Laura K. Nelson

@lauraknelson

Associate Professor @ UBC computational sociology machine learning is feminist You only have to look at the Medusa straight on to see her. And she’s not deadly. She’s beautiful and she’s laughing. www.lauraknelson.com

939 Folgt4.143 Follower1.964 Posts

Bluesky

Ist das dein Profil?Übernimm es — dann gehört dir lauraknelson.dirty.social.Profil übernehmen

Laura K. Nelson@lauraknelson· 9d

In sum, theories of idea diffusion need to take two additional processes seriously: The creation of ideas as a social process, and how the content of the idea impacts translation and diffusion. And fin.

Our applied creation-diffusion model (showing the Internal factors and External factors as input, going through Substantive content as an output, and the Diffusion as an output) with simplified effects laid over the lines.

Caption: Figure 5. Applied creation–diffusion model with (simplified) effects.
Note. The dotted lines capture pathways shaping idea creation. The dashed lines capture pathways shaping idea diffusion.

Caption:

Laura K. Nelson@lauraknelson· 9d

Specifically for our case, the ways in which an idea engages with gender and feminist scholarship matters for its diffusion. Engagement with "gender" diffuses more, while engagement with "feminism" and structure does not diffuse as widely. +

Marginal effects plot with loads of lines and columns. The point is to show that we measured engagement with different aspects of gender research: Feminism, Gender, Inequality, Structure, and Methods, and their correlation with diffusion.

Caption: Figure 3. Average marginal effects of internal and external factors on engagement with gender equity research.
Note. Full model specifications are in Online Supplemental Appendix 5.2.

Laura K. Nelson@lauraknelson· 9d

But wait, there's more theory! Typical diffusion theories model the internal/external axis (rows), to this, we added the context-content axis (columns). In short, the context of the idea matters for how it diffuses, but so does its content. We measured content via Concept Movers Distance (CMD) +

A classic 2x2 table, with Internal and External as rows, Context and Content as columns.

Caption: Figure 6. Two-way categorization of factors influencing diffusion.

Laura K. Nelson@lauraknelson· 9d

Specifically, we propose the existing theory, the emergence-diffusion model (top panel) assumes ideas are a given, and then models how they spread once they exist. We suggest that we have to model how ideas themselves emerge, which is itself a social process. Thus, our Creation-Diffusion model. +

Diagram with two panels, The top is the emergence-diffusion model, with "Internal factors" and "External factors" as input on the left, pointing to "Diffusion" as an outcome on the right. The bottom panel, our contribution, as "Internal factors" and "External factors" as input leading to "Substantive content" as an intermediary output leading again to "Diffusion" as tn output. The arrows flow through and around "Substantive content"

Caption:
Figure 1. Two models of diffusion.
Note. The top panel (Emergence–Diffusion) is our interpretation of the existing diffusion model latent in many quantitative diffusion studies. The bottom panel (Creation–Diffusion) is our proposed feminist-inspired diffusion model.

Laura K. Nelson@lauraknelson· 9d

Who's up for a good old fashioned (but short, bc I'm tired) paper thread? Thrilled to be in this special issue of Gender & Society (an absolute dream!), where we propose a feminist-inspired extension of how to theorize and model the spread of ideas in science + journals.sagepub.com/doi/10.1177/...

What Gets Lost in Translation? Epistemic Tensions between Translation and Diffusion in Practice-Oriented Scholarship

Abstract
This paper proposes a feminist-informed metascience theory to explain the interconnected processes of idea creation, translation, and diffusion. Drawing on feminist critiques of science centered on perspective, power, and praxis, we develop a creation–diffusion model that jointly examines how ideas are created and then how the content of ideas influences their spread across academic fields. Using a nationally funded gender equity program as a case of a practice-oriented knowledge-production community, we analyze how gender equity research concepts were translated and diffused through the program-funded publications. We operationalize substantive engagement with gender equity research using word embeddings and Concept Mover’s Distance. Then, with additional contextual variables, we measure (1) the predictors of this engagement (ideas) and (2) the impact of engagement on citation counts and citation interdisciplinarity (the diffusion process) across social science and STEM fields. Our findings show that, in addition to context, the content of the idea matters for diffusion: Engagement with critical dimensions of gender equity research, such as feminism and structural concepts, diffused less often, whereas other dimensions, such as gender and methods, achieved broader uptake. Our empirical findings highlight the role of perspectives and epistemic power in shaping both knowledge production and diffusion. Our theoretical framework and methods demonstrate how metascience can incorporate feminist theory directly into diffusion studies, and why it should.

