Source
Grindrod, Porter and Hansen (2024), LIT-208 (arXiv 2405.12084 v2; also a chapter in Cappelen and Sterken (eds.), OUP, not read), read in full in NOTE-158.
What was actually shown
The objection and the reply. The instability objection to meaning holism (Fodor and Lepore 1992; Pagin’s “total change thesis”) says any change ripples through the system, so speaker and hearer never share meanings. The paper takes holism as determination by relations (Dresner 2012), not priority, and reads meaning similarity off nearest neighbours, which survive rotation and change of dimensionality and can be compared across spaces, as in Hamilton et al. (2016). Fodor and Lepore’s dimension-selection objection lapses, since distributional dimensions are not semantic properties (§4; C2, C3 moderate).
Count model (§4, pp. 12–17). A ±10-word window on Stein’s The Making of Americans (about 522k tokens). “know” has neighbours understand 0.95231, it 0.94286, see 0.93759, think 0.93293 (Table 4). After adding Hemingway’s “A Clean, Well-Lighted Place” (under 1,500 words) all ten values move in the fourth or fifth decimal place (e.g. understand 0.95231 → 0.95284) and the order is unchanged (Table 6). “glass”, 2 uses in Stein and 5 in Hemingway, changes from chandeliers, prisms, brocade to saucer, waiter, table (Table 5).
Word2vec model (§4, pp. 18–20). CBOW on 372,194 Wikipedia articles (about 143M words), then with all 1,763 SEP articles added (about 21M words). “atomism”, 9 uses in Wikipedia and 662 in the SEP, reorders and gains neighbours (Table 7). “can” keeps its neighbours, with values changing by at most 0.003 and a swap at positions 9–10 (Table 8).
The stronger thesis (§5). The residual instability is a virtue: it registers idiolect and diachronic change (C8, weak, by appeal to Hamilton et al.).
What could have come out otherwise. “know” or “can” could have lost neighbours under the addition, or “glass” and “atomism” could have kept theirs. Neither happened in the one run shown.
What this does not say
- That resistance to change is proportional to frequency. The abstract says “roughly proportional”; the body shows four words and one before-and-after pair per model (C6, weak). That models grow more stable with size rests on a citation to Antoniak and Mimno (2018), not on the authors’ two corpus sizes (C7).
- That the contrasts exceed seed noise. The authors raise run-to-run randomness and report no repeated runs, so “can“‘s swap at positions 9–10 cannot be separated from noise.
- That relative distances are preserved. The definition of differential stability is relative distances, but the cosine values do change; what the tables show preserved is the ranked neighbour list. The paper does not reconcile the two (NOTE-158, Corrections).
- That transformer models behave the same way. Asserted (C10), with no experiment on contextual embeddings, where a word has no single neighbourhood.
- That distributional models capture meaning, or what similarity suffices for communication. The first is bracketed (§2). The second is answered informally, as graded success or an untested empirical threshold (C9).
- Anything new for machine-learning practice. Comparing embedding spaces by neighbour sets, and expecting rare words to be unstable, is already practice, with Antoniak and Mimno as its operational source. NOTE-158 recommends no anthology filing.
Connections
- The paper acknowledges Brandom’s inferentialism as a holist alternative and does not adopt it (fn. 15). Brandom’s account of content as inferential role is THEORY-665; LIT-106 (Marabini on Brandom) is read only in a seeded skim.
- Saussure’s value fixed by relations among coexisting terms (THEORY-159) is the structuralist form of the same relational view of meaning. This paper does not cite it, and no relation is declared.