ESG anchor words and risk
用深度学习和财报电话会议分析ESG锚定词的情感模式，发现环境情感能降低股价崩盘风险，社会情感效果取决于叙事语境，治理情感影响微弱。
Abstract The present study provides a novel methodology to quantify ESG, using deep machine learning along with earnings conference calls, to examine the extent to which sentiment patterns around ESG anchor words convey information about future stock crash risk. We define ESG anchor words as terms related to the three pillars of ESG, with the sentiment surrounding these words revealing firms' disclosure strategies and information transparency. As Bordalo et al. (Q J Econ 135:1399–1442, 2020) show, an anchor may relate to higher attention being paid by an observer. In specific, they show that when a risk surpasses a known threshold, identified as a risk-anchor, people seem to pay more attention. For the task in hand, we first compile expanded dictionaries of ESG-related terms using word2vec and then employ FinBERT, to extract sentiment from sentences containing these ESG anchor words. Using stock crash risk as our primary measure of information asymmetry and bad news hoarding, our results show that environmental sentiment operates as a credible signal that reduces crash risk, with effects amplified in environmentally sensitive industries and attenuated by financial opacity; social sentiment exhibits context-dependent patterns, increasing crash risk in optimistic narratives but reducing it in balanced communications, while governance sentiment shows low impact. Our findings likely imply that not all ESG communication reduces information asymmetry; credibility depends critically on dimension-specific materiality, narrative consistency and financial transparency.