Extracting Temporal Information from Short Messages

Abstract

Information Extraction, the process of eliciting data from natural language documents, usually relies on the ability to parse the document and then to detect the meaning of the sentences by exploiting the syntactic structures encountered. In previous papers, we have discussed an application to extract information from short (e-mail and text) messages which takes an alternative approach. The application is lightweight and uses pattern matching rather than parsing, since parsing is not feasible for messages in which both the syntax and the spelling are unreliable. The application works in the context of a high level database schema and identifies sentences which make statements about data describable by this schema. The application matches sentences with templates to identify metadata terms and the data values associated with them. However, the initial prototype could only manage simple, time independent assertions about the data, such as "Jane Austen is the author." This paper describes an extension to the application which can extract temporal data, both time instants and time periods. It also manages time stamps - temporal information which partitions the values of time varying attributes, such as the monarch of a country. In order to achieve this, the original data model has had to be extended with a temporal component and a set of sentence templates has been constructed to recognise statements in this model. The paper describes the temporal model and the extensions to the application, concluding with a worked example.