The Labeled Segmentation of Printed Books

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 737–747 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics The Labeled Segmentation of Printed Books Lara McConnaughey Computer Science Division University of California, Berkeley laram@berkeley.edu Jennifer Dai Computer Science Division University of California, Berkeley jenniferdai@berkeley.edu David Bamman School of Information University of California, Berkeley dbamman@berkeley.edu Abstract We introduce the task of book structure la- beling: segmenting and assigning a fixed category (such as TABLE OF CONTENTS, PREFACE, INDEX) to the document struc- ture of printed books. We manually anno- tate the page-level structural categories for a large dataset totaling 294,816 pages in 1,055 books evenly sampled from 1750– 1922, and present empirical results com- paring the performance of several classes of models. The best-performing model, a bidirectional LSTM with rich features, achieves an overall accuracy of 95.8 and a class-balanced macro F-score of 71.4. 1 Introduction The availability of large-scale book corpora (such as those created by Google Books, the Inter- net Archive and the HathiTrust) has driven a flurry of work in cultural analytics over the past decade, in which the text contained in historical books has provided the raw material for the analy- sis of genre (Underwood, 2016), literary charac- ter (Bamman et al., 2014), geographic attention (Wilkens, 2013), fame (Michel et al., 2010), and much more. Books, however, are not undifferentiated streams of text in the same way that born-digital documents like tweets or emails are; they are phys- ical objects with materiality (Werner, 2012) and are arranged in a complex structure steeped in a long design tradition, with the core content of the work placed between structured frontmatter (such as a table of contents and introduction) and back- matter (such as an appendix and index). Not all of this content is desirable for all analyses; as we show below, 11% of all pages in books belong to the peritext (Genette, 1987) that surrounds the core content, with wide variability from book to book. For other analyses, such as those addressing ques- tions in book history (Kirschenbaum and Werner, 2014), isolating this structure in a consistent way across historical documents can enable research into how the form of the printed book has, for ex- ample, changed over time. While other work has focused on extracting the idiosyncratic structure inherent in each book, such as recognizing chapter boundaries in order to au- tomatically generate a table of contents, or link a parsed table of contents to positions in a book (D´ejean and Meunier, 2005, 2009; Wu et al., 2013; Gao et al., 2009), labeling document segments with a fixed typology has complementary bene- fits, allowing researchers to identify consistent cat- egories in all books regardless of the names as- signed by a specific author or publisher, or popular at a given time.1 At the same time, book structure labeling presents real challenges to automatic identifica- tion. While large-scale digital collections origi- nate in page scans of the books, the most com- mon form …

Content truncated to 500 words