Expert system rule extraction from legal documents

Abstract

Working in the legal domain requires collecting a substantial amount of information from diverse sources and reasoning about the implicit knowledge stored within such information to make informed decisions. Manually extracting the needed information from unstructured sources is a tedious and time consuming process. This overhead can be significantly reduced if legal knowledge were stored in a structured and machine-readable manner. The system presented in this project processes legal documents which are initially in the unstructured form of natural language, extracts legal knowledge from such documents, and stores the newly discovered knowledge in a structured manner. The system was built as a series of independent but inter-connected modules in such a manner that at each stage in the execution process every module builds upon the work of previous modules by discovering extra knowledge regarding the text. Extracting legal knowledge from unstructured sources involves several steps. First of all, the system splits the input document into a set of legal rules in such a manner that each rule can be processed independently from the rest and makes use of Natural Language Processing techniques to expose and make explicit the linguistic structure of each rule. Each rule is then passed through a classification algorithm which assigns a legal provision type to each rule from a set of predefined provision types defined by an ontology which the system uses to represent knowledge. The system then identifies and extracts fragments of text that are used to fill the specific slots in an ontology that correspond to the rule's assigned provision type. The knowledge is then stored in the form of a relational database to be readily available for future use. The system performs well and achieves results that are up to standard with what is expected from modern Natural Language Processing applications. When evaluated on a set of legal rules that were manually annotated by human experts in the legal domain the system achieves levels of 98% of recall and precision in the classification task and levels of 93% of precision and 94% of recall in the knowledge extraction task.