US Patent:
20110231384, Sep 22, 2011
Inventors:
MAKSIM KOROTEYEV - BROOKLYN NY, US
VLADIMIR KOROTEYEV - Brooklyn NY, US
OLEKSANDR PASICHNYK - Kiev, UA
Assignee:
EVTEXT, INC. - Brooklyn NY
International Classification:
G06F 17/30
G06N 3/12
G06F 15/18
US Classification:
707709, 707796, 707793, 706 12
Abstract:
The invention is a process, system, workflow system for data retrieval processes, software, Web Site, service and SaaS (Software as a Service) created to support a data retrieval process from various document types to custom or preset retrieval data structures. The program supports manual, automatic and semiautomatic data retrieval using its internal features or external add-ons. It links data points in the structure to the corresponding data points in the document, stores documents, structures and links between them and outputs results in various formats. Links between a document and a retrieval data structure are established either automatically or manually by the user. After all required links are set, results can be retrieved from the program as an XML (Extensible Markup Language) structure with required data or as a PDF (Portable Document Format) or HTML (Hypertext Format Language), in MS Office formats and others containing a/the retrieval data structure, the original document or both with links between corresponding data points.The system incorporates a Text Mining engine, which provides automatic information retrieval capabilities. The engine implements Text mining technology that is based on Evolutionary Bayesian Ontology Classification. This technology uses Bayesian Ontology for modeling the problem's domain and applies Evolutionary Search for the most plausible classification decision.The ability to learn from data is a key feature of Bayesian Ontology, and for our embodiment. The complexity and size of semantic and format dependencies between elements in a natural language text is too high for analytical descriptions. Plus, we intend to save the user the trouble of building their own data retrieval models. Instead, we rely on an algorithm that automatically links user's data selections to the closest categories in pre-built ontologies and generates selection specific classifiers. Every individual ontology keeps learning from user corrections during its life cycle. The system is specifically built with the ability to accumulate data models learned from various types of documents. The more documents have been processed by the system, the higher generalization capabilities it possesses for automatic processing of new, unseen documents.