Learning Document Representations for Ranked Retrieval in the Legal Domain

Abstract

In this work I detail the compilation of a unique corpus of Norwegian court decisions. I utilize this corpus to train several different machine learning models to produce dense semantic vectors for both words and documents. I use the document vectors to perform ranked document retrieval, and evaluate and demonstrate the performance of the vectors for this task using a purposely-built ranked retrieval model utilizing the document references in the documents. Furthermore, I explore the interplay between pre-trained semantic word vectors and convolutional neural networks and conduct several hyperparameter optimization experiments using convolutional neural networks to produce document vectors.