A Visual Interface for Exploring Language Use in Slave Narratives

A Visual Interface for Exploring Language Use in Slave Narratives
Muralidharan, Aditi, Department of Computer Science UC Berkeley, aditi@cs.berkeley.edu
Abstract
The increasing prevalence of digitized source material in the humanities has led to uncertainty about how this suddenly available information will change scholars' research methods. What balance will scholars strike between in-depth examination of a few sources, and a more "distant reading" (Moretti 2005) of a large number of them? Our focus is specifically on text collections: comparing texts, and identifying and tracing patterns of language use. These tasks are not widely supported by any current software, but if humanities researchers want to use digitized text collections on a larger scale, they will need to do exactly such things.

We restrict ourselves to a particular collection: the North American antebellum slave narratives, written by fugitive slaves in the decades before the Civil War with the support of abolitionist sponsors. Scholars agree about the slave narrative's most basic conventions but it is likely that these narratives, with their extreme repetitiveness, may also manifest other regular features that have yet to be detected by scholars. This project aims to assist literary scholars in uncovering these patterns with computational techniques.

In collaboration with English scholars, we have built WordSeer ( http://bebop.berkeley.edu/wordseer), a system that can compare two or more narratives' grammatical features, and analyze the distribution of textual patterns throughout an entire collection. Our goal is for English scholars to be able to use our system to gather accurate information about language use patterns in a way that is intuitive, and natural to them.

We will present the system currently under development, and share the lessons we have learned while building a text exploration interface for use in the humanities.

System Description: Computation
The goal of the computation is to extract grammatical structure from text. We use established techniques in computational linguistics that allow fast and reliable extraction of phrase boundaries, inter-word relationship categories (e.g. subject, verb, object, modifier etc.) and parts of speech. For example, the sentence, "Marsupial mammals have pouches." contains the meaningful unit "marsupial mammals" that functions as a noun and is therefore a noun-phrase. It further contains the word "marsupial" which is an adjective modifier of the word "mammals", which is a noun. The process is completely general, and can be applied to any TEI-encoded text collection.

System Description: Interface
Based on the features that our literary scholar collaborators found useful, our current interface supports three tasks: searching for grammatical or word-based patterns in text (Figures 1 and 2), visualizing their distribution in the entire collection (Figure 3), annotating sections of text with tags and notes (Figure 4).

Figure 1: Easily expressing the query, "How was God described?"

Full Size Image

Figure 2: A list of search results augmented with a frequency visualization.

Full Size Image

Figure 3: The distribution of the exact phrase "I was born" through the collection.

Full Size Image

The power of extracting grammatical information is that a list search results is no longer an opaque list: trends and comparative frequencies can be extracted and displayed, giving an instant high-level picture: a guide to further exploration. For example, if a researcher were interested in the relationship between slaves and God, he or she might be interested in how God was described in the collection. Grammatical search makes this query easy to express: just type in "God" and choose the "described as" relationship, as shown in Figure 1

The search results are augmented with the simple but powerful visualization of relative frequencies shown in Figure 2: all the words that God is "described as" arranged from most to least frequent. The presence of the adjectives "great, true, just" immediately evokes a picture of the relationship - one very different from the picture more negative adjectives might paint. While this is no substitute for careful literary analysis, it can be a quick way to judge the extent to which an entity or event is represented a certain way in a collection, and so help formulate new hypotheses.

While investigating stylistic similarities between documents in a collection, it is useful to be able to investigate occurrences of patterns of interest and compare their distributions across documents. We use a visualization called heat maps, which uses the visual metaphor of text as a brick wall, with each brick, a section of text, and each column of bricks a document. Typing in a phrase shows its distribution throughout the entire collection, and patterns, such as the overwhelming occurrence (Olney 1984) of the exact phrase "I was born" at the very beginnings of narratives ( Figure 3) are easily apparent.

