Source-linked AI summary
Tracking Sentiment in Mail: How Genders Differ on Emotional Axes
Saif M. Mohammad, Tony, Yang
TL;DR
The paper addresses how emotions in large amounts of personal email and other mail can be quantified and tracked. It combines a crowdsourced emotion lexicon with sentiment analysis and visualizations, finding gender differences in workplace email and demonstrating comparative and tracking applications.
Problem
The paper asks how sentiment analysis and visualizations can quantify and track emotions in the large amounts of text contained in email and other mail.
Method
The authors create a crowdsourced word–emotion lexicon and use it with emotion analysis and visualizations to study mail corpora and workplace email.
Results
Women use and receive relatively more joy and sadness words, whereas men use and receive relatively more trust and fear words in workplace email.
Takeaways & Limitations
Emotion lexicons and visualizations can support comparisons across mail types and help people track emotion words in sent and received email.
Abstract
from arXiv · showhide
With the widespread use of email, we now have access to unprecedented amounts of text that we ourselves have written. In this paper, we show how sentiment analysis can be used in tandem with effective visualizations to quantify and track emotions in many types of mail. We create a large word--emotion association lexicon by crowdsourcing, and use it to compare emotions in love letters, hate mail, and suicide notes. We show that there are marked differences across genders in how they use emotion words in work-place email. For example, women use many words from the joy--sadness axis, whereas men prefer terms from the fear--trust axis. Finally, we show visualizations that can help people track emotions in their emails.
1 Introduction
The paper applies sentiment analysis and visualizations to track emotions in letters and emails, introducing a crowdsourced lexicon and several comparative and practical applications.
- Email provides unprecedented quantities of self-written and received text for tracking emotional states.
- The paper describes applications including suicide-note risk analysis, gendered communication analysis, emotion tracking over time, affect-based search, and emotion-focused writing assistance.
- The authors create a large word–emotion association lexicon through crowdsourcing with quality-control measures.
- They compare emotion words in love letters, hate mail, and suicide notes, and use these corpora to evaluate the lexicon and visualizations.
- The emotion analyzer identifies positive, negative, and neutral polarity and associates words with eight basic emotions.
2 Related work
Prior sentiment-analysis work covers polarity and selected emotions, while email affect analysis has often relied on surface features or small lexicons.
- Sentiment-analysis research has focused on identifying positive or negative polarity and, increasingly, emotions such as anger, joy, sadness, fear, surprise, and disgust.
- Prior email-affect analysis primarily addressed automatic gender identification using surface features and very small emotion lexicons.
3 Emotion Analysis
The authors build a crowdsourced, word-sense-aware emotion lexicon with quality controls, then use lexicon matches to estimate emotion distributions in larger texts.
- The lexicon uses Roget’s Thesaurus targets and annotates words occurring more than 120,000 times in the Google n-gram corpus.
- Word-sense annotations begin with a word-choice question that guides annotators toward the target word’s intended sense.
- Annotators answer ten association questions covering polarity and eight emotions after the sense-selection question.
- About 10% of annotations are discarded after incorrect sense-selection responses, filtering likely unreliable emotion annotations.
- The resulting lexicon contains about 24,200 word–sense pairs and about 14,200 word types, combining five annotators by majority vote.
- For a target text, the system counts lexicon matches and calculates emotion-word ratios, which are more reliable for comparing large texts than individual sentences.
4 Love letters, hate mail, and suicide notes
The paper compares emotion-word patterns across love letters, hate mail, and suicide notes using corpus-level percentages, difference bar graphs, and relative-salience word clouds. Love letters contain more joy and trust words, whereas hate mail contains more fear, sadness, disgust, and anger; additional visualizations identify salient words in cross-corpus comparisons.
- Corpora: 348 love-letter postings, 279 hate-mail pieces, and 21 suicide notes form the three comparison corpora.The corpora were compiled from publicly available sources and were intended for continued expansion.
