When studying how humans or machine learning models process language, real words carry heavy baggage. They arrive with pre-existing semantic meanings, historical contexts, and highly variable frequency biases. If a listener recognizes a word quickly, is it because the underlying phonetic structure is inherently easy to process, or simply because they encounter that word every day?
To isolate the core mechanics of phonology, perception, and language learning, researchers must strip away this lexical baggage. The solution is the systematic design of phonotactically matched pseudowords (or nonce words)—linguistic tokens that are entirely meaningless but structurally flawless.
The Mechanics of Phonotactic Matching
Designing a valid nonsense stimulus is more complex than randomly stringing characters together. A sequence like bglpx violates the structural rules (phonotactics) of English, causing the human brain or a language model to reject it immediately as an error. This triggers neural pathways associated with error detection rather than standard language processing.
To ensure that experimental stimuli are processed naturally, researchers construct pseudowords that mirror the statistical properties of a target language. As Baoguo Chen et al. demonstrate in their work on second-language acquisition (Rapid L2 Word Learning through High Constraint Sentence Context: An Event-Related Potential Study), this requires generating novel tokens that precisely mirror the syllabic structure and transition frequencies of real words. The resulting stimuli are completely unfamiliar yet entirely pronounceable and linguistically plausible.
[Random String] ----> bglpx (Violates phonotactics; triggers error handling)
[Pseudoword] ----> flimp (Obeys English phonotactics; isolates structural processing)
This structural matching can be tailored to isolate highly specific variables:
- Syllable Geometry: Toshiko Isei-Jaakkola et al. constructed nonsense words with rigid syllable configurations (such as CVCV or CVVCVV) to isolate how word structure independently influences vowel duration across different languages, completely independent of lexical meaning (A cross-linguistic study of vowel quantity in different word structures: Japanese, Finnish and Czech).
- Statistical Transitions: In computational linguistics, Gaël Le Godais et al. generated non-words adhering strictly to character-transition rules to evaluate whether character-level neural language models implicitly learn a mental dictionary or merely exploit surface-level n-gram statistics (Comparing Character-level Neural Language Models Using a Lexical Decision Task).
Achieving Lexical-Phonological Isolation
The primary power of this methodology lies in lexical-phonological isolation. By stripping away semantic meaning, researchers force the cognitive apparatus—whether a human brain or an artificial neural network—to rely entirely on bottom-up structural processing rather than top-down lexical retrieval.
For instance, when studying speech motor planning, the brain is simultaneously busy retrieving word meanings and planning articulation. Ina Kodrasi utilized pseudoword stimuli to ensure that observed electroencephalogram (EEG) signals reflected the pure construction of speech sounds rather than the retrieval of existing vocabulary entries (ERP Signals During Speech Articulation: Does Auditory Feedback Mask Other Ongoing Cognitive-motor Processes?).
In perceptual studies, this isolation prevents "lexical frequency bias," where a listener in a noisy environment guesses a word based on how common it is rather than what they actually heard. Chang Liu et al. leveraged phonotactically legal non-words to map the precise confusion matrices of Mandarin syllables under noise, ensuring the results reflected pure auditory processing rather than vocabulary familiarity (Perceptual confusability of Mandarin sounds, tones and syllables).
Probing the Boundaries of Grammar
Pseudowords do not just isolate; they can also be systematically varied along a gradient of structural legality to map the internal rules of a language system. By creating a spectrum from real words to "accidental gaps" (words that could exist but don't) to outright illegal sequences, researchers can pinpoint exactly where a cognitive system draws grammatical boundaries.
[ Real Word ] --------> [ Accidental Gap ] --------> [ Illegal Sequence ]
(e.g., cat) (e.g., lat) (e.g., lkat)
Yunjuan He et al. used this hierarchical design to show that language learners apply complex tone sandhi rules via generalized mental computation rather than rote memorization (Mechanisms of Tone Sandhi Rule Application by Non-Native Speakers).
This methodology also exposes how native language structures warp our perception of foreign sounds. Adriana Guevara-Rukoz et al. used recorded nonce words to demonstrate that Japanese speakers literally "hear" illusory vowels to break up illegal consonant clusters, proving that loanword adaptation happens at the perceptual level rather than through abstract grammatical post-processing (Speech Perception and Loanword Adaptations: The Case of Copy-Vowel Epenthesis).
Testing Abstract Biases in Artificial Intelligence
Recently, this psycholinguistic toolkit has been adapted to audit large-scale Vision-and-Language Models (VLMs). Humans exhibit strong sound-symbolic biases, famously exemplified by the bouba-kiki effect, where jagged shapes are instinctively paired with harsh, high-frequency sounds and rounded shapes with soft, low-frequency sounds.
To test whether models like CLIP possess similar cross-modal intuitions, researchers cannot use real words like "sharp" or "round," because the models have already seen millions of paired image-text examples containing those words during training. Instead, they synthesize novel, non-lexical labels based on specific phonetic properties like consonant voicing or vowel rounding.
Morris Alper et al. generated phonetic pseudowords to eliminate prior lexical knowledge and found evidence of sound symbolism implicitly embedded within VLM text embeddings (Kiki or Bouba? Sound Symbolism in Vision-and-Language Models). However, follow-up work by Tom Kouwenhoven et al. used rigorous psycholinguistic testing and attention mapping to show that these associations are highly inconsistent across different architectures, highlighting a fundamental divergence between human cognitive biases and machine representations (Cross-modal Associations in Vision and Language Models: Revisiting the Bouba-Kiki Effect).
Key Foundational Work
Adult and infant sensitivity to phonotactic features in spoken Japanese (2004) — contributed 1 concept elements to this direction.
Phoneme-based word activation in spoken-word recognition: evidence from Japanese school children (2004) — contributed 1 concept elements to this direction.
New evidence for prelexical phonological processing in word recognition (2001) — contributed 1 concept elements to this direction.