Embedding cosine similarity mostly tracks length, not AI homogenization
Original title: Embedding Cosine Measures Text Length Before It Measures Homogenization
Length controls reproduced most of a reported post-2022 political asymmetry, and corrected measures show the camps stopped converging before ChatGPT.
The paper argues that mean pairwise cosine similarity of mean-pooled encoder embeddings largely measures text length when the encoder was not trained for sentence similarity. On 6.06 million Reddit comments, truncating text and length-matched placebos account for most of an apparent conservative shift toward homogeneity. On 575,000 Stack Exchange answers, a reported post-ChatGPT technical homogenization is fully reproduced by a length placebo. A census of 31 studies locates the risk in anisotropic encoders with unequal lengths or time series.
- Truncating comments from 10 to 150 words raised mean cosine by eight times the reported break.
- An equal-length random swap reproduced 81 percent of the asymmetry; a fixed-content simulation reproduced 78 percent.
- The leading principal direction carried 99.9 percent of raw DeBERTa homogeneity.
- After length correction the asymmetry disappeared under every encoder.
- Cross-camp replies engaged their targets more closely than same-camp replies.
From the original article
1 min read at arxiv.orgMean pairwise cosine similarity between transformer embeddings is a standard measure of whether online discourse is converging under generative AI. When the embeddings are mean-pooled outputs of an encoder never trained for sentence similarity, this measure mostly tracks text length. In 6.
By Fengming Liu · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.