Zero-shot pseudowords memorability via representational content analysis.

Area of Science:

• Cognitive Psychology
• Computational Linguistics
• Neuroscience

Background:

• Novel strings of letters (pseudowords) lack inherent meaning but can influence behavior.
• Distributional properties of words are known to affect memorability.

Purpose of the Study:

• To investigate if distributional determinants of word memorability extend to pseudowords.
• To determine if subword statistics predict the memorability of novel letter strings.

Main Methods:

• Utilized a word-embedding model to represent words and pseudowords in a vector space using character n-grams.
• Trained a ridge model on word memorability norms to map embeddings to recognition memory.
• Applied the trained model zero-shot to predict memorability for pseudowords, incorporating orthographic and frequency features.

Main Results:

• The model achieved strong performance in predicting memorability for known words.
• Zero-shot prediction of pseudoword memorability using distributional scores significantly improved baseline models.
• Mnemonic information was derived from subword statistics, independent of orthographic familiarity.

Conclusions:

• Distributional representations from subword statistics contain mnemonic information.
• Novel strings are processed within a shared representational space shaped by language experience.
• Memorability is an intrinsic attribute predictable from representational information, even without established meanings.