A generative approach to extrapolate word concreteness ratings.

Area of Science:

• Psycholinguistics
• Computational Linguistics
• Cognitive Neuroscience

Background:

• Normative ratings of lexicosemantic properties are crucial for psycholinguistic and clinical research.
• Collecting these ratings (e.g., word concreteness, valence, age-of-acquisition) is labor-intensive.
• Existing resources for Chinese words are limited, hindering research.

Purpose of the Study:

• To computationally expand concreteness ratings for simplified Chinese words.
• To develop a generative model that predicts both the mean concreteness (con.mean) and variability (con.var) of word ratings.
• To validate the model's predictions against human ratings and assess its contribution to understanding Chinese lexical processing.

Main Methods:

• Utilized a mixture density network (MDN), a generative approach.
• Trained the MDN on different word embeddings to predict the probability density of trial-level concreteness ratings.
• Generated extrapolated concreteness ratings (con.mean and con.var) for over 78,000 Chinese words.

Main Results:

• The MDN successfully predicted both the central tendency and variability of human concreteness ratings.
• The model's predictions closely aligned with human judgments and captured key representational features.
• The study demonstrated the concreteness effect on event-related potentials in Chinese visual word recognition.

Conclusions:

• The developed MDN provides a scalable method for generating reliable word concreteness ratings.
• The released dataset enhances resources for Chinese psycholinguistic research.
• Computational approaches can effectively expand and validate lexicosemantic properties for large word sets.