Relevance Feedback Search Based on Automatic Annotation and Classification of Texts

The idea behind Relevance Feedback Search (RFBS) is to build search queries as an iterative and interactive process in which they are gradually refined based on the results of the previous search round. This can be helpful in situations where the end user cannot easily formulate their information needs at the outset as a well-focused query, or more generally as a way to filter and focus search results. This paper concerns (1) a framework that integrates keyword extraction and unsupervised classification into the RFBS paradigm and (2) the application of this framework to the legal domain as a use case. We focus on the Natural Language Processing (NLP) methods underlying the framework and application, where an automatic annotation tool is used for extracting document keywords as ontology concepts, which are then transformed into word embeddings to form vectorial representations of the texts. An unsupervised classification system that employs similar techniques is also used in order to classify the documents into broad thematic classes. This classification functionality is evaluated using two different datasets. As the use case, we describe an application perspective in the semantic portal LawSampo – Finnish Legislation and Case Law on the Semantic Web . This online demonstrator uses a dataset of 82 145 sections in 3725 statutes of Finnish legislation and another dataset that comprises 13 470 court decisions.