Towards Deep Learning from Twitter for Improved Tsunami Alerts and Advisories

Data from social-networking services increasingly complements that from traditional sources in scenarios that seek to 'cultivate' situational awareness. As false-positive alerts and retracted advisories appear to suggest, establishing a causal connection between earthquakes and tsunamis remains an extant challenge that could prove life-critical. Because posts regarding such natural disasters typically 'trend' in real time via social media, we extract tweets in an effort to elucidate this cause-effect relationship from a very different perspective. To extract content of potential geophysical value from a multiplicity of 140-character tweets streamed in real time, we apply Natural Language Processing (NLP) to the unstructured data and metadata available via Twitter. In Deep Learning from Twitter, words such as "earthquake" are represented as vectors embedded in a corpora of tweets, whose proximity to words such as "tsunami" can be subsequently quantified. Furthermore, when use is made of pre-trained word vectors available for various reference corpora, geophysically credible tweets are rendered distinguishable by quantifying similarities through use of a word-vector dot product. Finally, word-vector analogies are shown to be promising in terms of deconstructing the earthquake-tsunami relationship in terms of the cumulative effect of multiple, contributing factors (see figure). Because diction is anticipated to differ in tweets that follow a tsunami-producing earthquake, our emphasis here is on the re-analysis of actual event data extracted from Twitter that quantifies word sense relative to earthquake-only events. If proven viable, our approach could complement those measures already in place to deliver real-time alerts and advisories following tsunami-causing earthquakes. With climate change accelerating the frequency of glacial calving, and in so doing providing an alternate, potential source for tsunamis, our approach is anticipated to be of value in broader contexts.