Societal and personal values are transmitted to younger generations through interaction and exposure. Traditionally, children and adolescents learned values from parents, educators, or peers. Nowadays, social platforms serve as a significant channel through which youth (and adults) consume information, as the main medium of entertainment, and possibly the medium through which they learn different values. In this paper we are interested in the values of TikTok videos created by influencers targeting children and adolescents. We aim to extract these values manually, and develop methods to identify them automatically using artificial intelligence.
We curated a dataset of N = 885 TikToks and annotated them according to the Schwartz Theory of Personal Values. We identified high frequency of self-focus values, such as stimulation, hedonism, and power. We identified lower percentage of social focus values, such as universalism and conformity. We then experimented with an array of Masked and Large language models (MLM vs LLM, respectively), investigating their utility in value identification. We compare a direct approach, in which LLM's extract values from videos, and a 2-step approach, in which LLMs extract scripts of videos, which are then analyzed by either an MLM or an LLM. Achieving state-of-the-art results, we find that the 2-step approach performs better than the direct approach and that using a few-shot application to Gemini 1.5 outperforms a trainable Masked Language Model as a second step significantly.