Open Information Extraction Using Constraints over Part-of-speech Sequences

In 2010’s several exabytes of data are produced daily. Approximately between 1⁄5 and 1⁄3 of these data is text. To make use of such huge amounts of textual data, we need to be able to detect, to extract, to structure, and to process important information conveyed through this data flow in a fast and scalable manner. Open information extraction (Open IE) is a solution for detection, extraction, and initial structuring of information. Open IE is open-domain and relation-independent paradigm for information extraction performed in an unsupervised manner. This makes high-speed performance and scalability its main advantages, converting Open IE into a very perspective field for research and for applications to other text data processing tasks. In this work we have conducted an extensive research on various methods for Open IE and on its application to other tasks, and contributed in several ways. First, we have introduced an Open IE method requiring minimal pre-processing of input that assures its speed and robustness. Additionally, we proposed this method for Spanish language and showed it to be superior to other methods at least in one of the two aspects: either in terms of precision or in terms of robustness. As an additional contribution, we also introduced a method for performance comparison of Open IE systems implemented for different languages by comparing outputs for parallel datasets. Next, we introduced a method for Open IE with additional semantic pre-processing that allows semantic interpretation of the extracted relations, which was not possible with other Open IE methods. We showed that this method has a very high precision, although at a cost of yield. Most importantly, we demonstrated the application of this method to semantic structuring of extractions by introducing a novel procedure of extraction presentation in RDF/XML format that is a standard format maintained by W3C. Further, we showed that Open IE can serve for measuring of Web document informativeness that is one of the aspects of document quality. Not only did we show that it can be applicable to the arbitrary domain documents extracted from the Web “as is”, without additional pre-processing, we also showed that Open IE can serve for a complex text processing task that has a direct impact on an end-user. And the last but not least, we made publicly available the software and evaluation resources developed as parts of this work.