Researchers propose NTDH method for complex affective analysis with fewer training records than comparable systems

Comprehensive affective analysis requires reconciling heterogeneous prediction tasks—continuous, ordinal, and multi-label outputs—while accounting for context-dependent affective meaning and conflicting cues. Existing methods typically learn a direct mapping from inputs to labels without explicitly modeling this reconciliation.

The authors propose NTDH, a method that treats affective analysis as a complex-reasoning problem. NTDH introduces a unified output interface across heterogeneous label spaces and a reasoning trajectory over which a verifiable reward can be optimized. The authors state this is the first such treatment covering both sentiment and emotion.

A key obstacle is data: affective reasoning traces must be synthesized, but generic synthesis is misaligned with the targets, tolerances, and phenomena of affect, and may discard or leak failure cases. NTDH addresses four failures through four components: Naturalisation, Tolerance-aware gate, Domain-aware strategies, and Directional Hints.

NTDH’s components include Naturalisation, which sets the training answer to the gold label so it is correct by construction; a Tolerance-aware gate that checks each answer against the task’s scoring margin; Domain-aware strategies that refine reasoning using ideas from affective science; and Directional Hints that report only the type and direction of an error without exposing the target.

The authors train Qwen3-8B using supervised fine-tuning followed by GRPO, using the same tolerance for verification. A component ablation quantifies the data-quality effect of each part. The final policy improves over its SFT checkpoint on five of six official-test metrics and achieves the strongest EI-reg result among compared systems, with a Pearson correlation of 0.862.

The study uses 16,302 training records, about 14 times fewer than comparable instruction-tuned systems, while still improving performance on five of six official-test metrics.