistic coordination• Use coordination measure from
(Danescu-Niculescu-Mizil, 2012)
• We find significant coordination of texter towards counselor for all perspective change markers (e.g. future)• Counselor can help facilitate perspective change
23
4. Facilitate perspective change
Conclusion
24
Applying NLP techniques can teach us a lot about counseling1. Adapt to the conversation
2. Be creative in responses
3. Work towards making progress
4. Facilitate perspective change
Full study in paper: Large-Scale Analysis of Counseling Conversations: An Application of Natural Language Processing to Mental Health
Since This Research…• Project no longer active at Stanford (no data access), but…
• The counseling organization has a research fellowship program• Looking for AI/ML/NLP Experts, 3-6 months working on site
• Lots of new research on applying NLP to mental health• Especially on identifying/risk-asessing mental illness, depressions, etc.• CLPsych: Computational Linguistics and Clinical Psychology Workshop
• Also a growing number of startups in this space
25
26
Thanks!Any Questions?
New Research on the data
27
• Outside event causing increased volume
New Research on the data
28