Hidden Cognate Traps May Be Skewing Bilingual Word Recognition Studies

Every word in a bilingual mind carries invisible baggage. When a Spanish-English speaker sees the word “hospital,” their brain does not simply retrieve the English meaning; it simultaneously stirs the Spanish “hospital,” a near-identical twin that shares spelling, sound, and sense. These cross-language twins, known as cognates, have long been recognized as a source of facilitation in bilingual word processing. Now, a new tutorial published in Behavior Research Methods argues that the field has been overlooking a subtler and potentially more insidious version of this problem, one that could be quietly distorting the results of countless second-language experiments.

Kelly Kendro and Alexander J. Holmberg of Northern Arizona University focus their attention not on the target words themselves, but on the words that surround them in experimental tasks. In semantic priming studies, researchers present participants with a prime word followed by a target, measuring whether the prime speeds recognition of the target. When studying second-language vocabulary, primes are often drawn from the participant’s first language while targets come from the second. The conventional wisdom holds that if you control for obvious confounds, such as whether the prime and target look or sound alike across languages, any remaining speedup reflects genuine semantic priming. Kendro and Holmberg contend that this logic has a hidden flaw: the semantic relatives of a target word can themselves be cognates, and nobody has been systematically checking.

The mechanism at play is spreading activation, a foundational concept in cognitive psychology dating back to the classic network models of the 1970s. When a person processes a word, activation radiates outward to semantically related concepts, pre-activating them and making subsequent recognition faster. In a bilingual lexicon, however, this activation does not respect language boundaries. Decades of research, from early work on cognate recognition to eye-tracking studies showing covert activation of the non-target language, have established that both of a bilingual’s lexicons remain active even when only one language is in use. If a prime word in the first language is a cognate of a word that is semantically related to the second-language target, it can trigger a cascade of activation that boosts the target through a back door, inflating apparent priming effects for reasons that have nothing to do with the semantic relationship the researcher intended to measure.

Consider a concrete scenario. An English-Spanish bilingual sees the Spanish prime “médico” (doctor) followed by the English target “nurse.” Suppose “médico” happens to resemble an English cognate that is itself semantically linked to “nurse.” The prime activates its English look-alike, that look-alike spreads activation to related English concepts, and the target benefits from a head start that masquerades as crosslinguistic semantic priming. The authors note that prior research has already shown cognate synonyms can complicate word learning and that cognate status shapes facilitation in vocabulary assessment, yet stimulus-selection procedures in L2 priming research have rarely screened for this dimension in the semantic neighborhood of the target.

To close this gap, the tutorial lays out a seven-step procedure that researchers can apply to any crosslinguistic target pair. The workflow begins by establishing the semantic similarity between the prime and the target, since priming studies fundamentally depend on related pairs being genuinely related. The authors suggest embedding-based measures, including sentence-transformer models derived from the BERT architecture, while noting that researchers who prefer distributional approaches such as word2vec can substitute their own tools. From there, the procedure moves to the heart of the problem: identifying whether the prime or its translations share cognate status with words in the target’s semantic neighborhood.

Quantifying cognate status requires measuring form similarity across languages, and the tutorial draws on well-established computational machinery. Orthographic overlap can be estimated with Levenshtein edit distance, the count of character insertions, deletions, and substitutions needed to transform one string into another, an approach with a long history in predicting intelligibility and perceived linguistic distance between languages. Phonological similarity can be assessed by converting words to phonetic transcriptions using grapheme-to-phoneme tools such as Epitran before applying distance measures, which matters because two words may look different but sound alike. The authors also recommend consulting large-scale cognate databases such as CogNet, which catalogs etymologically related word pairs across thousands of languages, and crosslinguistic neighborhood resources like CLEARPOND that quantify how many similar-sounding and similar-looking words surround a given item in each language.

Etymology introduces its own complications. Words can descend from a common ancestor yet drift so far apart in meaning that they no longer co-activate each other’s lexical representations, a phenomenon the authors flag with reference to prior work on crosslinguistic lexical influence. Conversely, loanwords borrowed directly between languages may look alike without deep shared history. The tutorial therefore treats etymological relatedness as one signal among several rather than a verdict in itself, and the accompanying database-cleaning scripts accommodate languages where certain checks, such as gender agreement in German, need modification or omission.

To make the procedure practical rather than merely prescriptive, Kendro and Holmberg provide working Python code that automates much of the screening. The pipeline integrates the deep-translator module, which interfaces with Google Translate, chosen for its broad language coverage and ease of integration, though the authors explicitly welcome researchers to substitute alternative translation services and allow manual entry when multiple translations exist. Word frequency estimates, which matter because frequency modulates both lexical access and cognate facilitation, are drawn from the wordfreq library, which aggregates frequency data from many corpora. The code then computes semantic similarity, form similarity, and cognate status for each candidate pair, flagging items where hidden cognate activation could contaminate the intended manipulation. The authors demonstrate the full workflow on a sample set of English-Spanish items, showing how pairs that appear clean on the surface can fail screening once the semantic neighborhood is examined.

The broader significance of this work extends beyond any single experiment. Large lexical databases such as the English Lexicon Project, the British Lexicon Project, and the Spanish SPALEX database have standardized how researchers control for word frequency, length, and prevalence in monolingual studies, and word prevalence norms covering tens of thousands of English lemmas have sharpened estimates of how widely known a word actually is. But crosslinguistic research has lacked an equivalent guardrail for the interactions between a participant’s two lexicons. Studies of individual differences in bilingual word recognition have shown that experiential factors and word frequency both modulate cross-language priming, and language dominance shapes how orthographic and phonological similarity play out in cognate recognition. If stimulus sets systematically contain unrecognized cognate confounds, effect sizes across the literature could be noisy in ways that no amount of statistical control after the fact can fully repair.

The tutorial arrives at a moment when replication concerns have pushed psycholinguistics toward more rigorous stimulus validation, and its message is disarmingly simple: the words you did not choose can matter as much as the words you did. By screening for semantic similarity, orthographic similarity, phonological similarity, and etymological relatedness in a single automated workflow, researchers can rule out the possibility that a cognate synonym lurking in the prime’s neighborhood is doing the work that the experiment claims to measure. The code and materials are freely available through the Open Science Framework, lowering the barrier for labs working in any language pair to audit their stimuli. For a field whose theoretical claims about the bilingual mind rest on millisecond differences in recognition times, ensuring that those milliseconds reflect the process under study, rather than an accidental echo of a hidden twin word, is a small step in stimulus design with an outsized payoff for the validity of the science.

Subject of Research: Hidden cognate confounds in second-language lexical retrieval and semantic priming stimulus design

Article Title: Identifying hidden cognate effects in L2 lexical retrieval tasks: A tutorial for stimulus design

Article References: Kendro, K., & Holmberg, A. J. (2026). Identifying hidden cognate effects in L2 lexical retrieval tasks: A tutorial for stimulus design. Behavior Research Methods, 58(11), Article 324. https://doi.org/10.3758/s13428-026-03210-x

Image Credits: AI Generated

DOI: 10.3758/s13428-026-03210-x

Keywords: bilingualism, cognates, semantic priming, second language, lexical retrieval, psycholinguistics, stimulus design, spreading activation, orthographic similarity, Levenshtein distance, word recognition, Behavior Research Methods