Source-linked AI summary
Modeling public mood and emotion: Twitter sentiment and socio-economic phenomena
Johan Bollen, Alberto Pepe, Huina Mao
TL;DR
The paper examines whether public mood expressed in tweets can be quantitatively related to major social, political, cultural, and economic events. It analyzes public tweets from August 1 to December 20, 2008 using an extended POMS-based sentiment method, finding immediate and specific mood effects while noting time-series comparability constraints from changing tweet volumes.
Problem
The study asks whether aggregate public mood expressed in microblogs has a quantifiable relationship with major social, economic, and cultural events.
Method
The authors apply a syntactic, term-based sentiment analysis using extended POMS, extract six mood dimensions per tweet, aggregate them daily, and compare them with event and economic timelines.
Results
Events across social, political, cultural, and economic spheres have significant, immediate, and highly specific effects on public mood dimensions.
Takeaways & Limitations
Large-scale analyses of user-generated content may support modeling and prediction of collective emotive trends in relation to social and economic indicators.
Takeaways & Limitations
Changing tweet volumes alter the variance of mood estimates over time, making direct comparisons of mood-vector changes problematic without normalization.
Abstract
from arXiv · showhide
Microblogging is a form of online communication by which users broadcast brief text updates, also known as tweets, to the public or a selected circle of contacts. A variegated mosaic of microblogging uses has emerged since the launch of Twitter in 2006: daily chatter, conversation, information sharing, and news commentary, among others. Regardless of their content and intended use, tweets often convey pertinent information about their author's mood status. As such, tweets can be regarded as temporally-authentic microscopic instantiations of public mood state. In this article, we perform a sentiment analysis of all public tweets broadcasted by Twitter users between August 1 and December 20, 2008. For every day in the timeline, we extract six dimensions of mood (tension, depression, anger, vigor, fatigue, confusion) using an extended version of the Profile of Mood States (POMS), a well-established psychometric instrument. We compare our results to fluctuations recorded by stock market and crude oil price indices and major events in media and popular culture, such as the U.S. Presidential Election of November 4, 2008 and Thanksgiving Day. We find that events in the social, political, cultural and economic sphere do have a significant, immediate and highly specific effect on the various dimensions of public mood. We speculate that large scale analyses of mood can provide a solid platform to model collective emotive trends in terms of their predictive value with regards to existing social as well as economic indicators.
1. INTRODUCTION
Microblogging supports varied public communication, and tweets can encode authors’ moods even when mood is not their explicit subject. The study asks whether aggregate public mood relates quantitatively to major social, economic, and cultural events.
- Microblogging lets users broadcast brief updates publicly or to selected contacts, supporting daily chatter, conversation, information sharing, and news reporting.
- Tweets can convey authors’ moods through explicit self-expression or affective reactions embedded in information-sharing messages.
- The study examines whether public mood patterns from Twitter relate to fluctuations in macroscopic social and economic indicators.
- The analysis focuses on relationships between public mood and events including stock-market drops, oil-price rises, and a political-election outcome.
2. RELATED LITERATURE
Prior sentiment research spans natural-language, machine-learning, and term-based approaches applied to reviews, blogs, and other online content. The paper builds on this literature by applying an extended POMS system to microblogging, where scale and short texts pose methodological challenges.
- Sentiment analysis has been applied to public user-generated text, including reviews, blogs, online journals, and social content.
- Natural-language approaches use word constructs, while machine-learning methods such as SVMs classify sentiment from text.
- Online mood analyses have estimated happiness, seasonality, mood trends, sales, and public mood indicators from blogs and other web materials.
- Microblogging remained less studied than blogs, while very short texts and limited training data can challenge machine-learning sentiment classification.
- The study uses an extended, term-based Profile of Mood States system previously applied and validated on online textual corpora.
3. METHODS
The study builds daily public-mood time series from 9,664,952 tweets using an extended POMS instrument, then compares them with socio-economic events and indicators. It applies normalization to address changing tweet volumes and support comparisons across time.
- Data and instruments: The analysis uses POMS-ex, an extended POMS instrument that maps tweets onto Tension, Depression, Anger, Vigour, Fatigue, and Confusion.POMS-ex expands the original 65 mood adjectives to 793 terms using synonyms and related word constructs from WordNet and Roget’s Thesaurus.
