Honglong Yang

Honglong Yang

Auto-generated profile, based on publicly available information

Honglong Yang


Research Focus & Metrics

Academic Career Length

1Year (2026-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
20 October 2026 in SSRN Electronic Journal
Verified
Preprint
Open access
QCQiuchan Chen
Qiuchan Chen
WXWeikang Xie
Weikang Xie
MHMenglu Han
Menglu Han
HYHonglong Yang
Honglong Yang
ZLZiyi Leng
Ziyi Leng
CXCai Xinyuan
Cai Xinyuan
Narrative understanding requires going beyond local content to capture the overall development and organization of a story. Emotion dynamics and narrative organization reflect how a story is experienced and constructed, respectively, and constitute two interrelated dimensions of multi-image narrative understanding. However, existing work typically examines them in separate task and evaluation settings, leaving their relationship within the same narrative underexplored. In this work, we focus on multi-image narrative understanding, using picture books as a representative setting, and formulate two interrelated tasks: reader-side affective experience and author-side narrative organization. We introduce Emo-FOLIO, a benchmark comprising 234 picture books drawn from Chinese- and English-language sources and 4,173 image--text pages, with annotations and evaluation at the page, transition, and stage levels. We further propose Multilevel Organization of Story-Aware Information for Cross-level Reasoning (MOSAIC), a structured inference framework that constructs a reusable narrative scaffold from multimodal evidence, jointly infers narrative stages and their communicative intents, and conditions local emotion estimation on the predicted structure while preserving cross-window continuity. Experiments with eight multimodal backbones reveal marked capability divergence: strong page-level recognition does not reliably translate into cross-page trend estimation or stage-level understanding. Across four representative backbones, MOSAIC improves over direct whole-book inference by averages of 2.6, 7.2, and 6.8 percentage points in page-level multilabel emotion recognition, cross-page emotion trend estimation, and joint stage-and-intent recovery, respectively. These results show that explicit organization of narrative evidence and structure-guided inference provide an effective way to connect local and global understanding in sequential multimodal narratives.

5 October 2026 in SSRN Electronic Journal
Verified
Preprint
Open access
MHMenglu Han
Menglu Han
KWKang Wang
Kang Wang
QCQiuchan Chen
Qiuchan Chen
HYHonglong Yang
Honglong Yang
KZKuo Zhang
Kuo Zhang
CXCai Xinyuan
Cai Xinyuan
Visual affect analysis typically predicts a static emotion label for an image. However, emotional responses to artworks often evolve as viewers move from rapid visual impression to detailed observation and contextual interpretation. To model this evolution, we formulate artistic image affective trajectory prediction, which predicts a three-stage sequence of emotion labels for the same artwork as observational and contextual evidence progressively increases. We construct ArtArc, an artistic image affective trajectory benchmark with 3,210 artworks. Each sample includes a title, a description, and human-annotated affective trajectories, covering over 13K stage-level emotion-label entries. ArtArc allows both single- and dual-label annotations to reflect the ambiguity of artistic affect. We further propose Artistic-image Affective Trajectory Inference with Dynamic Evidence (ArtTIDE), a training-free structured reasoning framework that guides multimodal large language models to update affective judgments through global visual evidence, visual-semantic evidence, and textual-contextual evidence. Experiments on five representative multimodal models show that the proposed benchmark remains challenging, especially for fine-grained emotion recognition. ArtTIDE consistently improves stage-level prediction, trajectory matching, and context-aware updating over the base setting, increasing the three-stage soft matching rate and contextual supplementation gain rate by 9.2 and 9.0 percentage points, respectively, while reducing the context regression rate by 5.8 percentage points. Ablation experiments and case studies further indicate that the structured integration of visual, semantic, and textual evidence is important for modeling affective trajectories in artistic images.