Imagine you inherit a filing cabinet full of survey questions — hundreds of personality items like "I am the life of the party" and "I sympathize with others' feelings" — but zero completed surveys. Traditionally you're stuck: factor analysis requires people to actually answer the questions. This paper asks whether you can skip the people entirely by treating each item's meaning as a proxy for how respondents would cluster their answers. The mechanism is straightforward. Each survey item gets converted into a high-dimensional embedding vector by a language model (here, Qwen3-Embedding-8B). You then compute cosine similarities between all item pairs, forming a correlation-like matrix. Standard factor-analytic machinery — eigendecomposition, rotation, retention criteria — runs on that matrix instead of on response data. The R package semanticfa wraps this pipeline end-to-end: factor extraction, interpretation, redundancy flagging, short-form construction, and new-item screening. The worked example is the 50-item IPIP Big Five Factor Markers, the workhorse personality inventory. Five factors emerge semantically, and three of them map cleanly onto single Big Five domains. Two factors blend Extraversion with Agreeableness items — meaning the embedding space doesn't fully distinguish "outgoing" from "warm." Factor congruence with human-response factors ranges from .92 (Openness) to .52 (Agreeableness). That .52 is barely above the conventional threshold for factor similarity (.85 is typically considered good), which the authors acknowledge. The most interesting diagnostic is the correlation between semantic similarity and response similarity at the item-pair level: r = .46 across 1,225 item pairs, benchmarked against 874,434 respondents from the Open-Source Psychometrics Project. That's meaningful but far from deterministic — meaning similarity explains roughly 21% of the variance in how people actually co-endorse items. Reverse-keyed items expose a genuine limitation: items phrased as opposites are semantically similar (they're about the same concept), but respondents answer them oppositely. Sign-flipping the embeddings doesn't fix this and actually makes things worse for reverse-scored items. The paper's most consequential conceptual move is proposing "semantic-behavioral coherence" as a measurable property of any scale–embedding-model pair. This reframes the question from "does semantic FA replace response FA" (the authors explicitly say no) to "how much of a scale's structure is already baked into the language of its items?" That's a genuinely useful diagnostic for scale developers, especially early in item writing when you have items but no respondents yet. Integrity is above average for a methods paper: 874K respondents as the human baseline, public data and code on Zenodo, a 96-page manuscript with 14 figures that doesn't hide the Agreeableness problem. The main weakness is that only one embedding model (Qwen3-Embedding-8B) and one personality scale are tested. The authors frame this as a worked example, not a definitive validation, but readers should note the generalizability gap. The practical upshot: if you're building a new psychometric scale and want a fast, free check on whether your items cluster the way you think they do before running an expensive data collection, semanticfa gives you that. It's a complement to human data, not a replacement — and the authors are admirably clear about the boundary.
