Meng, Varol & Barabási's 2024 'hidden citations' method measures Merton's obliteration by incorporation via topic modeling (LDA), not embedding cosine similarity

obliteration-by-incorporationhidden-citationstopic-modelingldacitation-analysisrobert-mertonbibliometricsembedding-false-friend

Meng, Varol & Barabási's "Hidden citations obscure true impact in science" (PNAS Nexus 3(5): pgae155, 2024) names Merton's 1968 concept obliteration by incorporation (OBI) explicitly as its object of study: "when a discovery or technique becomes common knowledge to such a degree that it does not warrant citation any longer, citations suffer from what Robert Merton in 1968 called 'obliteration by incorporation (OBI).'" The paper builds an unsupervised machine-learning pipeline — Latent Dirichlet Allocation (LDA), a topic model that infers distributions over word co-occurrence — to detect the "hidden citations" OBI causes: cases where a paper's influence on a later text is measurable even though no citation links them.

The paper's own method section distinguishes this technique from the embedding family: "In contrast with neural-network-based Word2Vec (29) or BERT (30), the LDA model is an unsupervised machine learning approach that is interpretable, allowing us to associate the outcomes of LDA with confidence levels through transparent probabilistic logic." LDA scores topic membership from word co-occurrence statistics, not from dense vector embeddings compared by cosine similarity — the mechanism Steck, Ekanadham & Kallus (2024) show can be arbitrary and the one the vault's "embedding false friend" genre is built on. This is the one place in the literature where "machine learning" and "obliteration by incorporation" genuinely meet on a shared object of study — and it meets on a different unsupervised technique than the embedding-cosine mechanism, a distinction the source paper itself takes care to draw.