Trend Analysis in AI Research over time Using NLP Techniques

Abstract

The dramatic rise in the number of publications in machine learning related studies poses a challenge for companies and new researchers when they want to focus their resources effectively. This thesis aims to provide an automatic pipeline to extract the most relevant trends in the machine learning field. I applied unsupervised topic modeling methods to discover research trends from full NIPS conference papers from 1987 to 2018. By comparing the Latent Dirichlet Allocation (LDA) topic model with a model utilizing semantic word vectors (sHDP), it was shown that the LDA performed better in both quality and coherence. Using the LDA, 50 topics were extracted and interpreted to match the key concepts in the conference publications. The results revealed three distinct eras in the NIPS history as well as the steady shift away from the neural information processing roots towards deep learning.