/tweet-sentiment-analysis

Sentiment Analysis of Tweets in 2016 U.S. Election Day

Primary LanguageTeX

Tweet Sentiment Analysis

This is a course project that analyzed the sentiment of tweets posted in 2016 U.S. Election Day.

We try to figure out whether using the social media can help predict the election result.

Tweet Hydrator

Due to the Twitter's ToS, the data published only contains tweet IDs, so we need to hydrator it (aka, get the full tweet information).

Install requirements:

pip install -r requirements.txt

To hydrate, first you need a CSV file with only ID in each row. Then edit the tweets_fetch.py to fill information, and run it.

Usage: tweets_fetch.py -i input_file -o output_file -p proxy_address

Options:
  -h, --help           show this help message and exit
  -p str, --proxy=str  Proxy address
  -i FILE, --in=FILE   Input CSV file
  -o FILE, --out=FILE  Output CSV file

For example, I have a CSV file called "tweet_id_1.csv" and want to get an output of "full_tweets_1.csv", then run:

python tweets_fetch.py -i tweet_id_1.csv -o full_tweets_1.csv

It also supports proxy. Use the -p option.

Sentiment Analysis

In this project, we utilized https://github.com/aalind0/NLP-Sentiment-Analysis-Twitter, which uses nltk and Sklearn to train and provides the best optimized sentiment analysis. To run the analysis, you need to do the following...

  1. Install required packages and data

    1. Install sklearn with pip install scikit-learn
    2. Install nltk with pip install nltk
    3. Open a fresh python interpreter, run 
      > import nltk
> nltk.download('stopwords')
> nltk.download('movie_reviews')
> nltk.download('averaged_perceptron_tagger')
> nltk.download('punkt')

  2. Run the train_classifiers.py file to train models. Or you may use the pretrained models in this repo.

  3. Run sentiment_calculation_multithread.py (it will use 1/4 of all your CPU cores to calculate) or sentiment_calculation.py (it will only utilize one core using one thread) to calculate the sentiment. You need to use this syntax: python xxx.py <index> and replace the <index> with the number of csv file. The filename is hardcoded so you may change it yourself.

Resulting Accuracy

The accuracy varies because we randomly our training sets. But it should be stable at around $[65, 75]$. This is a demo run:

  • Original Naive Bayes: 72.9607250755287
  • Sklearn Multinomial Naive Bayes: 70.2416918429003
  • Sklearn Bernoulli Naive Bayes: 72.35649546827794
  • Sklearn Logistic Regression: 70.69486404833837
  • Sklearn Linear SVC: 67.97583081570997
  • Sklearn SGD classifier: 67.06948640483384

Voted Classifier: 71.75226586102718

Reference