Financial narrative summarization

Abstract

The aim of this thesis is to introduce and evaluate different techniques for the summarization of Financial Documents. These techniques included the use of Kmeans and DBSCAN algorithms for the selection of the most important sentences while ensuring that the developed pipeline produces summaries that are representative of the whole document, is Domain - agnostic and can generalize also on documents that do not have a Table of Contents. The produced summaries were evaluated using the Rouge 2 F1 score and compared with the submissions of the FNS 2023 challenge. The approach using Kmenas algorithm achieved a higher Rouge 2 F1 score compared to three out of seven submissions having an equal score with the fourth-best submission while the one using the DBSCAN achieved a higher Rouge 2 F1 score compared to three out of the seven submissions.