Page 1
Subword and spatiotemporal models
for identifying actionable information
in Haitian Kreyol
Robert Munro
Stanford University
CoNLL 2011
Munro, Robert. "Subword and spatiotemporal models for identifying actionable
information in Haitian Kreyol." Proceedings of the Fifteenth Conference on
Computational Natural Language Learning. Association for Computational Linguistics,
2011.
http://www.robertmunro.com/research/munro11kreyol.pdf
Page 3
Messages start streaming in
Page 4
Messages start streaming in
Page 11
Crowdsourced (Mission 4636)
Page 12
Feedback
US Marines
◦ “Saving lives every day.”
FEMA:
◦ “The most comprehensive and up-to-date
map available to the humanitarian
community.”
World Food Program
◦ “We delivered food to an informal camp of
2500 people that you identified for us.”
Page 13
Sudden-onset language
processing
Page 14
Sudden-onset language
processing
Page 15
Sudden-onset language
processing
Page 16
Sudden-onset language
processing
Page 17
Prioritization
Only 2% of messages were
‘actionable’
◦ An identifiable location
◦ Medical, S+R, water, clustered food
requests, security, unaccompanied
children.
How can we prioritize the
actionable items in the original
Haitian Kreyol?
Can we leverage the models for
more-sparse information sources?
Page 18
Evaluation data
Mission 4636. 40,811 text-messages
sent to a free number, ‘4636’, in Haiti.
(predominantly in Haitian Kreyol, with
translations, UN-defined categories,
and geolocation)
Radio Station. 7,528 text-messages
sent to a Haitian radio station.
Twitter. 63,195 Haiti-related tweets.
Page 19
Variation is the norm
mesi mèsi mèci meci merci
Kreyol French
Abbrv. Full Form Pattern Meaning
s’on se yon sVn is a
av`en av`eknou VvVn with us
relem rele mwen relem call me
wap ouap uVp you are
map mwen ap map I will be
zanmimzanmi mwen zanmim my friend
lavel lave li lavel to wash (it)
Page 21
Combining spatio-temporal
models
Page 22
Streaming architecture
Build from initial items
timeModel
Page 23
Streaming architecture
Predict (and evaluate) on incoming
items
◦ (penalty for training)time
Model
Page 24
Streaming architecture
Repeat / retrain
timeModel
Page 25
Streaming architecture
Repeat / retrain
timeModel
Page 26
Streaming architecture
Repeat / retrain
timeModel
Page 27
Streaming architecture
Repeat / retrain
timeModel
Page 28
Streaming architecture
Repeat / retrain
timeModel
Page 29
Features
G : Words and ngrams
W : Subword patterns
P : Source of the message
T : Time received
C : Categories (c0,...,47)
L : Location (longitude and latitude)
L : Has-location (a location is written
in the message)
Page 30
Features
Subword models
◦ Full-forms and normalizations (Munro and
Manning 2010)
Abbrv. Full Form Pattern Meaning
s’on se yon sVn is a
av`en av`eknou VvVn with us
relem rele mwen relem call me
wap ouap uVp you are
map mwen ap map I will be
zanmimzanmi mwen zanmim my friend
lavel lave li lavel to wash (it)
Page 31
Time / Space / source
Timestamp (in place of discounting)
Phone-number of sender
Spatial tile-membership:
Page 32
Additional streaming models
Message contains an identifiable
location
timeModel
Page 33
Additional streaming models
Message contains an identifiable
location
Prediction for 47 categories
timeModel
timeModel
timeModel
timeModel
…
Page 34
Final streaming model
Prediction for ‘is actionable’
timeModel
Page 35
Final streaming model
Prediction for ‘is actionable’
timeModel
timeModel
timeModel
timeModel
…
Combines features with predictions
from Category and Has-Location
models
Page 36
Evaluation
100 training epochs
◦ Calculated on predictions over epochs 2-
100
Comparison of two-tier architecture
with Oracle ‘has-location’ and
‘Categories’
Identification of actionable messages
in Radio Station and Twitter messages
(full results in paper)
Page 37
Feature-based improvements
Subwords and Source (G, T,W, P)
Temporal feature (G, T)
Words/Ngrams (G)
0.326
0.252
0.207
Page 38
Combined models
All Features/Models
Spatial clusters (L)
Words/Ngrams (G)
0.855
0.756
0.207
Page 39
Outperforming the oracle
Location (two-teir prediction)
Location (oracle)
Words/Ngrams (G)
0.310
0.274
0.207
Page 40
Negative results from filtering
Oracle true-neg filtering
All Features/Models
Words/Ngrams (G)
0.428
0.855
0.207
Page 41
Social media
Twitter
Radio station
All Features/Models
0.969
0.904
0.855
Page 42
Conclusions (usability)
Subword and spatio-temporal models
can give a 10-fold increase in
prioritization
Adding multi-tiered streaming models
can give a 50-fold increase in
prioritization
Cross-domain adaptation is possible
for need(le)-in-haystack information
extraction from social media
Page 44
Appendix: abstract
Crisis-affected populations are often able to maintain digital
communications but in a sudden-onset crisis any aid organizations
will have the least free resources to process such communications.
Information that aid agencies can actually act on, ‘actionable’
information, will be sparse so there is great potential to
(semi)automatically identify actionable communications. However,
there are hurdles as the languages spoken will often be
underresourced, have orthographic variation, and the precise
definition of ‘actionable’ will be response-specific and evolving.
We present a novel system that addresses this, drawing on 40,000
emergency text messages sent in Haiti following the January 12,
2010 earthquake, predominantly in Haitian Kreyol. We show that
keyword/ngram-based models using streaming MaxEnt achieve up to
F=0.21 accuracy. Further, we find current state-of-the-art subword
models increase this substantially to F=0.33 accuracy, while
modeling the spatial, temporal, topic and source contexts of the
messages can increase this to a very accurate F=0.86 over direct
text messages and F=0.90-0.97 over social media, making it a viable
strategy for message prioritization.