Abstract
Classifying corporate environmental, social, and governance (ESG) disclosures is challenging because they combine technical terminology with subtle, domain-specific distinctions. Pretrained Transformers such as ClimateBERT perform well but require many parameters and substantial GPU memory. This paper introduces TC-HGAT, a lightweight hierarchical graph attention network with task-semantic conditioning. TC-HGAT uses frozen FastText-300D embeddings and builds, for each document, a graph of word nodes, virtual cluster nodes, and a task-semantic node obtained by mean-pooling the embeddings of a natural-language task description. On three ESG tasks (Claim, Specificity, and Commitment), TC-HGAT achieves macro-F1 scores of 94.6%, 77.8%, and 79.3%, with the highest mean scores on Claim and Commitment, although no difference from ClimateBERT is statistically significant. Most of the gain comes from the task-semantic node, which improves macro-F1 significantly on all three tasks (by 5.7, 2.4, and 4.5 points), whereas the cluster layer alone gives a modest improvement that is significant only on Claim. Random vectors in the same graph position, including a norm-matched one, do not reproduce this gain, indicating that it depends on the content of the task representation. TC-HGAT trains about 1.4 million parameters, far fewer than BERT-base and ClimateBERT, and needs substantially less peak GPU memory during training.