An interactive multi-task ESG classification method for Chinese financial texts

In view of the problems existing in the ESG classification task of Chinese financial texts, such as feature loss caused by excessively long texts, this paper proposes an interactive multi-task model AmultiESG for ESG classification of Chinese financial texts. The model divides Chinese financial text ESG classification and financial sentiment dictionary expansion into primary and secondary tasks. First, BiLSTM model is used to learn the original representation of the text. Then, in the secondary task, the attention mechanism and full connection layers are combined with the domain dictionary to realize the extraction of emotional words. In the main task, in order to prevent feature loss due to the excessively long texts, we process the text again and divide it into blocks according to the period. Meanwhile, we learned new feature representation of the text by combining text label representation, text block representation, BiLSTM output features and domain dictionary features. And we introduce an interactive information transfer mechanism to iteratively improve the predicted results of the two tasks and strengthen the association between them. It has been experimentally demonstrated that the proposed method shows superior performance compared to other baselines for the ESG classification task of Chinese financial text, especially for long-text classification tasks.