Wolfram Computation Meets Knowledge

Wolfram Summer School


Lawrence Temlock

Technology and Innovation

Class of 2018


Larry Temlock is a capital markets finance expert with many years of experience in non-US debt and alternative investment products. Prior to his current role as CFO /cofounder at Sun Exchange, he was Managing Director, Capital Markets at Daiwa Securities Hong Kong, and served in a variety of positions at JP Morgan, Merrill Lynch and UBS Securities. Larry has an MSc in financial engineering and a BA in political science, both from Columbia University. At Sun Exchange, Larry is working to promote global investment in emerging markets’ solar projects.

Computational Essay: Improving Time Series Model Fit With Variance Modeling

Project: Computing Text Reading Levels

Project Overview

Classical readability indices use counts of words, syllables, and word frequencies. Recent systems such as Lexile leverage reading research and psychometric theory to score and match readers and books. Text context, subject matter and other subtle characteristics don’t generally factor into readability indices commonly used in education, state policy, the military, and the media. In this project, three approaches were used to test if a neural network can be trained to generate meaningful readability scores. 97,000 text samples, 50-500 characters long, were collected and targeted with Lexile L-scores ranging from 430L (elementary) to 1350L (high school). In Trial 1, text data was assigned 100-column feature vectors from a GloVe net model*, and input to a network embedding layer. Trial 2 input a 1-column word frequency vector; Trial 3 combined these two elements with NetEncoding. The common network architecture assembled two gated recurrent layers, followed by a sequence last layer and a linear layer. The three network training trials were unable to produce networks capable of meaningfully scoring text reading level. The result of Trial 1 was overfitting. Trials 2 and 3 both failed to reduce the output from the loss function. Random sample results from the networks compared unfavorably with those generated by the Dale-Challs Raw Score and Automated Readability Index.