Bolun Zhang, Zhejiang University
The influx of Natural Language Processing (NLP) technologies from the technology sector into academia has catalyzed the expansion of computational social science in the last decade. However, the significant contrasts in the political-economic regimes between these domains have been largely overlooked by researchers. Algorithms commonly deployed in both spheres perform divergently, yielding variable outcomes due to the disparate roles they fulfill in the technology industry versus academic research. Our study focuses on topic modeling, a popular NLP technique in computational social science, to illustrate this gap. Social scientists grapple with the algorithmic instability when applied to identical corpus—a non-issue for tech industry professionals. To understand this discrepancy, we conducted a historical genealogical analysis of topic modeling's development and its diffusion from industry to social science. Our findings reveal that identical algorithms, when situated in distinct organizational settings and aimed at different objectives, are subject to alternative evaluative standards. In short, the governing political-economic regimes of it are different in these two domains. Reflecting on these findings, we propose that methodological transfers in technology are not merely technical in nature. Algorithms and their application have political and economic presuppositions. In this sense, computational social science has many to learn from digital sociology and critical algorithm studies.
No extended abstract or paper available
Presented in Session 52. Digital Methods: Archives, Algorithms, AI