Research

This research addresses the challenge of improving the quality of data we use to train AI, especially in the context of incivilities, where the results could be controversial. This concept of "controversial labels" is characterized by high disagreement among labelers, which cannot be solved with simply more labelers. I propose using cognitive empathy priming to enhance label quality and consistency. Through two studies, I demonstrate that empathy-based treatments can significantly improve labelers' ability to identify controversial sexism. This priming results in more accurate and consistent annotations. My findings show substantial increases in model performance, with the F1 score of models trained with the empathy-primed models showing improvements of up to 28.29% compared to models trained with control group labels without the priming.

This paper examines how different levels of incivility affect subsequent user engagement. I collected a unique dataset from a community site with a chronological structure that minimizes endogeneity concerns. I categorize incivility into extreme, mild, and no incivility using fine-tuned large language models. My theoretical framework, based on Zajonc's (1980) theory of emotion and cognition, distinguishes extreme and mild incivility treatment effects. I find that mild incivility, which triggers emotional arousal without significant cognitive dissonance, enhances user engagement and comment originality. In contrast, extreme incivility decreases engagement and comment novelty, likely due to cognitive dissonance. Interestingly, such self-regulation is not expressed in upvotes and downvotes, which involve less cognitive process. Also, both levels of incivility increase subsequent uncivil comments while decreasing civil ones.

While "Tainting the Discourse: The Role of Incivility in Shaping Subsequent User Engagement” investigates the effect of different "degrees" of incivility, this paper examines the impact of the "target" of incivility. Building on social identity theory, I suggest two mechanisms that lead to incivility: competitive and group-conforming incivility. I examine how these mechanisms influence later user involvement and incivility. Results indicate that group-conforming incivility is a greater determinant of uncivil content than competitive incivility, particularly towards women. I also show that incivility towards uncivil users does not prevent subsequent uncivil comments.

I investigate the consequences of accountant shortages by examining the link between job vacancy duration and internal control deficiencies. Using the entire online job postings posted online, I estimate the vacancy duration to fill the positions. I find that longer vacancy duration is positively associated with the probability of material internal control weaknesses. This association is stronger for smaller firms, firms with lower labor intensity, and firms in sectors with more complex accounting standards. I also explore whether labor shortages prompt firms to invest more in technology and how such investments affect their future labor needs and internal control quality.

This paper investigates how individual game live streamers strategically compete for viewer attention under the exploration-exploitation framework. I apply AI and machine learning methods to quantify streamers' strategic decisions in multiple dimensions, including streaming content and themes. To measure the distance between the content, I conducted domain transfer by pre-training and fine-tuning large language models with a large corpus of games and live streaming. Using these quantified strategies, I develop a market equilibrium structural model that incorporates viewer demand and streamer content generation costs. My findings indicate that exploration can enhance viewer utility, offering insights into content creation and consumption dynamics.

I investigate the effects across the heterogeneous platforms (gaming and live streaming) in one ecosystem. I observe that these platforms benefit each other through subtle, indirect connections. By drawing upon existing research on the indirect impact of symbiosis in ecology, which refers to the way species influence one another through intermediary species, I demonstrate that these indirect symbiotic effects have an influence on other platforms.  In addition, leveraging symbiotic plasticity from the same literature, I investigate how these cross-platform indirect symbiotic effects adapt to natural shock (i.e., COVID-19 as a natural experiment). I demonstrate the existence of an interconnected gaming ecosystem with symbiotic relationships that fluctuate over natural shocks.

I examine the driver of employers' salary disclosure decisions in AI job postings and how these decisions impact the hiring process in the AI labor market. Leveraging large language models, I extracted various data from the AI job postings; I find that labor market competition increases the likelihood of salary disclosure, while proprietary information concerns and internal pay disparity risks reduce it. I find that salary disclosure is associated with longer vacancy durations, particularly for smaller firms and complex AI roles. This extended duration, however, appears to lead to better match quality, as evidenced by lower subsequent employee turnover.