Bachelor's Thesis Tobias Eyl

An Approach for a Semantic Information Extraction of Decisions of the Federal Court of Justice

Abstract

While the amount of documents is also continuously growing in the legal sector, there are lacking possibilities for effectively using these resources compared to other sectors like finance sector. Most legal documents like court decisions, legal literature or the law texts itself are mostly published in plain text with little or without any additional metadata that might enable a more efficient usage.

Within the scope of this work, an approach is developed which extracts the legal parties and their legal relations among them and finally displays the extracted data in a graph-like form enabling the legal professional to conduct a more efficient research. In order to achieve this, at the beginning, a linguistic analysis will be performed to elicitate judgment specific linguistic features and subsequently to build a set of legal keywords indicating certain legal relations like a specific contractual agreement between two parties. Following, an ontology representing all the required semantic information within the sentences containing these keywords is built. In order to do this, a broad literature research is conducted and its results will be analyzed in the next. The developed ontology will then be implemented on the basis of a NLP-Technique called Dependency Parsing. For this, a model for spaCy’s neuronal-network based dependency parser is trained which subsequently is applied to the respective section of judgments. On the basis of the semantic dependency model, extraction rules for every defined legal relation are implemented to enable the extraction of the information once the annotations has been set correctly. Finally, a visual representation will be implemented providing a well-arranged overview of the extracted semantic information.

The results of the evaluation show that this approach delivers remarkable high precision results despite being based on a relatively small set of training data with 38 training sentences and 25 sentences for evaluation.

Research questions

1. How an ontology for semantic information of court decisions can look like?
2. How the key information of a court decision can automatically be extracted using NLP?
3. How a prototype for a semantic analysis of court decisions can be implemented?