Thesis
French
ID: <
10670/1.2bw41h>
Abstract
This thesis focuses on entity recognition in documents recognized by OCR, driven by a database. An entity is a homogeneous group of attributes such as an enterprise in a business form described by the name, the address, the contact numbers, etc. or meta-data of a scientific paper representing the title, the authors and their affiliation, etc. Given a database which describes entities by its records and a document which contains one or more entities from this database, we are looking to identify entities in the document using the database. This work is motivated by an industrial application which aims to automate the image document processing, arriving in a continuous stream. We addressed this problem as a matching issue between the document and the database contents. The difficulties of this task are due to the variability of the entity attributes representation in the database and in the document and to the presence of similar attributes in different entities. Added to this are the record redundancy and typing errors in the database, and the alteration of the structure and the content of the document, caused by OCR. To deal with these problems, we opted for a two-step approach: entity resolution and entity recognition. The first step is to link the records referring to the same entity and to synthesize them in an entity model. For this purpose, we proposed a supervised approach based on a combination of several similarity measures between attributes. These measures tolerate character mistakes and take into account the word permutation. The second step aims to match the entities mentioned in documents with the resulting entity model. We proceeded by two different ways, one uses the content matching and the other integrates the structure matching. For the content matching, we proposed two methods: M-EROCS and ERBL. M-EROCS, an improvement / adaptation of a state of the art method, is to match OCR blocks with the entity model based on a score that tolerates the OCR errors and the attribute variability. ERBL is to label the document with the entity attributes and to group these labels into entities. The structure matching is to exploit the structural relationships between the entity labels to correct the mislabeling. The proposed method, called G-ELSE, is based on local structure graph matching with a structural model which is learned for this purpose. This thesis being carried out in collaboration with the ITESOFT-Yooz society, we have experimented all the proposed steps on two administrative corpuses and a third one extracted from the web