By conservative estimates, several hundred thousand poems appeared in early American and U.S. newspapers from the eighteenth through the early twentieth centuries. Counting snippets of verse that appeared in death notices, advertisements, and articles makes the presence of poetry in historic newspapers even more pervasive. Feminist scholars and others performing recovery work routinely resurrect authors and works from newspaper pages, but until recently this rich trove of newspaper verse as a corpus of its own has been outside the scope of literary study and a footnote in histories of American newspapers. In the last decade, however, scholars have made significant inroads in studying the importance of newspaper verse as a form and the public role of poetry in American culture. Underpinning this scholarship is a growing recognition that the evaluation and history of American poetry should not be based on less than one percent of the poetic record. In addition, this new scholarship values and explores the role of poetry in the daily lives of people, including making sense of what it means to be human and in processing national, social, and individual experiences. To the extent that these new histories depend on traditional methods of archival discovery and analysis, however, they will remain anecdotal—individual narratives extrapolated from a miniscule subset of the whole, with limited means of situating the anecdote as either representative or idiosyncratic. In short, the magnitude of the corpus requires new modes of discovery and analysis.
A fundamental problem in the reappraisal of newspaper verse has been finding and processing poetic content in an efficient manner, which is essential for developing new interpretations, analyses, and literary histories. The primary means of finding this content typically involves paging through original issues of newspapers, scrolling through reels of microfilm, and browsing digital images to scan, by human eye, each page for graphical features that resemble poetry. Dealing with only daily newspapers for a single year, 1860, would require visually scanning nearly half a million newspaper pages. Certainly no individual in a lifetime could complete a count—to say nothing of a comprehensive bibliography or macro-level analysis—of newspaper verse using this strategy.
While the digitization of historic newspapers has mitigated some issues of access, the main avenues for discovery in these collections are browsing and text-based searching. Browsing for poetic content in such collections follows the strategy outlined above: going image by image through digitized pages and visually scanning the images for features typical of printed poetry. Ironically, web interfaces and variations in Internet connection speeds can make digitally paging through a newspaper a slower process than either scrolling through microfilm or flipping physical pages.
How, then, might one discover poetic content in digitized historic newspapers? Our research shows that digital page images hold significant promise for scholarly inquiry with regard to poetic content. The basis for our approach is that the appearance of poetic content usually follows certain patterns that can be visually differentiated from other published texts in newspapers. Given a newspaper page, a person can survey the page and figure out quickly whether it contains a poem, to a certain degree of accuracy, without having to read or understand the text. Our project has the computer do the same visual processing as the human eye and brain when a person moves through a newspaper issue looking for poetic content. This image processing approach can also be used as a powerful filter, removing materials from further consideration that do not meet the specified criteria. That is, not only does the process work to identify pages that appear to include poetry, but it discards those that do not, weeding out much of the noise.
The image processing component of the project consisted of two important phases: training and deployment. During the training phase, the goal was to produce a classifier able to categorize an image as either a poem or non-poem image. To produce the classifier, a training dataset was prepared and fed into a machine learning-based classifier. After the classifier was produced, we moved to the deployment phase. During this phase, the steps used to prepare the training sets were streamlined to automatically process and classify new images. In this paper, we focus on the four stages of the training phase: (1) pre-processing, (2) feature extraction, (2) neural networks learning, and (4) testing.
The first stage involved manual extraction of image snippets from digitized newspapers. For training, we developed three sets of snippets: (1) at least part of a poem appears in the snippet, (2) the snippet contains no poetic content, and (3) the snippet has visual cues that are similar to poetic content. Because each snippet is inherently noisy and could be of low quality, we performed 3x3, 5x5, and 7x7 averaging to smooth out noisy pixels, a step known as blurring. To convert the blurred snippet into a binary image—effectively identifying the object pixels from background pixels—we used a bi-Gaussian (or bi-normal) curve approximation (Haverkamp et al., 1995) to obtain the binary segmentation threshold. (See Figure 1 and Figure 2.)
Fig. 1: Binary non-poem image snippet; 5x5 blurring.
Fig. 2: Binary poem image snippet; 7x7 blurring.
After obtaining binary images, the next task was representing and extracting visual cues as salient features. We evaluated three sets of attributes: (1) the left and right margins (number of columns without a dark pixel); (2) the vertical white spacing between adjacent lines of text (mean and standard deviation); (3) the jaggedness of the ends-of-lines for poetic content (mean and standard deviation).
Then, with this imagery data re-represented as numeric data, we used machine learning techniques to train a classifier. Specifically, we used the machine learning approach known as artificial neural networks (ANNs) (Hopfield, 1988; Yegnanarayana, 2009). An ANN learns a vector of weights on features in the dataset to choose labels for new data. Such a network consists of multiple nodes connected to threshold functions or additional layers of nodes. The network is updated iteratively until it correctly predicts labels for training data. We used a back- propagation ANN (Hecht-Nielsen, 1989) where weights connecting the nodes update iteratively based on how the network classifies known data points in the training dataset—whether its classification matches the labels of training data points.
