“Matching Learning”: Profiling and Clustering Users on Tinder Based on Emotion and Sentiment Analysis

Abstract

Emotion and sentiment analysis tools offer the possibility of detecting emotion in several ways. In this paper, we study IBM Watson’s Natural Language Understanding (Emotion and Sentiment), and CoreNLP’s Sentiment Analysis accuracy levels against an annotated dataset, so as to observe any difference when comparing a discrete emotion classification approach (Watson’s Emotion) versus a one-dimensional approach (Watson and CoreNLP’s Sentiment). We have found that one-dimensional approaches were more accurate (85.5% for Watson, 62.7% for CoreNLP) than the discrete approach (37.6%). Being aware of those accuracy rates, and how they classified those emotions, we use those same services to analyse Tinder biographies, and observe how users present themselves, and finally applying clustering algorithms to analyse any trends between the emotion and sentiment of their biographies and other markers, such as age, gender, location, etc. After analysing the biographies, we have observed a tendency for more Neutral presentations (45–70%), followed by Positive (15–38%) presentations, though IBM Watson’s sentiment over classified joyful presentations (73%). In terms of clustering, we observe three distinct groups according to the emotional tone of their biographies, and the information provided. With this, we aim to contribute a better understanding on how two widely used emotion and sentiment analysis tools compare against each other, and how users can be classified according to the emotional/sentiment tones of their biographies.