Laura K. Nelson@lauraknelson· 20d

I haven't read this yet but come on with that headline. The study of sociology. The study, of soci-ology.

Screenshot of an article in the WSJ.
Image: two podiums facing one another with a crack in the ground beneath them. A call-out with protest signs above the first podium, a call-out with a stack of old looking books above the second.

Text: THE SATURDAY ESSAY

Is It Time to ‘Liquidate’ the Study of Sociology?
Sociology has long been tasked with investigating society’s most pressing challenges. But these days, sociologists’ greatest obstacle may be themselves.

Laura K. Nelson@lauraknelson· 24d

Link! journals.sagepub.com/doi/10.1177/...

Article
What Gets Lost in Translation? Epistemic Tensions between Translation and Diffusion in Practice-Oriented Scholarship
Steven Lauterwasser1, Laura K. Nelsonhttps://orcid.org/0000-0001-8948-300X2, Jessica R. Gold1, Alexander J. Gates3, and Kathrin Zippel4
Abstract
This paper proposes a feminist-informed metascience theory to explain the interconnected processes of idea creation, translation, and diffusion. Drawing on feminist critiques of science centered on perspective, power, and praxis, we develop a creation–diffusion model that jointly examines how ideas are created and then how the content of ideas influences their spread across academic fields. Using a nationally funded gender equity program as a case of a practice-oriented knowledge-production community, we analyze how gender equity research concepts were translated and diffused through the program-funded publications. We operationalize substantive engagement with gender equity research using word embeddings and Concept Mover’s Distance. Then, with additional contextual variables, we measure (1) the predictors of this engagement (ideas) and (2) the impact of engagement on citation counts and citation interdisciplinarity (the diffusion process) across social science and STEM fields. Our findings show that, in addition to context, the content of the idea matters for diffusion: Engagement with critical dimensions of gender equity research, such as feminism and structural concepts, diffused less often, whereas other dimensions, such as gender and methods, achieved broader uptake. Our empirical findings highlight the role of perspectives and epistemic power in shaping both knowledge production and diffusion. Our theoretical framework and methods demonstrate how metascience can incorporate feminist theory directly into diffusion studies, and why it should.
Keywords
metascience, feminist theory, idea diffusion, translational research, word embeddings, Concept Mover’s Distance, power, praxis, science of science

Laura K. Nelson@lauraknelson· 27d

The free version of ChatGPT automatically searches the web. The prompt "What are the most relevant sources for the claim that LLMs hallucinate?" returns the following. All those link buttons in the output are live hyperlinks that were found, not generated. If this isn't a search engine, what is it?

The foundational paper (introduced the term in the LLM context)
Ji et al. (2023)

Ji, Ziwei, et al. “Survey of Hallucination in Natural Language Generation.” ACM Computing Surveys 55, no. 12 (2023): 248.

This is the standard survey article.

It defines hallucination as generated content that is nonsensical or unfaithful to the provided source/input.

It distinguishes intrinsic hallucinations (contradict the source) from extrinsic hallucinations (unsupported by the source).

Useful when you need a general definition and overview of causes, detection, and mitigation strategies.

The paper most often cited for why hallucinations occur
Maynez et al. (2020)

Maynez, Joshua, et al. “On Faithfulness and Factuality in Abstractive Summarization.” Proceedings of ACL (2020).

Shows that neural generation systems often produce statements that are not supported by the source text.

Helped establish factuality/faithfulness as a core evaluation problem for generative models.

Frequently cited as evidence that hallucination is not an edge case but a structural property of neural text generation.

The key benchmark paper
TruthfulQA

Lin, Stephanie, Jacob Hilton, and Owain Evans. “TruthfulQA: Measuring How Models Mimic Human Falsehoods.” ACL (2022).

Evaluates whether models give truthful answers to questions that commonly elicit misconceptions.

Larger models often performed worse on truthfulness than smaller ones, because they better imitate patterns in web text.

This is one of the strongest empirical demonstrations that scaling alone does not eliminate hallucinations.

The classic “stochastic parrots” critique
Bender et al. (2021)

Bender, Emily M., et al. “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” FAccT (2021).

Does not use the term “hallucination” as the central concept.
...

Laura K. Nelson@lauraknelson· 27d

Definitionally

Screenshot of the Wikipedia page for "Search Engine":

A web search engine or Internet search engine is a software system that provides hyperlinks to web pages and other relevant information on the Web in response to a user's query.

