Subword and Spatiotemporal Models for Identifying Actionable Information in Haitian Kreyol

Proceedings of the Fifteenth Conference on Computational Natural Language Learning, pages 68–77, Portland, Oregon, USA, 23–24 June 2011. c⃝2011 Association for Computational Linguistics Subword and spatiotemporal models for identifying actionable information in Haitian Kreyol Robert Munro Department of Linguistics Stanford University Stanford, CA, 94305 rmunro@stanford.edu 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 agen- cies can actually act on, ‘actionable’ informa- tion, will be sparse so there is great poten- tial to (semi)automatically identify actionable communications. However, there are hurdles as the languages spoken will often be under- resourced, 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 earth- quake, predominantly in Haitian Kreyol. We show that keyword/ngram-based models us- ing streaming MaxEnt achieve up to F=0.21 accuracy. Further, we find current state-of- the-art subword models increase this substan- tially to F=0.33 accuracy, while modeling the spatial, temporal, topic and source contexts of the messages can increase this to a very ac- curate F=0.86 over direct text messages and F=0.90-0.97 over social media, making it a vi- able strategy for message prioritization. 1 Introduction The recent proliferation of cellphone technologies has resulted in rapid increases in the volume of in- formation coming out of crisis-affected regions. In the wake of the January 12, 2010 earthquake in Haiti local emergency response services were inop- erable but 70-80% of cell-towers were quickly re- stored. With 83%/67% of men/women possessing a cellphone, the nation remained largely connected. People within Haiti were texting, calling and inter- acting with social media, but aid agencies were not equipped to process so much information. This is because in any sudden onset crisis, this flood of in- formation out of the region coincides with crisis- response organizations already working at capacity as they move in. The problem definition is clear: how can we filter this information to identify action- able intelligence to support relief efforts? The solution is complicated. There may be few resources for the language(s) of the crisis- affected population and the ratio of actionable to non-actionable information will often be very large, especially when reported through social-media and other non-official channels. In the absence of ex- isting electronic resources models must be built on- the-fly and account for substantial spelling varia- tions. The definition of what constitutes action- able intelligence will often be context-specific and changing, too. In the data used here ‘actionable’ changed quickly from Search and Rescue to any medical emergency and then to include clusters of requests for food and water. The models will there- fore need to be time-sensitive or otherwise adaptive. The system presented here attempts to address these problems, finding that the accurate identifica- tion of actionable information can be achieved with subword models, the automatic identification of top- …

Content truncated to 500 words