Towards Learning from User Feedback for Ontology-based Information Extraction

Kurzfassung

Many engineering projects involve the integration of various hardware parts from different suppliers. In preparation, parts that are best suited for the project requirements have to be selected. Information on these parts' characteristics is published in so called data sheets usually only available in textual form, e.g. as PDF files. To realize the automated processing, these characteristics have to be extracted into a machine-interpretable format. Such a process requires a lot of manual intervention and is prone to errors. Domain ontologies, among other approaches, can be used to implement the automated information extraction from the data sheets. However, ontologies rely solely on the experiences and perspectives of their creators at the time of creation. To automate the evolution of ontologies, we developed ConTrOn - Continuously Trained Ontology - that automatically extracts information from data sheets to augment an ontology created by domain experts. The evaluation results of ConTrOn show that the enriched ontology can help improve the information extraction from technical documents. Nonetheless, the extracted information should be reviewed by experts before using it in the integration process. We want to provide an intuitive way of reviewing, in which the extracted information will be highlighted on the data sheets. The experts will be able to accept, reject, or correct the extracted data via a graphical interface. This process of revision and correction can be leveraged by the system to improve itself: learning from its own mistakes and identifying common patterns to adapt in the next extraction iteration. This paper presents ideas how to use machine learning based on user feedback to improve the information extraction process.