Laura K. Nelson@lauraknelson· Jul 10

Good paper! You could have maybe included Sociological Methods & Research and/or Sociological Methodology tho? Sad to see sociology completely left out of this review 🥲 journals.sagepub.com/doi/abs/10.1... journals.sagepub.com/doi/full/10....

a screenshot from the text reading: "We also searched top sociology journals, but did not find any papers involving prompting LLMs in American Journal of Sociology, American Sociological Review, or the Annual Review of Sociology."

Laura K. Nelson@lauraknelson· Jun 29

Digging into this one, enjoying it. One argument from China is the Marxist argument that AI can free us from intellectual labor, the way machines can free us from physical labor. Not judging that claim, but I'm curious: is anyone seriously grappling with that line in North America? Jacobin, maybe?

Picture of the front cover of the book "Machine decision is not final: China and the history and future of artificial intelligence"

Laura K. Nelson@lauraknelson· Jun 27

I really enjoy walking into the Vancouver School of Economics

Laura K. Nelson@lauraknelson· Jun 24

Got curious and looked up the Wikipedia page on Generative AI and boy oh boy semantic confusion all the way down. You're absolutely right and I'm already giving up this battle, but I still will maintain we can all be more responsible when we teach this stuff (especially in non-NLP courses).

Screenshot of a Wikipedia page:
From Wikipedia, the free encyclopedia
"GenAI" redirects here; not to be confused with Artificial general intelligence.
Generative artificial intelligence (GenAI) is a subfield of artificial intelligence (AI) that uses generative models to generate text, images, videos, audio, software code (vibe coding) or other forms of data.[1] These models learn the underlying patterns and structures of their training data, and use them to generate new data[2] in response to input, which often takes the form of natural language prompts.[3][4]

Laura K. Nelson@lauraknelson· Jun 24

And one more, for good measure (why not?)

Screeshot from Blei 2012:

Lda and probabilistic models. LDA
and other topic models are part of the
larger field of probabilistic modeling.
In generative probabilistic modeling,
we treat our data as arising from a
generative process that includes hid-
den variables. This generative process
defines a joint probability distribution
over both the observed and hidden
random variables. We perform data
analysis by using that joint distribu-
tion to compute the conditional distri-
bution of the hidden variables given the

Laura K. Nelson@lauraknelson· Jun 24

From the same Blei 2012 article: dl.acm.org/doi/10.1145/...

Screenshot from Blei 2012:

We should explain the mysterious name, “latent
Dirichlet allocation.” The distribution that is
used to draw the per-document topic distribu-
tions in step #1 (the cartoon histogram in Figure
1) is called a Dirichlet distribution. In the genera-
tive process for LDA, the result of the Dirichlet
is used to allocate the words of the document to
different topics. Why latent? Keep reading.

Laura K. Nelson@lauraknelson· Jun 24

A generative model specifies a data-generating process: it assumes that the observed data were produced by some latent variables and stochastic steps. Topic models are generative in exactly this sense. (I don't think this a contested fact at all. Topic models are generative.) From Blei 2012:

Screenshot of the Blei 2012 article:

LDA is a statistical model of docu-
ment collections that tries to capture
this intuition. It is most easily described
by its generative process, the imaginary
random process by which the model
assumes the documents arose. (The
interpretation of LDA as a probabilistic
model is fleshed out later.)

Laura K. Nelson@lauraknelson· Jun 24

My spicy ghost pepper hot, but 100% genuine take: scholars in my specific subfield (sociology of culture) have been using generative AI responsibly in research since at least 2013. I'll even put a date on it: Mohr and Bogdanov, 2013. + www.sciencedirect.com/science/arti...

Screenshot of an article published in Poetics in December 2013: Introduction—Topic models: What they are and why they matter John W. Mohr, Petko Bogdanov

Laura K. Nelson@lauraknelson· Jun 19

Image of Emile Durkheim with a speech bubble at the bottom

Laura K. Nelson@lauraknelson· Jun 17

Summer reading stack. What should I add??

Stack of books vertically on a book shelf. From top to bottom: Medium Hot by Hito Steyerl, A an edited volume called Machine Decision is not Final: China and the History and Future is Artificial Intelligence, Algorithms of Resistance by Tiziano Bonini and Emiliano Trere, Staying with the Trouble by Donna Haraway, Atlas of Anomalous AI, Data-Spheres of Planetary Urbanization, Other Networks by Lori Emerson

Laura K. Nelson@lauraknelson· Jun 2

Fascinating thread. Post after post of students describing how, pre-AI, they got As on book reports and courses without reading a single book. Either this is all a bit I don't understand (very possible, I don't get social media) or we* need to really sit with how students move through HS. *Just me?

