A Controlled Reanalysis of KIDO: Text Variants, Image-Model Configurations, and Cross-Modal Score Discrepancy

Abstract

KIDO pairs children’s prompted happy and sad drawings with written reflections, creating a multimodal setting in which the text and image channels provide very different evidence for the same labels. We reanalyze KIDO by comparing Turkish, original English, and LLM-generated Turkish-grounded English text; replicating image-model configurations; and measuring text–image score discrepancy on 1630 drawings from 815 children. Turkish-grounded English is numerically higher than original English across emotion, gender, and educational-level classification under the controlled English protocol. For emotion, the respective three-seed macro-F1 scores are 0.9808±0.0004 and 0.9798±0.0034; however, the paired ensemble difference is small, and its 95% child-cluster bootstrap interval includes zero. A blinded, tail-enriched bilingual audit of 120 child-distinct pairs provides source-grounded support for the generated text: 88.3% of ratings receive adequacy scores of 4–5, while major meaning errors occur in 7.5% of ratings and are confined to the deliberately sampled low-similarity tail. Text performance also depends strongly on explicit emotion cues: such cues occur in approximately two-thirds of test reflections, and masking them at evaluation reduces macro-F1 by 0.054–0.073, while performance remains above 0.90. Image-emotion models achieve a substantially lower mean macro-F1, ranging from 0.7683 to 0.7867 across replicated DenseNet121 and ConvNeXt configurations. Cross-modal score differences remain substantial: the Multimodal Discrepancy Index, equivalent to total variation in this binary task, ranges from 0.2300 to 0.2753 across image ensembles. These results show that text preparation, lexical cues, and image-model configuration materially affect conclusions drawn from KIDO and support reporting channel-specific evidence before interpreting multimodal agreement or fusion.