Spatial Information Extraction from Short Messages

Texts in addition to maps and satellite images, have become an important spatial data resource in recent years. Electronic written texts used in mediated interactions, especially short messages, have triggered the emergence of new ways of writing. Extracting information from such short messages, which represent a rich source of information, is highly important in order to discover domain-relevant information in the text and facilitate information retrieval. However, short messages are hard to analyse because of their brief, unstructured and informal nature. This paper focuses on the kinds of special or unique spatial entities and relations are contained in short messages. A new entity extraction method specifically dedicated to French short messages (SMS and tweets) is outlined to address this issue. The method is then tested on more traditional sources, like newspaper texts. This work is crucial in order to take advantage of the vast amount of geographical knowledge expressed in heterogeneous unstructured data. Firstly, we propose a process in which new spatial entities are extracted (e.g. motpellier, montpelier, Montpel are associated with Montpellier). Secondly, we identify new spatial relations that precede spatial entities (e.g. sur, par). Finally, we propose general patterns for the extraction of spatial relations. The task is very challenging and complex due to the specificity of short message language, which is based on weakly standardized modes of writing. The experiments were carried out on the three French corpora (i.e. 88milSMS, tweets, and Midi Libre) and highlight the efficiency of our proposal for identifying new kinds of spatial entities and relations.