Post with 385 reposts, 30 quotes, and 1.8k likes. The post says: "Oh, you can't write a 5-page book report without AI? I fucking wrote 5-page book reports without having read the goddamn book. Get your finger out of your ass."

Laura K. Nelson@lauraknelson· May 27

Air Canada gives these little foam tic tac toe games to kids when they fly and it has the most irritating design I've ever seen and I'm determined they did this on purpose solely to drive adults bonkers on long haul flights because Whyyyy?

Tic-tac-toe board with a black background and blue lines, with one red "O"placed that fits perfectly in the square and one white "X" placed that totally inexplicably does not fit perfectly in the square even though it could be cut to fit and no matter which way you place it there's a tiny gap that I'm determined was designed solely to drive the adult bonkers on long haul flights

Laura K. Nelson@lauraknelson· May 21

Effective data visualisation showing the amount of land required to produce 100g of various proteins. The red one (beef) goes up two full flights of stairs, then around the corner, continuing into an exhibit. @ the Natural History Museum in Vienna

Bar chart printed on stairs showing how much land is required to produce 100g of various protiens. The bars start on the ground and are printed up the stair case. The proteins are tofu, grain, nuts, milk, and beef. The title, printed on the ground: How much ground have you eaten today?

Laura K. Nelson@lauraknelson· May 9

Laura K. Nelson@lauraknelson· May 8

Wow this figure. AI reviews appear to be (subjectively) much better along almost all dimensions 👀

Figure 2: Survey responses: AI vs. human review comparisons, The figure is a horizontal bar chart with a marked 0 line, and bars to the left indicated human preferred and bars to the right indicating AI preferred, on a continuous scale from -0.6 to 0.6. The figure shows the differences in the mean response score between AI and human reviews for each of the nine review-quality criteria. In six out
of nine criteria, AI reviews were rated higher than human reviews. The preference towards AI reviews was stronger for authors than for PC, SPC, and ACs. All p-values show strong statistical significance at the α = 0.01 level.

Laura K. Nelson@lauraknelson· Apr 27

Image showing how much space is required to transport 60 people in cars (panel 1), bus (panel 2), and bicycle (panel 3). Not surprisingly cars take up an enormous amount of space.

Laura K. Nelson@lauraknelson· Apr 23

See also pubs.aeaweb.org/doi/pdfplus/... (note the different y-axis scales across plots)

This figure shows weighted citation rates from each of five social sciences to the other four. Plots are smoothed with five-year moving averages. Papers cited were published between 1955 and 2015. The five social sciences are Economics, Psychology, Political Science, Anthropology, and Sociology. Economics' weighted citation share sociology is just above 0.01 in 2015. Sociology's weighted citation share to Economics is just below 0.10 in 2015, 10x greater.

Laura K. Nelson@lauraknelson· Apr 17

I'm ok with this. The Conversation makes academic work accessible to the public. I, an employed academic, am paid to produce academic work, and work with The Conversation to translate it for the public, with my portion of that work paid via my (public) salary. They're not publishing freelancers.

From the Wikipedia article: Except in "exceptional circumstances", it only publishes articles by "academics employed by, or otherwise formally connected to, accredited institutions, including universities and accredited research bodies".

https://en.wikipedia.org/wiki/The_Conversation_(website)

Laura K. Nelson@lauraknelson· Apr 2

This figure is the one that is actually embarrassing for sociology:

Fig. 2: Data and code availability by field. Left, data and code availability as a percentage of papers. Right, the raw counts of papers with data and code available and not available. Restricted data (purple) did not count as available data, but might be accessible in principle.

This shows political science has thee most papers with open data/code, followed by economics, then much further behind, psychology and then sociology (sad). Sociology is well below all field average.

Laura K. Nelson@lauraknelson· Apr 1

You can't drop that one in the chat and not the other one. Sociology sits differently in this one. The two graphs together are much more interesting.

A chart showing the replication rate (same methods, different data). In this graph, unlike the other, sociology is over 50% replicable.

Laura K. Nelson@lauraknelson· Mar 31

Bluesky, pink flowers I wish I could also share the olfactory experience walking down this corridor. It's glorious (if you don't get seasonal allergies)