2. Fmylife.com
• A website devoted to sharing unfortunate
stories about one’s day
• The format is always the same: “Today, ...
FML.”
• Votes: “I agree, your life sucks” or “You
totally deserved it”
3. Examples
• “Today, I found out that I am being sued for
losing a set of wedding photos that I took. I
lost them by being mugged on the way
home after the shoot and £10,000 worth of
equipment was stolen from me. FML”
• “Today, my boyfriend broke up with me
over Facebook. He was sitting right next to
me. FML”
4. Examples
• “Today, I was absently chewing a torn nail off
my finger. Not thinking, I spit the nail out... and
watched it land in my boss's coffee mug. FML”
• “Today, I tried to trick people into thinking that
I was in a relationship by changing my
relationship status on facebook to "in a
relationship". Only one person commented on
it. They said "HAHAHAHA yeah right!" FML”
5. Sentiment Analysis
• Also “opinion mining”
• Detect the attitude of the author
automatically
• Ex: Positive or negative movie review
• Often done by examining keywords and
adjectives, good indicators of emotion
• Highly relevant for business
6. Dataset
• Mined Fmylife.com for 600 pages of stories
• Tag by votes of the users
• 7816 stories total:
• 71.91% were sympathetic
• 22.69% were “you deserved it”
• 5.40% were neutral
7. Methods
• Used Python’s Natural Language Toolkit
(NLTK) to do analysis
• Throw out stopwords
• NaïveBayes using different featuresets:
• lengths of words, stories
• most frequent words (unigrams)
• Only improves performance 1-2%
8. Why FML is hard
• High amounts of pragmatics and inferred
knowledge
• Adjectives don’t suffice, and emotion/
sentiment of author is the same across all
stories
• Overall goal is to infer an emotional
reaction from the reader, not from the
author
9. Possible Next Steps
• Bigrams, POS tagging, sentence parsing, etc.
• Examining comments on the stories
• Other ML models
• ???