Back-propagation ANNs have two phases. During phase 1, a training instance or data point is fed into the ANN’s multi-layer structure, generating activations in the nodes and resulting in a final output label. During phase 2, "rewards" are propagated back to the input layer from the output layer based on whether the final output label matches the ground-truth label of the instance. If the output label is correct, all the linkages within the structure contributing to the correct output are rewarded with an increased weight. If the output label is incorrect, all contributing linkages are penalized accordingly with a reduced weight. In this manner, the network learns incrementally to find a combination of weights for these links.
Finally, to validate the accuracy of the classifier, a ten-fold cross validation process was used. In this process, the total data set was broken into ten groups. The classifier was trained using nine of the groups and then tested on the single remaining set. This process of training and testing was repeated until each group had been used as the test set once. The results were then aggregated and the accuracy computed.
In addition to discussing the importance of newspapers and newspaper verse for American literary history and the need for new modes of discovery in digitized collections, this paper will report on three results of our research: (1) basic analysis, (2) training and classification analysis, and (3) a comparative study. The basic analysis will present the algorithms we used to extract visual features from the snippets, and the correlation analyses we did to ascertain the feasibility of our image-based approach. This will show the potential discrimination power of our visual features in distinguishing poem or non-poem snippets. The training and classification analysis will document our experimentation with different ANN configurations and report on our training processes. This will include the number of hidden nodes in our ANNs used, learning weights and momentum parameters investigated, convergence rates of the different configurations, and training and classification accuracies. Finally, we will report on our comparative study in terms of the usefulness of our blurring processes and how better to fuse them. In particular, we will show whether submitting each image as a single data point with three sets of attributes or submitting each blurred image with its set of attributes as a single data point leads to more effective training and higher classification accuracy.
See, for example, Bennett (2003); McGill (2003); Barrett (2012); Barrett and Miller,(2005); Gardner (2009); Cohen (2010); Garvey (2012); Chasar (2012); Rubin (2007).
In his foundational history of American journalism, Frank Luther Mott estimated the number of U.S. newspapers in existence in 1860 at 3,000, 11 percent of which were dailies. Most dailies in this period were four pages long.
Barrett, F. and Miller, C. (2005). Words for the Hour: A New Anthology of American Civil War Poetry. Amherst, University of Massachusetts Press.
Bennett, P.B. (2003). Poets in the Public Sphere: The Emancipatory Project of American Women's Poetry, 1800-1900. Princeton, Princeton University Press.
Chasar, M. (2012). Everyday Reading: Poetry and Popular Culture in Modern America. New York, Columbia University Press.
Cohen, M. (2010). Contraband Singing: Poems and Songs in Circulation During the Civil War, American Literature, 82(2): 271-304.
Gardner, E. (2009). Unexpected Places: Relocating Nineteenth-Century African American Literature. Jackson, University Press of Mississippi.
Garvey, E. G. (2012). Writing with Scissors: American Scrapbooks from the Civil War to the Harlem Renaissance. Oxford, Oxford University Press.
Haverkamp, D., L.-K. Soh, and C. Tsatsoulis. (1995). A Comprehensive, Automated Approach to Determining Sea Ice Thickness from SAR Data, IEEE Transactions on Geoscience and Remote Sensing, 33(1): 46-57.
Hecht-Nielsen, R. (1989). Theory of the Backpropagation Neural Network, Proceedings of the International Joint Conference on Neural Network (IJCANN’1989), Washington, DC, June 1989.
Hopfield, J. J. (1988). Artificial Neural Networks, IEEE Circuits and Devices Magazine, 4(5): 3-10.
McGill, M. (2003). American Literature and the Culture of Reprinting, 1834-1853. Philadelphia: University of Pennsylvania Press.
Mott, F. L. (1942). American Journalism: A History of Newspapers in the United States through 250 Years, 1690 to 1840. New York, Macmillan.
Rubin, J. S. (2007). Songs of Ourselves: The Uses of Poetry in America. Cambridge, MA, Belknap Press.
Yegnanarayana, B. (2006). Artificial Neural Networks. New Delhi, Prentice-Hall of India.
If this content appears in violation of your intellectual property rights, or you see errors or omissions, please reach out to Scott B. Weingart to discuss removing or amending the materials.
Hosted at École Polytechnique Fédérale de Lausanne (EPFL), Université de Lausanne
July 7, 2014 - July 12, 2014
377 works by 898 authors indexed
XML available from https://github.com/elliewix/DHAnalysis (needs to replace plaintext)
Conference website: https://web.archive.org/web/20161227182033/https://dh2014.org/program/
Attendance: 750 delegates according to Nyhan 2016
Series: ADHO (9)