- Data preparation: Each tweet is normalized through token separation, punctuation removal, lowercasing, stop-word removal, and Porter stemming before mood scoring.The procedure retains tweets representing explicit sentiment or an individual’s present status, producing a subset of 1.1M normalized tweets for scoring.
- Mood scoring and aggregation: POMS scoring counts matches between tweet terms and the adjective sets for six mood dimensions, producing a six-dimensional mood vector that is then unit-normalized.Daily aggregate mood vectors are obtained by averaging the normalized tweet mood vectors for tweets submitted on each date.
- Time-series normalization: Because daily tweet counts vary, early time-series values have larger variance; the study therefore uses local z-score normalization and variance normalization for temporal comparison.Z-scores emphasize local short-term deviations, while variance normalization uses a variable mean to compare general mood levels across periods while maintaining normalized variance.
4. RESULTS
The study examines public mood around major events and across economic-indicator periods using six daily mood dimensions extracted from Twitter. Election and Thanksgiving produced sharp, specific changes, while economic-indicator relationships were delayed, dimension-dependent, and inconclusive for long-term trends.
- 4.1 Case studies: The analysis first tests mood-series validity against the 2008 U.S. presidential election and Thanksgiving, then examines longer-term effects of DJIA and WTI changes.The case studies use short event windows, while economic comparisons use four periods of indicator change.
- 4.1 Case studies: Around the election, Depression and Confusion spiked before November 4, while Fatigue fell and Tension peaked on election day.The broader series shows nearly 4-standard-deviation movements: Vigour from -1 to +3 and Tension from -2 to +2.
- 4.1 Case studies: Thanksgiving left nearly all mood dimensions near baseline, but Vigour rose sharply and Fatigue dipped, marking the period's largest positive Vigour spike from 0 to +4 standard deviations.The pattern is consistent with a happy, energetic holiday.
- 4.2 General correlation drivers versus public mood trends: The six mood dimensions were not statistically significantly correlated across all 141 days, indicating differentiated emotional responses during the period's turmoil.The study reports no statistically significant pairwise correlations across the full period.
- 4.2 General correlation drivers versus public mood trends: Across DJIA periods, Depression increased from a median of -0.983 standard deviations in DJIA-I to 4.532 in DJIA-IV, while Anger rose from -0.198 to 2.043.Both changes were reported with p < 0.0001; Fatigue declined continuously across all DJIA periods.
- 4.2 General correlation drivers versus public mood trends: DJIA-IV had significantly higher Tension, Depression, and Anger than earlier periods, even after the index had stabilized at a relatively low plateau.Negative sentiment increased after the DJIA's fall, while Depression rose throughout the three-month period.
- 4.2 General correlation drivers versus public mood trends: WTI-III to WTI-IV aligned better with mood-curve discontinuities than the corresponding DJIA boundary and showed more significant differences for Vigour and Fatigue.The authors suggest WTI changes may correspond more closely in time to mood changes but affect different dimensions than DJIA changes.
- 4.2 General correlation drivers versus public mood trends: Economic effects appeared delayed and cumulative, whereas short-term mood variability was predominantly associated with news, elections, holidays, and other short-lived events.The long-term relationship between economic indicators and mood remained inconclusive.
5. CONCLUSION
The study uses Twitter’s brief, time-stamped messages to investigate public mood and emotive trends through daily sentiment analysis. It argues that a syntactic, term-based approach and established psychometric measures can support large-scale modeling of collective mood in social, economic, and cultural context.
- Tweets from August 1 to December 20, 2008 were analyzed as temporally specific signals of public mood.The study treats each tweet as a microscopic, temporally authentic sentiment observation and aggregates mood components daily.
- The analysis extracted six mood dimensions from each tweet and compared daily trends with cultural, social, economic, and political events.Comparisons included stock-market and crude-oil indices, the U.S. Presidential Election, and Thanksgiving Day.
- Social, political, cultural, and economic events were correlated with significant, sometimes delayed fluctuations in public mood levels.The conclusion preserves an association rather than claiming that these events causally produced the fluctuations.
- The paper argues that sentiment analysis of very short text corpora can be performed efficiently with a syntactic, term-based approach requiring no training or machine learning.It presents this as a methodological contribution for analyzing minute text corpora such as tweets.
- The authors emphasize established psychometric instruments and speculate that large-scale user-generated content could help model and predict collective emotive trends.They caution that such results should be interpreted within the social, economic, and cultural spheres where users are embedded.