RoBERTa-Based Sentiment Mining for Engineering Technical Reports

Abstract

Engineering technical reports document the operations and changes, providing a basis for decision-making. Due to the highly specialized nature of these documents, sentiment extraction becomes challenging because they contain complex jargon, formal structures, and the implicit nature of evaluative language. This paper uses a customized RoBERTa model to build a sentiment analysis system for engineering technical reports. These three methods are used to extract subtle evaluative expressions in technical papers. These methods include expanding domain-specific vocabularies, adaptive segmentation and hierarchical embeddings, as well as customized attention mechanisms. A multi-source engineering corpus containing over 30,000 real technical documents was evaluated, and detailed annotation and segmentation were performed. The proposed model significantly outperforms traditional baseline models on datasets with higher term density, achieving an average accuracy of 93% and a macro F1 score of over 0.81 in negative sentiment detection. Without domain vocabulary adaptation and structural encoding, the F1 score would drop by more than 10%. Targeted model adaptation can improve the accuracy of sentiment mining in the engineering field and support the implementation of more reliable safety management, compliance assessment, and other measures.