Inferring Ontology Fragments From Semantic Role Typing Of Lexical Variants

[Context and Motivation] Information systems depend on personal data to individualize services. To manage privacy expectations, companies use privacy policies to regulate what data is collected, used and shared. However, different terminological interpretations can lead to privacy violations, or misunderstandings about what behavior is to be expected. [Question/Problem] A formal ontology can help requirements authors to consistently check how their data practice descriptions relate to one another and to identify unintended interpretations. Constructing an empirically valid ontology is a challenging task since it should be both scalable and consistent with multi-stakeholder interpretations. [Principle Ideas/Results] In this paper, we introduce a semi-automated semantic analysis method to identify ontology fragments by inferring hypernym, meronym and synonym relationships from morphological variations. The method employs a shallow typology to categorize individual words, which are then matched automatically to 26 reusable semantic rules. The rules were discovered by classifying 335 unique information type phrases extracted from 50 mobile privacy policies. The method was evaluated on 109 unique information types extracted from six privacy policies by comparing the generated ontology fragments against human interpretations of phrase pairs obtained by surveying human subjects. The results reveal that the method scales by reducing the number of otherwise manual paired comparisons by 74% and produces correct fragments with a 1.00 precision and 0.59 recall when compared to human interpretation. [Contributions] The proposed rules identify semantic relations between a given lexeme and its morphological variants to create a shared meaning between phrases among end users.