Information extraction from non-segmented text (on the material of weather forecast telegrams)

Both the domain and sublanguage specific approach to text analysis and information extraction is proposed. Texts under consideration are weather forecast telegrams written in Russian. Telegrams are an example of deviant text type, with lack of text segmentation means, a lot of abbreviations, syntactic and spelling mistakes. The presented work pursues the problem of text segmentation: a procedure for the recovery of text structure is proposed that results in a sequence of topically coherent text fragments suitable for semantic interpretation. Topical mechanisms combined with narrative structure analysis allow disambiguation of circumstantial (locative and temporal) modifiers attachment.