Automatic office document classification and information extraction

TEXPR.OS (TEXt PROcessing System) is a document processing system (DPS) to support and assist office workers in their daily work in dealing with information and document management. In this thesis, document classification and information extraction, which are two of the major functional capabilities in TEXPROS, are investigated. Based on the nature of its content, a document is divided into structured and unstructured (i.e., of free text) parts. The conceptual and content structures are introduced to capture the semantics of the structured and unstructured part of the document respectively. The document is classified and information is extracted based on the analyses of conceptual and content structures. In our approach, the layout structure of a document is used to assist the analyses of the conceptual and content structures of the document. By nested segmentation of a document, the layout structure of the document is represented by an ordered labeled tree structure, called Layout Structure Tree (L-S-Tree). Sample-based classification mechanism is adopted in our approach for classifying the documents. A set of pre-classified documents are stored in a document sample base in the form of sample trees. In the layout analysis, an approximate tree matching is used to match the L-S-Tree of a document to be classified against the sample trees. The layout similarities between the document and the sample documents are evaluated based on the edit distance between the L-S-Tree of the document and the sample trees. The document samples which have the similar layout structure to the document are chosen to be used for the conceptual analysis of the document. In the conceptual analysis of the document, based on the mapping between the document and document samples, which was found during the layout analysis, the conceptual similarities between the document and the sample documents are evaluated based on the degree of conceptual closeness degree . The document sample which has the similar conceptual structure to the document is chosen to be used for extracting information. Extracting the information of the structured part of the document is based on the layout locations of key terms appearing in the document and string pattern matching. Based on the information extracted from the structured part of the document the type of the document is identified. In the content analysis of the document, the bottom-up and top-down analyses on the free text are combined to extract information from the unstructured part of the document. In the bottom-up analysis, the sentences of the free text are classified into those which are relevant or irrelevant to the extraction. The sentence classification is based on the semantical relationship between the phrases in the sentences and the attribute names in the corresponding content structure by consulting the thesaurus. Then the thematic roles of the phrases in each relevant sentence are identified based on the syntactic analysis and heuristic thematic analysis. In the top-down analysis, the appropriate content structure is identified based on the document type identified in the conceptual analysis. Then the information is extracted from the unstructured part of the document by evaluating the restrictions specified in the corresponding content structure based on the result of bottom-up analysis. The information extracted from the structured and unstructured parts of the document are stored in the form of a frame like structure (frame instance) in the data base for information retrieval in TEXPROS