towards a data-driven approach to identify crisis-related topics in social media streams
TRANSCRIPT
Towards a Data-driven Approach to Identify Crisis-Related Topics in Social Media Streams
Muhammad Imran (@mimran15) and Carlos Castillo (@ChaToX)
Qatar Computing Research Institute
Doha, Qatar.
SWDM’15 : WWW’15 May 18th 2015
Information Variability on Social Media
• Different events present different information categories
• Even for recurring events, categories proportion change
Information Variability on Social Media
• Different events present different information categories
• Even for recurring events, categories proportion change
Information Variability on Social Media
• Different events present different information categories
• Even for recurring events, categories proportion change
Information Variability on Social Media
• Different events present different information categories
• Even for recurring events, categories proportion change
Information Variability on Social Media
• Different events present different information categories
• Even for recurring events, categories proportion change
Different Classification Approaches
• Various classification approaches exist:– Manual classification by human experts– Automatic classification using unsupervised or
supervised approaches(needs training data)– Hybrid: Automatic + Manual
• Retrospective vs. real-time classification– Batch processing (offline, training data availability)– Stream processing (real-time, scarce training data)
Real-time Stream Classification (Supervised )
• Fewer categories are better– Decrease workers dropout – More training data for each category, more accuracy– “7 plus/minus 2” rule [G. A. Miller, 56]
• Categories need to be defined carefully– Empty categories (waste space and efforts of workers)– Categories that are too large introduce heterogeneity
Problem Statement
• How can we classify items arriving as a data stream into a small number of categories, if we cannot anticipate exactly which will be the most frequent categories?
Our research improves crowdsourcing-based and supervised learning-based systems (e.g. AIDR) by finding latent categories in fast data streams.
Our Approach (top-down + bottom-up)
1. An expert defines information categories (top-down)2. Messages are categorized into the initial set plus an
extra “Miscellaneous” category3. Identify relevant and prevalent categories from the
messages in the “Miscellaneous” category (bottom-up)
1. Generate candidate categories2. Learn characteristics of good categories3. Rank categories on good characteristics
How do we identify relevant categories?
Candidate Generation
We propose to apply Latent Dirichlet Allocation (LDA) on the Miscellaneous category:• Input: A set of n documents (all messages in
the Misc. category) and a number m (# of topics to be generated)
• Output: n x m matrix in which cell(i, j) indicates the extent to which document i corresponds to topic j.
Candidate Evaluation
To reduce the workload of experts to decide which categories to pick or not, we propose the following criteria:• Volume: a category shouldn’t be too small• Novelty: a category must not overlap or be too
similar to the existing categories• Cohesiveness (intra- and inter-similarity): a
category should be cohesive (should have small intra-topic and large inter-topic values)
Experimental Testing• We used Twitter data of 17 crises (from the
CrisisLexT26 dataset at crisislex.org)
A. Affected individuals, deaths, injuries, missing, found.
B. Infrastructure and utilities: buildings, roads, services damage.
C. Donation and volunteering: needs, requests of food, shelter, supplies.
D. Caution and advice: warnings issued or lifted, guidance and tips.
E. Sympathy and emotional support: thoughts, prayers, gratitude, etc.
Z. Other useful information not covered by any of the above categories.
Candidate Generation Setup
• Applied LDA on the messages in the “Z” category of each crisis
• 5 topics were generated for each crisis• Considered messages with LDA score > 0.06 in
each topic• Presented the LDA generated topics to experts
in random order
Candidate Annotation Setup
Recruited two experts from two Int. humanitarian organizations in the crisis response domain
Results• Topics with avg. score <= 2.5 considered as bad topics• Topics with avg. score >= 3.5 considered as good topics• Hit: if the metric value of good topics > bad topics
A crisis is not considered for evaluation, if all of its topics receive an average score either below or above 3.0.
Conclusion
• Novelty, intra-similarity and cohesiveness are useful in identifying good topics
• Our approach combines top-down (manual) and bottom-up (automatic) elements.
• Learned important characteristics of good topics
• Future work includes candidate ranking including recommendation for adding, merging, dropping new unseen categories
Data used in this study can be requested:Contact: Muhammad Imran at
[email protected] OR @mimran15
Thank you!
Authors contact:Muhammad Imran @mimran15Carlos Castillo @ChaToX