Random forests, sound symbolism and Pokémon evolution

Area of Science:

• Computational Linguistics
• Artificial Intelligence
• Cognitive Science

Background:

• Sound symbolism, the idea that sounds in words correlate with meaning, is a recognized linguistic phenomenon.
• Machine learning models can potentially learn and apply these sound-meaning correspondences.
• Pokémon names offer a unique dataset for studying sound symbolism due to their varied origins and evolutionary transformations.

Purpose of the Study:

• To construct and train machine learning algorithms (random forests) to classify Pokémon based on sound symbolism.
• To compare the classification accuracy of these algorithms against human participants.
• To investigate the role of name length as a feature in sound-meaning classification.

Main Methods:

• Training random forests on names of Japanese, Chinese, and Korean Pokémon to predict pre-evolution and post-evolution status.
• Conducting an elicitation experiment where Japanese participants named novel Pokémon.
• Comparing machine learning performance against human classification of elicited Pokémon names, incorporating name length as a feature and using a novel cross-validation method to address overfitting.

Main Results:

• Machine learning models demonstrated efficient learning of systematic sound-meaning correspondence patterns.
• The trained random forests achieved higher classification accuracy than human participants in distinguishing Pokémon evolutionary stages.
• A novel cross-validation method was developed to resolve overfitting issues identified in initial experiments.

Conclusions:

• Machine learning algorithms, particularly random forests, are adept at identifying and utilizing sound symbolism for classification tasks.
• AI models can surpass human performance in recognizing subtle sound-meaning relationships within linguistic data.
• The study validates the efficacy of sound symbolism in predicting semantic categories, even in artificial linguistic systems like Pokémon names.