The third research behavior we support is that of organizing sections of text into sets that illustrate a point. With our reading and annotation interface ( Figure 4), researchers can highlight sections of text, and add tags and detailed notes.

Figure 4: Reading and annotating documents with tags and notes.

Full Size Image

Adding a tag to a highlighted section of text is like adding it to a set. Researchers can use the sets they make in other parts of the application: searching within a set, or restricting the heat map visualization to documents in a set.
Related Work
In the digital humanities, the closest work to our project comes from two well- known text analytics efforts. The first is the MONK project (MONK) incorporating the SEASR analysis toolkit. These projects offers two computational linguistics tools in addition to word distribution and frequency statistics: tagging words with their parts of speech and extracting named entities. Users can visualize occurrence patterns of word sequences within a chosen text, and plot networks of how often named entities occur near each other. This research led to visual text-mining analyses of Emily Dickinson's correspondence (Catherine Plaisant et al. 2006), and of Gertrude Stein's "The Making of Americans" (Don et al. 2007) and an interface for exploring the parts of speech used near query words of interest (Vuillemot et al. 2009).

The second is Voyeur (Voyeur), which operates entirely at the word level. It allows users to plot word frequencies, see concordances (contexts in which words occur) and create tag clouds.

Other digital humanities projects have used more advanced language processing, but have not developed them into user interfaces or combined them with visualizations. Topic modeling is being applied to 19th Century British and American novels (Jockers 2010). These novels were also the subjects of cutting-edge computational linguistics research that showed how to automatically extract social networks from free text (Elson et al. 2010). Topic modeling is also being applied to the compendium of Danish, Norwegian, and Swedish folklore collected by Evald Tang Kristensen. In the field of visualization, applications to text in the humanities have been limited to word clouds, and node-and-link diagrams of named entities, and co-occurrences.

Outside the realm of text, but in the domain of comparative exploration, LISA, a comparison search interface for cultural heritage artifacts was created by (Amin et al. 2010).

The digital humanities work described above comes from the application of ideas from human-computer interaction and natural language processing. We are informed by general principles of search user interface design described by Hearst (Hearst 2009), and of visual exploration of large data-sets described by Shneiderman (Shneiderman 1996).

References:
Amin, A.K. et al. 2010 “Designing a thesaurus-based comparison search interface for linked cultural heritage sources, ” Proceeding of the 14th international conference on Intelligent user interfaces, 249-258

Voyeur Tools: See Through Your Texts|Hermeneuti.ca - The Rhetoric of Text Analysis, October 29, 2010 (link)

Don, A. et al. 2007 “Discovering interesting usage patterns in text collections: integrating text mining with visualization., ” Proceedings of the sixteenth ACM conference on Conference on information and knowledge management, Lisbon, Portugal, 213-222

Elson, D. K. Dames, N. McKeown, K. R. 2010 “Extracting social networks from literary fiction, ” Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, 138-147

Hearst, M. 2009 Search user interfaces, Cambridge Univ Press http://searchuserinterfaces.com

Jockers, M L. What is a Literature Lab: Not Grunts and Dullards|Matthew L. Jockers., October 29, 2010 (link)

Moretti, F. 2005 Graphs, Maps, Trees: Abstract models for a literary history, Verso Books

Olney, J. 1984 “"I Was Born": Slave Narratives, Their Status as Autobiography and as Literature, ” Callaloo, 46-73

Plaisant, C, et al. 2006 “Exploring erotics in Emily Dickinson's correspondence with text mining and visual interfaces, ” Proceedings of the 6th ACM/IEEE-CS joint conference on Digital libraries, Chapel Hill, NC, 141-150 (link)

Shneiderman, B. 1996 The Eyes Have it: A Task by Data Type Taxonomy for Information Visualizations, April 22, 2010 (link)

Vuillemot, R. et al. 2009 “What's being said near "Martha"? Exploring name entities in literary text collections, ” In Visual Analytics Science and Technology: VAST 2009. IEEE Symposium, 107-114