Attitude Identification through Parameter-Efficient Fine-Tuning
Proceedings of the Second Workshop on Building Educational Applications Using NLP
Abstract
We investigate automatic attitude detection in UN Security Council speeches using adapters. Following Martin and White’s Appraisal Theory, we identify three types of evaluative language: affect (emotional responses such as hope or concern), judgement (ethical evaluations of behavior), and appreciation (valuations of objects or situations). Training only 0.95% of BERT-large’s parameters, adapters achieve F1 scores ranging from 0.76 (affect) to 0.46 (appreciation), approaching full fine-tuning performance while enabling rapid task-specific experimentation. Differences in observed evaluation metrics mirror the pattern of the human inter-annotator agreement. This correlation suggests that computational difficulty reflects genuine linguistic ambiguity. Affect benefits from conventionalized diplomatic expressions, while appreciation faces context-dependent evaluation and severe class imbalance. Analysis demonstrates that evaluative intensity varies systematically across diplomatic contexts, with implications for corpus design in specialized discourse analysis.