Vector Traces of What Goes Unsaid

Indirect contextual signals in public comments and classical narratives When people share an experience, can their language carry traces of it even outside conversations that name it? We compared groups of public comments and literary passages along fixed directions in text-embedding space. After removing entire threads that explicitly discussed temperature, 80,723 comments from Toronto’s public forum still scored higher on cold days than on warmer days in two models. The contrast varied by year. In online discussions surrounding the September 11 attacks, the strongest time split along an attack-related direction fell on September 11 after exclusion of direct mentions, and on September 12 after exclusion of event-related threads. The change reversed sign. After sentences describing birds were removed from solar and comparison passages, solar passages scored higher along a bird direction in three initial literary samples. Only two of three new samples showed the same sign. In an ancillary test on 340 Chinese greetings, subtracting a shared greeting vector reduced agreement with annotators on the primary communicative purpose from 179 cases to 162. These results locate differences in particular corpora after specified exclusions. Their interpretation depends on what a filter removes, how texts are grouped, and whether the difference recurs in new material. 未言之事的向量痕迹 · 公开留言与古典叙事中的间接背景线索 当人们没有把同一件事说出口,汇集他们的言语,还能测到与它有关的向量痕迹吗?本文在公开语料中删去直接表述,沿预先确定的语义方向比较剩余文本。多伦多论坛排除明确谈冷热的讨论串后,80,723 条留言在两种模型下仍表现为冷日组偏冷,34 对日期中分别有 22 对和 21 对同向,分年比较出现反向。9·11 前后,只删直接提及事件的留言时,袭击方向的最大分界在 9 月 11 日向上。再去掉相关讨论串,最大分界在 9 月 12 日向下。在中国、希腊和印度的文学材料中,两边都删掉直接鸟句后,首批三组太阳摘段都比对照偏向“鸟”,换作品后只有两组同向。另对中文问候比较三种读法,从上下文向量中减去共同问候参考并未稳定改善用意识别。这些结果刻画了特定语料里可测的间接差异,也显示其解释会随留下哪些发言、选择哪些日期和作品而改变。 Files. English and Simplified Chinese editions of the same preprint. The ZIP contains both manuscripts, editable sources, aggregate data and an offline verification script. Author contribution and AI use. Ye Fang proposed the research question, directed the group comparisons, specified symmetric removal of explicit discussions, requested tests on new material and shaped the interpretation. OpenAI Codex assisted with acquisition and analysis scripts, synthetic calibration material, text screening, calculation checks, drafting and bilingual editing. The reported embeddings were produced by OpenAI API models. Two blinded AI annotation passes supplied the greeting labels. The paper does not claim independent human annotation of those labels.