Higher explanation-necessity scores are associated with longer written explanations of autonomous driving scenes

Explanations of automated-vehicle behavior may help occupants understand what prompted a maneuver, but how much an explanation should say remains an open design question. Retrospective annotation of recorded driving scenes can supply candidate content. We analyzed 1,103 training and test records from the release, matching the source-described corpus size; each paired human-written text with an aggregate perceived explanation-necessity score.

Explanations contained 9,541 whitespace-delimited tokens (mean 8.65; range 2–20) before lexical preprocessing. Higher scores were associated with longer explanations (Pearson’s \(r=0.374\); template-cluster bootstrap 95% interval [0.266, 0.448]), including after score-1.0 observations were excluded. We call this corpus-level length association progressive elaboration.

Bounded-model comparisons did not identify a unique latent-category count. For groupwise summaries, we used four descriptive partitions: three below-ceiling regions and 59 records scored 1.0. Record-weighted and unique-template analyses showed higher within-partition lexical distinctiveness in upper than in lower partitions.

In template-grouped cross-validation of these exploratory partitions, the unweighted text model exceeded token-count and stratified-dummy baselines in balanced accuracy (0.382 versus 0.327 and 0.247). The 1,280-record rerun preserved the score—length, lexical, and predictive findings and left the latent-category count unresolved. These results motivate prospective studies of how annotation-derived content should be selected, compressed, and delivered in real time.

We appreciate the high-performance GPU computing support of HPC-AI Open Infrastructure via GIST SCENT. The work is the authors’ own, and the funders had no role in study design, analysis, or manuscript preparation. This work was supported in part by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (RS-2025-16902996, RS-2024-00343397), and in part by the Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. 2019-0-01842, Artificial Intelligence Graduate School Program (GIST)).

Department of AI, College of AI, Gwangju Institute of Science and Technology, Gwangju, Republic of Korea Gwangbin Kim, Dohyeon Yeo, Yumin Kang, Minwoo Oh & SeungJun Kim Correspondence to Gwangbin Kim or SeungJun Kim. The authors declare no competing interests. Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.