[4K5-GS-6c] Language media processing

4:45 PM - 5:00 PM JST(7:45 AM - 8:00 AM UTC)

[4K5-GS-6c-06] Affective Alignment Optimization for Narrative Translation

〇Takayuki Shimotomai¹, Miwako Kamijo² (1. Headwaters inc., 2. Sagami Women's University)

Keywords:

Machine translation,affective engineering,alignment

Optimizing narrative translation with large language models (LLMs) requires surrogate metrics that can jointly assess (i) affective qualities valued by readers and (ii) structural consistency.This paper (1) analyzes the latent structure of human evaluations and (2) proposes an automatic metric for structural consistency.Our affective evaluation experiment shows that human judgments can be explained by two axes, ``general quality'' and ``faithfulness'' (PCA cumulative explained variance: 84\%), and that LLM-based evaluation tends to be biased toward general quality.We further propose KG-Check, which evaluates whether source facts expressed as subject--relation--object triplets are entailed by a translation, and observe a correlation with a rule-based ground-truth metric.