4. Word Embeddings to Explore Meaning Spaces- [E3] Nelson, L. K.

[lkcao]

Post questions here for this week's exemplary readings:

3. Nelson, L. K. 2022. “Leveraging the alignment between machine learning and intersectionality: Using word embeddings to measure intersectional experiences of the nineteenth century U.S. South.” Poetics, 88, 101539.

[chanteriam]

I really enjoyed this reading and it felt to me like an eye-opening way to apply vector algebra and ML to important qualitative theories, particularly one as difficult to quantify as intersectionality. In her discussion, Nelson talks about how these robust methods do not supplant the need for "standards around robustness and sensitivity checks"; I wonder: how large is the quantitative social sciences community, particularly in relation to using machine learning techniques? Are there enough scholars with both the theoretical and the quantitative/computer science background to fact-check these studies? Have there been issues with seemingly plausible, but incorrect, social scientific studies using newer machine learning methods being published because there were not enough individuals familiar with the methods to fact-check their usage?

[yueqil2]

Nelson empirically demonstrates an alignment between machine learning and intersectionality and I assess that she sets forth a potential approach to identity politics. When "Black women" served as one of the four averaged vectors representing four intersectional identities, it has a valence difference with other vectors (male/female, black/white). What if we are curious about the heterogeneity(or divergency) inside the black women group? If we want to know which factors are responsible for the plight of black women, such as race, gender, class, etc., can we also study the word embedding model?

[yuzhouw313]

I was truly surprised by the Dr.Nelson's finding that in the nineteenth century U.S. South black men would have more discursive authority over white women, but I was also amazed by the potential of machine learning techniques and specifically word embedding to uncover such "counter-intuitive" trend. However, the construction of social institution vectors confused me, as the author did so by adding certain pre-defined words to capture words similar. Specifically, how do these selected terms for constructing the vectors capture a comprehensive cultural and racial landscape in the 19-century South and how do they reflect the underlying theoretical assumptions about power dynamics within these social institutions?

[Dededon]

This paper show a good workflow to run an analysis from computational discovery of cultural patterns, to qualitative close-reading of the textual materials. The dictionary method is not perfect, but acceptable in building of the polity, economic, etc. vectors, and gender and race are very easy-to-define categories of intersectionality. I'm glad that the author is honest about the unrobust performance of word-embedding in the discussion session when compared with traditional statistics in measuring culture.

[michplunkett]

I appreciate the application of ML to the concept of intersectionality and the analysis of complicated social relationships as a whole. Perhaps it's a problem with my view on the scope of work required for this analysis, but is ML analysis adding that much to studies like this one? To what degree is the application of ML, or the request to apply it in most social science research, just the request for researchers to do ANOTHER THING?

There is only a finite amount of time to work on any individual question, and the paper doesn't entirely convince me that this addition of another laborious process to the research pipeline is giving a good enough ROI to validate its inclusion. Labs are limited by money, time, work hours, etc. and now they need to have experts on ML in addition to everything they had before. If they are just producing research that says "yeah, that's what we thought," then it feels like a potential waste of energy and time.

[Caojie2001]

The article is interesting as it applies the word embedding method to a field (intersectionality) that is traditionally occupied by qualitative methods. I am interested in the way the words are selected to represent a certain category or institution: is there a scientific way to produce the potential word lists, or are they just selected randomly?

[Carolineyx]

Given the complexity and nuances of intersectionality and the historical context, how does the choice of word embeddings as a method influence the findings of this study, particularly in terms of accurately capturing the lived experiences and societal roles of Black and white men and women? Furthermore, how does the study address the limitations of using machine learning to interpret qualitative data, especially when dealing with historical texts where language use and meanings have evolved over time?