The Expression of Emotion in Social Media

: Temporal Patterns and Stylometric Analysis

Abstract

Emotion significantly drives human behaviour and impacts the quality of life, studying
the interplay between public sentiments and social media is essential for understanding
and improving individual and societal well-being, and the interface between psychology
and data science is an important area to investigate. This thesis studies the expression of
emotions in social media content, with particular regard to temporal patterns in Twitter.
It also considers an issue likely to play a growing role in the future of this topic: the
difference between human-generated and machine-generated content, as we expect social
media to be populated by both types of content soon.
In the first part, we investigate periodic patterns in emotion in Twitter content, with
the aim of better understanding the underlying mechanisms of these patterns. Previous
research has shown the presence of periodic patterns in Twitter time series by analysing
specific lexicons (Linguistic Inquiry and Word Count) used for psychological measurement.
However, whether these periodic patterns are driven by internal factors (such as intrinsic
circadian rhythms) or external reasons remains unclear.
In this thesis, we compare the psychometric indicators of tweets before and during
the lockdown to further identify the causes of these temporal patterns of emotions. By
comparing tweets collected from 54 cities in the UK before and during the lockdown, we
want to discover whether the same temporal emotion patterns can still be observed after
most external factors are removed, providing more evidence to support these periodic
patterns driven by internal factors. Furthermore, our cross-countries comparison (between
the UK and Italy) further reinforced these findings, showing consistent emotional patterns
in both English and Italian tweets.
We also propose a new data analysis pipeline to detect periodic patterns in time series.
This will help us better observe the time series generated during the research process,
particularly in identifying periodic patterns within these series.
At the end, we also consider the stylometric differences (including sentiment) between
human-generated content and machine-generated content, using GPT-4.
This research contributes to a deeper understanding of diurnal patterns of psychometric
indicators in social media, providing stronger evidence and direction for future research.

Date of Award	1 Oct 2024
Original language	English
Awarding Institution	University of Bristol
Supervisor	Nello Cristianini (Supervisor) & I C G Campbell (Supervisor)