Manuscript in preparation
Does Word Memorability Generalize Across Languages?
Cross-linguistic memorability in English and Hindi
Working title: Memorability Emerges from Language-Specific Semantic Structure. Prajneya Kumar, Weizhen Xie, Wilma A. Bainbridge, and Vishnu Sreekumar.
- Participants
- 763 eligible native Hindi speakers, tested online
- Stimuli
- 221 Hindi words derived from an existing English memorability study
- Task
- Paired-associate learning, an arithmetic distractor, then cued recall
- Analyses
- Cross-language memorability correspondence, split-half reliability, word frequency, and graph measures of semantic networks
Research question
Some words are consistently more memorable than others across observers. Are these memorability differences stable across languages, or do they depend on the linguistic and semantic structures through which words are represented?
Methods
I designed and ran a cued-recall memory experiment with 763 eligible native Hindi speakers, using 221 Hindi words derived from an existing English memorability study. Participants studied word pairs, completed an arithmetic distractor task, and then recalled each target word from its cue. I built the custom JavaScript platform used to run the experiment remotely, including Hindi response entry and recall scoring.
We tested whether item-level memorability scores generalize between English and Hindi, and estimated within-language split-half reliability. To relate memorability to representation, I built language-specific semantic networks from distributional word embeddings and computed graph measures for each word: degree centrality, PageRank, clustering coefficient, and average path length. Word frequency was analyzed alongside these measures. Hypotheses and analyses were preregistered on OSF.
Findings so far
Cross-language generalization
Item-level English and Hindi memorability scores were not significantly correlated in our dataset (Spearman’s ρ ≈ −0.047, p = 0.487).
Semantic representations
We examined relationships between memorability and properties of language-specific semantic networks, including word frequency, centrality, and graph structure. Some Hindi network measures are strongly correlated with word frequency, and their associations with memorability become weaker when frequency is controlled.
Interpretation
These results challenge a simple assumption that a word’s memorability should remain stable across translation, and motivate a closer look at the representational context in which memories are formed.
What these results do not show
- They do not show that language causes differences in memorability.
- They do not show that semantic network structure causes memorability, or that network centrality independently explains Hindi memorability once frequency is accounted for.
- A non-significant correlation in one English–Hindi sample does not show that no memorability effects transfer between languages, and it does not generalize to every language pair.
Earlier presentations
- Word Memorability: A Cross-Cultural Perspective. Poster, 9th Annual Conference of Cognitive Science (ACCS9), IIT Delhi, 2022. Poster PDF (IIIT Hyderabad showcase copy)
- Analysing Memorability Features of Multilingual Analogous Words in a Semantic Network. Poster, Context and Episodic Memory Symposium – International (CEMSi), University of Melbourne, 2021.
- Word Memorability: A Cross-Language Semantic Network-Based Perspective. Poster, IIIT Hyderabad R&D Showcase, 2022–2024.