Revealing the Veil of Greenwashing in ESG Reports: Predicting the Degree of Corporate Greenwashing Based on Thematic and Sentiment Features of Text

The global rise of the green economy has positioned green financing as a critical method for enterprises to secure external funding. However, due to the high costs and imperfect regulations associated with environmental information disclosure, some enterprises establish the green image through greenwashing to attract external investment and minimize production inefficiencies, caused by abnormalities such as material shortages. The greenwashing behavior of enterprises is essentially a deception intended for investors, making it challenging for them to make informed decisions. To protect investors' interests, and maintain the market and social order, this paper conducts a study on corporate greenwashing based on the environmental module of ESG reports issued by Chinese Ashare listed companies. First, this paper utilizes MacBERT-LDA to extract topic distribution features (explicit features) and perform topic clustering. It then uses Bi-LSTM to extract topic sentiment features (implicit features) based on the clustering results. The two mutually independent feature extraction networks form a dual-channel structure, fully capturing the explicit and implicit information of the text data. Second, a Transformer model is utilized to fuse the dual-channel feature extraction results, producing richer and more comprehensive semantic features. Finally, these fused features are input into IT2F-BLS to achieve an accurate prediction of the degree of corporate greenwashing, which provides a powerful reference for investors to make decisions. We refer to this ensembled model as the IT2F-BLS greenwashing degree prediction model based on explicit and implicit textual features. Experimental results demonstrate that the model effectively captures both types of information and predicts corporate greenwashing with higher accuracy.