Paul Johannesson; Mong Li Lee; Liddle, S.; Opdahl, A.; Pastor López, O. (2017). Special issue on conceptual modeling - 34th International Conference on Conceptual Modeling (ER 2015). Data & Knowledge Engineering. 109:1-2. doi:10.1016/j.datak.2017.03.001

A considerable amount of clean semistructured data is internally available to companies in the form of business reports. However, business reports are untapped for data mining, data warehousing, and querying because they are not in relational form. Business reports have a regular structure that can be reconstructed. We present algorithms that automatically infer the regular structure underlying business reports and automatically generate wrappers to extract relational data.

We can extract and structure information from documents if we can match attributes with document data values and associate these matched attribute-value pairs as tuples in relations. In this paper we present a general approach to extracting and structuring information from unstructured documents that are data rich (have many recognizable constants). In our approach to this problem we start with an application ontology that describes the objects, relationships, and constraints in a domain of interest. We parse this ontology to generate recognition rules for constants and context keywords and to extract structural and constraint information. Given the generated rules and an unstructured document, we apply a recognizer to extract the constants and keywords, and we then apply a structure builder to match constant values with attributes, to associate attribute-value pairs as relations, and to populate a generated database schema with the extracted data according to the constraints of the application ontology. When applied to a list of several similar unstructured documents, the result is a populated database structured according to and ltered with respect to the application ontology. To make our approach general, we x all the processes and change only the ontological description for a di erent application domain. In experiments we conducted on two di erent types of unstructured documents taken from the Web, our approach attained recall ratios in the 80 % and 90 % range and precision ratios near 98%.