Tyler Atkinson

Using machine learning and natural-language-processing to determine whether someone is happy or not, based off of their tweets.

www.twittersentiment.info

1. Motivation for project
2. Collecting and Cleaning Data
3. Modeling
4. Why is sentiment important?
5. Tech Stack

What is sentiment analysis?
Sentiment analysis identifies if a piece of text is positive, neutral, or negative depending on the type of wording they use, any emoji's, etc. An example of positive and negative tweets would look like:

Positive Sentiment	Negative Sentiment

I collected my data using Twitter's streaming API. Over the course of about a week, I collected a total of 800k tweets that were sent out from the United States. After receiving a tweet from the API, I parsed it to pull out only the data I wanted: the text, time, location, etc. and saved it to a Mongo database.

To get my data ready for any exploratory analysis or modeling, I used regex to:

• Remove links
• Remove user '@' handles
• Remove hashtags

Sentiment analysis typically takes one of two forms: polarity-based, where pieces of texts are classified as either positive or negative, or valence-based, where the intensity of the sentiment is taken into account. The method I used to classify tweets as happy, neutral or sad uses both approaches.

To get a sentiment score, I used a package built into the python package nltk called Vader which stands for Valence Aware Dictionary and sEntiment Reasoner. It is a lexicon and rule-based sentiment analyzer that performs exceptionally well with sentiments expressed in social media. Social media is hard for natural language processing because people rarely use proper sentence structure and tend to use a lot of slang, and this is where Vader outperforms other tools. Using the examples from above, we can get a sentiment score of the tweet. After cleaning the text, we get a score composite score of .99 and -.98 respectively.

'Happy birthday sweet pea! 💛💛 I hope you have an amazing 18th birthday!
you’re such a kind soul and I hope you’re doing great (: and thank you for
listening to me when I needed advice or just needed someone to tell everything
was going to be okay, I love you lots 💛'
>>> {'compound': 0.9884, 'neg': 0.0, 'neu': 0.499, 'pos': 0.501}

'Ugh.  Worst decision by the Seahawks since Percy Harvin over  No, worst since
Jim Mora. Hell, maybe even the worst since The Boz. UGH!'
>>> {'compound': -0.9793, 'neg': 0.602, 'neu': 0.398, 'pos': 0.0}

After labeling my tweets on a scale from -1 to 1, I decided to create a range for positive, neutral, and negative tweets and chose the following scale

I then created combinations of the following models paired with both CountVectorizer and TfidfVectorizer: Naive Bayes, LinearSVC, and Random Forest Classifier. Ultimately, the LinearSVC along with CountVectorizer performed the best.

CountVectorizer works by turning a collection of text documents (in my case tweets) into a sparse matrix of token counts without any stop words (the, is, a, so, very, etc.).
For example, the tweet 'Happy birthday sweet pea! 💛💛 I hope you have an amazing 18th birthday!' would be turned into

This step also prepares my data for modeling, since you cannot enter raw text into LinearSVC. LinearSVC works by plotting your data, and it tries to find a hyperplane that separates classes in this feature space. For my application, these classes consisted of tweets that were positive, neutral, or negative. An optimal separating hyperplane separates classes and maximizes the margin between the classes. An example of this might look like this

You can adjust these margins depending on your use case, a hard (smaller) margin will be more accurate, and a soft (bigger) margin will be better at generalization. For my project, I did not have to adjust this parameter.

To answer this question, I broke my analysis into two parts:

• Sentiment of topics based on hashtags or keywords
• Sentiment and location

To show how one could use keywords and sentiment to analyze a topic, I decided to capture tweets during Super Bowl LII. I created two groups from these tweets: one group contained tweets about the Patriots, the other contained tweets about the Eagles. Below is a graph of the average sentiment of the two groups as the game progressed.

Once you have a happy or sad label for tweets, you can filter through them by specific keywords or hashtags and look at what "happy" and "not happy" people are saying. For example, if we wanted to look at what people were saying about the college football championship game on January 8th, we could filter by the keywords "football, Alabama, Georgia" and generate the following word clouds.

(hint: the top one is positive)

There are some impressive statistics about these tweets as well. In the above example, there were:

• 374 positive tweets with an average sentiment of .54
• 161 negative tweets with an average sentiment of -.49

To see if there was a relationship between someone’s location and the happiness of their tweet, I graphed the average sentiment of every county I had in my dataset. I did, however, find that there were counties with only one tweet, which obviously does not accurately represent that county.

https://public.tableau.com/shared/XJ5FFF4RR?:display_count=yes - Click to see an interactive version

For a better location example, I filtered tweets containing words and hashtags that related to the two NFL teams playing in Super Bowl LII. I created two groups:

• Group describing the Patriots (‘pats, brady, patriots’, etc.)
• Group describing the Eagles (‘eagles’,‘Philadelphia’,‘nick foles’, etc.)
  Not surprisingly, most people from Massachusetts were tweeting about the Patriots and same for Pennsylvania and the Eagles.

Hutto, C.J. & Gilbert, E.E. (2014). VADER: A Parsimonious Rule-based Model for Sentiment Analysis of Social Media Text. Eighth International Conference on Weblogs and Social Media (ICWSM-14). Ann Arbor, MI, June 2014.