The Art of Public Speaking: Machine Learning and Natural Language Processing To Analyze TED Talks

Abstract

Background: TED talks are influential videos from expert speakers who share ideas worth spreading through a speaker-to-audience framework. Researchers and philosophers have been interested in untangling the art of speech since Ancient Greece.

Objective: This study identifies and quantifies the textual and descriptive factors contributing to the success of TED talks.

Data and methods: TED talks and their transcripts published between January 8th, 2006 and December 9th, 2017 were retrieved from Kaggle and the official TED Website. A talk’s success was defined by the number of views of each video. A speaker’s passion was tracked by sentiment analysis using the AFINN lexicon. Factors contributing to the success of TED talks were identified using regression, random forest, and XGboost models.

Results: A total of 1, 398 talks were included in the analyses. Adjusting for publication date, the most predictive elements of a talk’s success are temporal attributes, storytelling, speaker charisma, and speed of delivery. However, there is weak evidence for the role of starting strong, stating facts and numbers, and the speaker's passion.

Conclusions: This study demonstrates how machine learning can be leveraged to analyze the art of public speaking. Based on these results, speakers can derive data-driven recommendations to enhance their presentation’s structure, delivery, and tone to fascinate the audience.

Keywords: machine learning; sentiment analysis; quantitative text analysis; quantitative research methods; TED talks; education technology; public speaking; storytelling; random forest; XGboost; natural language processing; data analysis; data mining