New Search | Return to Search Results | Print View

Paper: Topic Modeling Historical Sources: Analyzing the Diary of Martha Ballard

Blevins, Cameron, Stanford University, cblevins@stanford.edu


In 1991, historian Laurel Ulrich’s A Midwife’s Tale swept a little-known 18th-century midwife named Martha Ballard into the national historical consciousness. Ulrich’s work centered on the analysis of nearly 10,000 diary entries penned by Ballard between 1785 and 1812, leading to an exploration of issues such as shifting family structures, the professionalization of obstetrics, and debtor patterns in a rural economy (Ulrich 1991). My research examines the same diary, but instead of a traditional close reading of the source, I use topic modeling to mine a digitized transcription, iterating through hundreds of thousands of words in order to search for textual patterns.

One of the fundamental challenges to applying text processing techniques to historical sources is one of data quality. Older, hand-written documents are often difficult to transcribe into a digital format, while the shorthand style of diary writing is often filled with abbreviations and misspellings. For instance, Ballard employs a vocabulary peppered with variations: the word “daughter” is spelled fourteen different ways: “daught,” “dagt,” “dat,” etc. One way to overcome this challenge is to use topic modeling, a method of computational linguistics that attempts to group words together based on their appearance in the text.

My short paper session focuses on an analysis of a historical source using topic modeling (Blei and Lafferty 2009). As a form of linguistic analysis, topic modeling has been employed over the past several years to examine large-scale, multi-author textual databases, including historical newspapers (Block 2006), journal articles (Gerrish et al. 2010, Hall et. al. 2008), and social network data (Ramage et al. 2010). My application of topic modeling differs from many of these investigations by focusing on multiple, short texts by a single author: in this case, Ballard’s diary entries.

I employed the machine learning toolkit MALLET (McCallum 2002) in order to topic model each of Ballard’s entries as separate pieces of text. MALLET, identified thirty topics, which I then labeled for clarity. The following sample topics were some of the most coherent (my own labels in bold and uppercase):

Topics

           
Topic Label  Topic Words 
MIDWIFERY  birth deld safe morn receivd calld left cleverly pm labour fine reward arivd infant expected recd shee born patient 
CHURCH  meeting attended afternoon reverend worship foren mr famely performd vers attend public supper st service lecture discoarst administred supt DEATH: day yesterday informd morn years death ye hear expired expird weak dead las past heard days drowned departed evinn 
GARDENING  gardin sett worked clear beens corn warm planted matters cucumbers gatherd potatoes plants ou sowd door squash wed seeds 
SHOPPING  lb made brot bot tea butter sugar carried oz chees pork candles wheat store pr beef spirit churnd flower 
ILLNESS  unwell mr sick gave dr rainy easier care head neighbor feet relief made throat poorly takeing medisin ts stomach 

Although topic modeling was useful for overcoming some of the challenges of spelling variations, its real value lies in its ability to quantitatively measure the relative thematic content of each piece of text. In the case of Ballard’s diary, MALLET assumes that each diary entry is compromised of some combination of thirty topics. An entry in which Ballard attended a sermon and purchased supplies from the general store might contain, for instance, scores of 50% for the CHURCH topic, 25% for the SHOPPING topic, and minimal or zero scores for the remaining twenty-eight topics. Associated temporal metadata (day, month, year, day of the week) allowed me to chart the behavior of certain topics over time.

As a simple barometer of its effectiveness, I used one of the generated topics that I labeled COLD WEATHER, which included words such as cold, windy, chilly, snowy, and air. Aggregating its entry scores by month shows exactly what one would expect over the course of a year ( Figure 1).

Figure 1

Full Size Image

This approach also can chart patterns over the course of the diary, which covers the final twenty-seven years of Ballard’s life. Two topics tended to involve words related to HOUSEWORK. Aggregated by year, they demonstrate a steady increase in the frequency with which Ballard writes about daily chores ( Figure 2).

Figure 2

Full Size Image

Both topics moved in tandem and steadily increased as she grew older (excepting a curious divergence in the last several years of the diary). This is somewhat counter-intuitive, as one would assume the household responsibilities for an aging grandmother with a large family would decrease over time. Yet this pattern bolsters the argument made by Ulrich in A Midwife’s Tale, in which she points out that the first half of the diary was “written when her family’s productive power was at its height.” (Ulrich 1991, pp. 285) As her children married and moved into different households, and her own husband experienced mounting legal and financial troubles, her daily burdens around the house increased. Topic modeling quantifies and visualizes this pattern, one not immediately visible to a human reader.

Topic modeling allows for patterns to crystallize that are imperceptible to a human reader. One topic was particularly intriguing, and included the words: feel husband unwel warm feeble felt god great fatagud fatagued thro life time year dear rose famely bu good

These were words that seem to cover EMOTION and spiritual reflection – an abstract topic that is difficult enough for a human reader to describe. Yet the computer did a remarkable job in identifying a cohesive group of words. The topic follows a fascinating trajectory in Ballard’s diary ( Figure 3).

Figure 3

Full Size Image

Not only did Ballard write about this topic more as she grew older, but there was a dramatic leap from 1803 to 1804-1805. This corresponds quite well to the period of intense family travail: Her husband was imprisoned for debt and her son was indicted by a grand jury for fraud, causing a cascade effect in Martha’s own life. Topic modeling not only reveals the trajectory of tangible themes (housework, births, gardening, etc.), but also begins to quantify and visualize abstract themes by charting Ballard’s emotional state of being.

My short paper session focuses on the results of my existing work on topic modeling Ballard’s diary while outlining some of the future paths this research could take. In particular, I am interested in pairing trends in topics with trends in Ballard’s social network. What topics correlate with what kinds of people? Are women or men described alongside particular themes? In what broad context do ministers, doctors, neighbors, or family members appear? In conjunction with traditional research and analysis, topic modeling presents a valuable methodology for examining historical sources.

References:

Blei. D. Lafferty, J. 2009 “Topic Models, ” Text Mining: Classification, Clustering, and Applications., Srivastava, A. and Sahami, M. (ed.) Champan & Hall Boca Raton pp. 71-94

Block, Sharon 2006 “Doing More with Digitization: An Introduction to Topic Modeling of Early American Sources, ” Common-Place, 6.2 (link)

Gerrish, S. Llewellyn, C. 2010 “JSTOR Discipline Browser, ” JSTOR, (link)

Hall, D. Jurafsky, D. Manning, C. 2008 “Studying the History of Ideas Using Topic Models, ” Proceedings of Empirical Methods of Natural Language Processing., (link)

McCallum, Andrew 2002 MALLET: A Machine Learning for Language Toolkit, (link)

Ramage, D. Dumais, S. Liebling, D. 2010 ““Characterizing Microblogs with Topic Models.”, ” International Conference on Weblogs and Social Media., (link)

Ulrich, Laurel Thatcher 1991 A Midwife’s Tale: The Life of Martha Ballard and Her Diary, 1785-1812., Vintage New York