Discovering Higher-Level Patterns - Nelson 2015

[jamesallenevans]

Nelson, Laura K. 2015. “Persistent Political Logics: Geographical Differences and Temporal Continuities within the Women’s Movements in Chicago and New York City.” Working paper. (But you must also skim the framework she draws from her analysis in Nelson, Laura K. 2017. “Computational Grounded Theory: A Methodological Framework.” Sociological Methods & Research DOI: 10.1177/0049124117729703: 1-40.)

[di-Tong]

（1）Analysis Unit of Topic modeling: Nelson mentioned that in her corpus, each page is treated as one “document,” which is a relatively arbitrary division of text. I wonder why she chose to use page instead of the natural article as document unit for training the topic model. Is it a way to increase document number and standardize the length of each document? Do these two elements matter a lot for the performance of the topic model? What are the minimum document number and length required for the stable performance of a topic model? Would such an arbitrary division of text have a huge impact on the results?

（2）Metadata and topics: While Nelson uses structural topic modeling (STM), a model that allows topic prevalence and contents to vary according to document-level meta data, she seems to use the topic weights of each documents to identify the top topics (which can also be done with a simple LDA). I wonder why she chose STM instead of other forms of topic models here. In order to compare between different organizations over different time periods, would fitting four separate topic models for the four corpus does the same job? What is the strength of her specific design here?

（3）Choice of topic number: I wonder how Nelson decided upon the 20-60 topic number range for initial exploration in the first place. Is there some kind of rubrics about the corresponding relationships between the size of the corpus (both in terms of words and document counts) and the possible topic number range?

[ckoerner648]

Nelson 2015 analyzes political speeches, op-eds, pamphlets, policy recommendations of the women organizations in New York and Chicago and asks if the waves metaphor is suitable to describe the women’s movement. She finds that the groups have different strategies to campaign for change: The women’s organization in Chicago focused on concrete and particular issues and solutions, e.g. to pass a factory law in Illinois. Women in New York focused on abstract ideas and concepts. In their publications, the two groups use particular and very distinct groups of words: Typical words used by women in Chicago are ‘children’, ‘union’, club’, and ‘school–which all indicate concrete entities. Women in New York use expressions like ‘feminist’, ‘radical’, and ‘will’–which are used to describe general issues and solutions. The paper infers the theoretical assertion that institutionalized and heterogenous local beliefs that persist over time can be more influential than overarching macro social-structural beliefs. Nelson’s study is methodologically absolutely different, but the result fits in very well with historian Lisa Tetrault’s claim in “The Myth of Seneca Falls."

[acmelamed]

The proposed utility of grounded theory as included in the methodology outlined by Nelson here seems like it has the potential to alleviate much of the ontological problems commonly associated with computational content analysis. By incorporating subjective human input into the process at a point where the initial computational pattern detection process has already been conducted, but before the final confirmation of those patterns via NLP, this method makes a convincing argument for being superior to previous strategies which attempt to ontologically privilege either human or machine classifications in a non-dialectical sense.

[VivianQian19]

My question is, since Nelson used four albeit the most influential women’s organizations from each period, how representative are these four organizations of the rest of the women’s organizations in Chicago and New York? That is, can the analysis of these four organizations tell use something beyond what is true of them? are the results generalizable?