- Emotion comparisons: Love letters have many more joy and trust words, while hate mail has many more fear, sadness, disgust, and anger words.Figures 1–7 show polarity and emotion-word percentages for the three corpora; Figure 4 compares love letters with hate mail.
- Visualizations: Relative-salience word clouds supplement emotion-difference bars by showing which words contribute to an emotion’s prominence across two corpora.The method normalizes word frequencies by total token counts in the compared texts.
- Visualizations: For joy, love letters are more associated than hate mail with words such as loving, baby, beautiful, feeling, and smile.The authors describe this word cloud as a sanity check of the manually created emotion lexicon.
- Cross-corpus comparisons: Suicide notes show salient fear words including hell, kill, broke, worship, sorrow, afraid, loneliness, endless, shaking, and devil relative to love letters.These words are listed in decreasing order of salience.
- Cross-corpus comparisons: A disgust-word cloud contrasts hate mail with suicide notes, featuring words such as ignorant, quack, fraudulent, illegal, lying, and damage.Cancer and disease are prominent because the hate-mail source denigrates alternative-treatment websites and receives angry messages from some patients and physicians.
5 Emotions in email: men vs. women
The paper examines emotion-word differences in workplace email by gender and correspondence direction. Women communicate relatively more on the joy–sadness axis, whereas men prefer the trust–fear axis, with anticipation especially elevated in cross-gender communication.
- Data and scope: The analysis uses the Enron email corpus to study gender differences in workplace emotion-word use, while acknowledging that it does not prove the underlying propositions.The corpus contains more than 200,000 emails from 150 senior managers, with personal communication alongside official business.
- Sender gender: Men used many more trust words than women in emails they sent.Figure 12 compares emotion-word percentages in emails sent by men and women, and Figure 13 visualizes the trust-word difference.
- Recipient gender: Men received more trust words, whereas women received more joy words.This comparison concerns emails sent to women versus emails sent to men.
- Correspondence direction: Men used many more anticipation words when writing to women than when writing to other men.The same comparison also reports a marked difference in trust words.
- Discussion: Women use more joy–sadness words, men prefer trust–fear words, and anticipation is markedly higher in cross-gender communication than same-sex communication.The discussion synthesizes comparisons across sender gender, recipient gender, and correspondence direction.
6 Tracking Sentiment in Personal Email
The paper presents individual-oriented visualizations for tracking emotions in personal email. A Gmail-integrated application lets users inspect sentiment and emotion-word differences across correspondents and individual messages.
- Application: A Google Apps API application integrates with Gmail to help individuals track emotions toward people they correspond with.The visualizations are designed for users who generally have access only to emails they send and receive.
- Recipient-level views: For John Arnold, the visualizations show positive and negative-word percentages in emails sent to colleagues and differences for selected recipients.Selecting a bar reveals the difference between emails sent to a particular person and all emails sent out.
- Message-level views: The application also shows positive and negative-word percentages for each email John sent to Andy Zipper.This provides message-level tracking in addition to aggregate correspondent-level views.
- Privacy and future analysis: The proposed large-scale personal-email analysis would share emotion-word frequencies, gender, and age without sharing text, names, or email IDs.The future application was described as protecting user privacy by transmitting aggregate metadata rather than message content.
7 Conclusions
The paper develops a crowdsourced emotion lexicon and applies it with visualizations to analyze and track emotions across mail. Its analyses reveal gendered patterns in workplace email and support tools for monitoring sent and received messages.
- The authors created a large word–emotion association lexicon by crowdsourcing and used it to analyze and track emotion-word distributions in mail.
- Women use and receive relatively more joy and sadness words in workplace email, whereas men use and receive relatively more trust and fear words.
- Cross-gender workplace communication contains a markedly higher percentage of anticipation words than same-sex communication.
- Comparisons of love letters, hate mail, and suicide notes provide corpora for examining emotion-word distributions.
- Word clouds, visualizations, and a Gmail application help people interpret and track emotion words in sent and received emails.