Twitter Sentiment Analysis With Word2Vec

A Twitter sentiment analysis (TSA) project to show the effect of different training sizes on word2vec models.

This repository is based on the following two experiments:

1. Comparison between different sizes (number of tweets trained) of a self-trained word2vec model: The word2vec models are trained on a predefined number of tweets and then compared to each other.

2. Comparison between a specific word2vec model trained on the tweets themselves and a pre-trained word2vec model (word2vec-google-news-300): The domain-dependent (marked as specific in the filename) word2vec model trained on the largest number of tweets (from Experiment 1) is compared to the general (marked as nonspecific in the filename) pre-trained word2vec model.

Data

The data is taken from https://github.com/prateekjoshi565/twitter_sentiment_analysis and is originally from a competition about TSA (https://datahack.analyticsvidhya.com/contest/practice-problem-twitter-sentiment-analysis/).

However, you are not limited to using these datasets. Feel free to use your own datasets!

Getting Started

Installing

Use

pip -r requirements.txt

If this does not work, you will need the following dependencies:

• Python 3.8 (you can use Simple Python version management (pyenv) to easily manage different versions)
• Python Development Workflow for Humans (Pipenv)

and use

pyenv install 3.8.10
pyenv local 3.8.10
pipenv install

Executing Program

• python code/run.py (feel free to comment out parts before running).

---

Program structure:

The program will perform the following operations (in yellow or blue) on the data (in grey):

It will also generate visualisations to summarise the results. The generation of all models and visualisations will take about 5 min (depending on your hardware). Afterwards you will see the following additional file structure:

./
├── data
│   └── cleaned
│       ├── data.csv
│       ├── test.csv
│       ├── train.csv
│       └── validation.csv
├── models
│   └── word2vec
│       ├── 100_tweets_w2v.model
│       ├── 1000_tweets_w2v.model
│       ├── 10000_tweets_w2v.model
│       ├── 20000_tweets_w2v.model
│       └── 26971_tweets_w2v.model
└── plots
    ├── data
    │   ├── 1-class_distribution_pie_chart.pdf
    │   ├── 2-word_freq-pos_tweets.pdf
    │   ├── 3-word_freq-neg_tweets.pdf
    │   ├── 4-word_freq_with_tweets-pos_tweets.pdf
    │   ├── 5-word_freq_with_tweets-neg_tweets.pdf
    │   ├── 6-word_freq_with_words-pos_tweets.pdf
    │   └── 7-word_freq_with_words-neg_tweets.pdf
    ├── test
    │   ├── 1-class_distribution_pie_chart.pdf
    │   ├── 2-word_freq-pos_tweets.pdf
    │   ├── 3-word_freq-neg_tweets.pdf
    │   ├── 4-word_freq_with_tweets-pos_tweets.pdf
    │   ├── 5-word_freq_with_tweets-neg_tweets.pdf
    │   ├── 6-word_freq_with_words-pos_tweets.pdf
    │   ├── 7-word_freq_with_words-neg_tweets.pdf
    │   ├── 8-evaluation_f1-specific_w2v_models.pdf
    │   ├── 9-evaluation_mcc-specific_w2v_models.pdf
    │   ├── 10-evaluation_f1-unspecific_w2v_models.pdf
    │   └── 11-evaluation_mcc-unspecific_w2v_models.pdf
    ├── train
    │   ├── 1-class_distribution_pie_chart.pdf
    │   ├── 2-word_freq-pos_tweets.pdf
    │   ├── 3-word_freq-neg_tweets.pdf
    │   ├── 4-word_freq_with_tweets-pos_tweets.pdf
    │   ├── 5-word_freq_with_tweets-neg_tweets.pdf
    │   ├── 6-word_freq_with_words-pos_tweets.pdf
    │   └── 7-word_freq_with_words-neg_tweets.pdf
    └── validation
        ├── 1-class_distribution_pie_chart.pdf
        ├── 2-word_freq-pos_tweets.pdf
        ├── 3-word_freq-neg_tweets.pdf
        ├── 4-word_freq_with_tweets-pos_tweets.pdf
        ├── 5-word_freq_with_tweets-neg_tweets.pdf
        ├── 6-word_freq_with_words-pos_tweets.pdf
        ├── 7-word_freq_with_words-neg_tweets.pdf
        ├── 8-evaluation_f1-specific_w2v_models.pdf
        ├── 9-evaluation_mcc-specific_w2v_models.pdf
        ├── 10-evaluation_f1-unspecific_w2v_models.pdf
        └── 11-evaluation_mcc-unspecific_w2v_models.pdf