tailieunhanh - Báo cáo khoa học: "Reversing Controlled Document Authoring to Normalize Documents"

This paper introduces document normalization, and addresses the issue of whether controlled document authoring systems can be used in a reverse mode to normalize legacy documents. A paradigm for deep content analysis using such a system is proposed, and an architecture for a document normalization system is described. | Reversing Controlled Document Authoring to Normalize Documents Aurélien Max Groupe d Etude pour la Traduction Automatique GETA Xerox Research Centre Europe XRCE Grenoble France Abstract This paper introduces document normalization and addresses the issue of whether controlled document authoring systems can be used in a reverse mode to normalize legacy documents. A paradigm for deep content analysis using such a system is proposed and an architecture for a document normalization system is described. 1 Introduction Controlled Document Authoring is a field of research in NLP that is concerned with the interactive production of documents in limited domains. The aim of systems implementing controlled document authoring is to allow the user to specify an underlying semantic representation of the document that is well-formed and complete relative to its class of documents. This representation is then used to produce a fully controlled version of the document possibly in several languages. We distinguish controlled document authoring systems from what is referred to in Reiter and Dale 2000 as computer as authoring aid which are Natural Language Generation systems intended to produce initial drafts or only routine factual sections of documents in that the former can be used to produce high-quality final versions of documents without the need for further hand-editing. The question which motivated our work was the following can we reuse the resources of an existing controlled document authoring system to analyze documents from the same class of documents If so we could obtain the semantic structure corresponding to a raw document and then produce from it a completely controlled version. If the raw document is bigger in scope from the documents that the authoring system models then something similar to document summarization by content recognition and reformulation would be done. Incomplete representations after automatic analysis could be interactively .