A proposal for an architecture to extract information from sms messages during emergency situations

Abstract

In mass emergencies, a fair amount of information is exchanged via SMS messages. These messages tend to be informal and to contain abbreviations and misspellings, which makes them difficult to treat. This is a problem for current Information Extraction tools, especially for messages in Portuguese. This work proposes an architecture to extract information from SMS messages during emergencies. The architecture comprises four components: Linguistic Processing, Temporal Processing, Event Processing, and Information Fusion. We also defined an SMS corpus building process. From the proposal of this architecture, we conducted a case study, which included building BraCorpSMS, a corpus of SMS messages received by an electric utility company. We built a prototype in Python using NLTK to validate the architecture. The prototype had its Information Extraction components evaluated achieving Precision of 88%, Recall of 59% and balanced F-measure of 71%. The results indicate improvement opportunities, but as this is the first work for Portuguese facing processing SMS messages during emergency situations, it also serves as a roadmap for future work in the area.