Replicating Human Semantic Fluency

Artificial Intelligence
Winter 2019
CS 441/541

Authors

Alejandro Espinoza (azespinoza)
Michelle Duer (mkduer)
Alli Wong (wonal)
Carson Cook (cjc77)
Jacob Collins (jacobmcollins)

Project Overview

Inspired by the work done in "Predicting and Explaining Human Semantic Search in a Cognitive Model" [1], our team decided to create a semantic network off of word embeddings to simulate human semantic fluency. The various steps in our project included:

• NLP Pre-Processing: cleaning the corpus through tokenization and lemmatization
• Generating word embeddings using Word2Vec's skip-gram method
• Learning the embeddings and building a graph network where the edges exist based on cosine similarities above a specific threshold
• Algorithms for traversing the network and attempting to simulate human fluency
  • Weighted Random Walk
  • Simulated Annealing
  • Random Start/Re-start Hill Climbing
• Visualizations
  • Networkx plots for visualizing the network and similarities
  • t-SNE plot for starting corpus word embeddings and similar clusters
  • SVD and t-SNE plot for comparing image subspaces and corpus subspaces
  • Seaborn lineplot for algorithm walks
  • Seaborn barplot for total quantitative measurements
• Experimentation through parameter fine-tuning
• Conclusion that the random start/re-start hill climbing algorithm best simulated human semantic fluency
• Additional projects for visualizing and comparing subspaces of data

Test Datasets

Shakespeare: http://www.gutenberg.org/files/100/100-h/100-h.htm
Fairy Tales: https://www.gutenberg.org/files/19734/19734-h/19734-h.htm
Wine: https://www.kaggle.com/zynicide/wine-reviews

Resource

[1] F. Miscevic, A. Nematzadeh and S. Stevenson, "Predicting and Explaining Human Semantic Search in a Cognitive Model," Nov. 29, 2017. [Online]. Available: https://arxiv.org/pdf/1711.11125.pdf.