[4Yin-A] Poster 6

[4Yin-A-17] Rethinking Diversity in News Comment Ranking via Mainstream Separation and Constructive Minority Exposure

〇Zuofan Fang¹, Kota Tsubouchi², Tooru Shimizu², Nobuhiko Nishio¹ (1. Ritsumei University, 2. LY Corporation)

Keywords:

Diversity Ranking,Deep Clustering,Responsible AI

Diversifying news comments can broaden readers' perspectives and mitigate the dominance of a single narrative in online deliberation. However, comment distributions are characteristically long-tailed: a dense mainstream contains multiple closely situated rationale modes, while low-quality or off-topic outliers appear sparsely. Under this regime, off-the-shelf clustering and dissimilarity-based reranking are prone to either failing to separate mainstream sub-peaks or inadvertently promoting fragmented outliers. We introduce M-DEC, a majority-aware deep embedded clustering framework that modifies self-training with an outlier-concentration term, a mainstream compactness regularizer, and a soft upper-capacity constraint, followed by representative selection through an entropy-regularized coverage objective. Across 30 news articles, M-DEC yields stronger outlier concentration (lower entropy, higher max-recall) and generates top-10 slates that are rated as more diverse in human evaluation without degrading perceived quality. Overall, these results underscore the value of structured mainstream separation and exposure balancing for diversifying news comment ranking while avoiding outlier dominance.