Identifying Creative Content at the Page Level in the HathiTrust Digital Library Using Machine Learning Methods on Text and Image Features

Nikolaus Nova Parulian, Glen Worthey

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Abstract

Front-matter pages in a digitized book typically consist of mostly factual content that is not subject to copyright, and thus could potentially be opened to the public, even if the book itself is protected under copyright. However, the boundary of what is considered to be “front matter” is rather arbitrary, and some copyright-protected creative content can be found in the initial pages of a copyrighted volume. In this work, we conduct empirical research to evaluate machine learning approaches to detect creative content in the first 20 pages in a large sample of HathiTrust volumes. We start by analyzing different machine learning methods to distinguish creative from factual content using the statistically-expressed textual features from the HathiTrust Research Center’s Extracted Features dataset. From this experiment, we found that the random forest model had the best performance compared with logistic regression, support vector machine (SVM), or stochastic gradient descent (SGD) models. This experiment also reveals that textual data is not sufficient to reliably identify pages containing some kinds of creative content, e.g., images. Thus, we further trained an image detection model using YOLO-v3 to detect page types, thus creating an ensemble of textual and image features. Our findings show a promising result for the random-forest model trained on a combination of text and image features, increasing the accuracy from 85% to 89% compared with the model trained only on textual data.

Original languageEnglish (US)
Title of host publicationDiversity, Divergence, Dialogue - 16th International Conference, iConference 2021, Proceedings
EditorsKatharina Toeppe, Hui Yan, Samuel Kai Chu
PublisherSpringer
Pages478-489
Number of pages12
ISBN (Print)9783030712914
DOIs
StatePublished - 2021
Event16th International Conference on Diversity, Divergence, Dialogue, iConference 2021 - Beijing, China
Duration: Mar 17 2021Mar 31 2021

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume12645 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference16th International Conference on Diversity, Divergence, Dialogue, iConference 2021
Country/TerritoryChina
CityBeijing
Period3/17/213/31/21

Keywords

  • Copyright
  • Digital humanities
  • Digital library
  • Image processing
  • Machine learning

ASJC Scopus subject areas

  • Theoretical Computer Science
  • Computer Science(all)

Fingerprint

Dive into the research topics of 'Identifying Creative Content at the Page Level in the HathiTrust Digital Library Using Machine Learning Methods on Text and Image Features'. Together they form a unique fingerprint.

Cite this