Predicting Lexical Priming Effects from Distributional Semantic Similarities: A Replication with Extension

Area of Science:

• Cognitive Psychology
• Computational Linguistics
• Natural Language Processing

Background:

• Distributional semantics models quantify word similarity based on co-occurrence in large text corpora.
• Latent Semantic Analysis (LSA) and Hyperspace Analog to Language (HAL) are prominent models of distributional semantics.
• Lexical priming effects demonstrate that encountering a word influences the processing of subsequently encountered related words.

Purpose of the Study:

• To replicate and extend findings on the predictive power of LSA and HAL for lexical priming.
• To investigate the efficacy of different distributional semantics models in predicting priming effects.
• To improve upon previous methodologies by using larger corpora and refined experimental controls.

Main Methods:

• Two experiments were conducted using a pseudo-random item generation method.
• Systematic control of word similarities based on LSA and HAL cosine similarities.
• Utilized semantic spaces derived from significantly larger corpora compared to prior studies.

Main Results:

• Experiment 1 showed a significant effect of HAL cosine similarities on lexical decision times.
• Experiment 2 revealed significant effects for both LSA and HAL cosine similarities.
• Pooled data analysis indicated HAL cosines as a superior predictor of priming effects over LSA cosines.

Conclusions:

• Distributional semantic similarity measures, particularly HAL, effectively predict lexical priming.
• The findings support the utility of computational models in understanding human lexical processing.
• Methodological enhancements and larger corpora strengthen the reliability of these predictions.