A scientific benchmark and comparison of the performance of sentiment analysis models in NLP on small to medium datasets
Authors : Ihab Bendidi, Yousra Bourkiche, Clément Siegrist, Kaouter Berrahal
In general, documents with similar sentiments, would be close to each other in the embeddings feature space. This can become another method to judge the performance of sentiment analysis models.
In this work, we aim to perform a benchmark of recent sentiment analysis works and models, reproduce their results, and judge their performance in comparison to baseline methods.
The following work in made on a jupyter notebook, that you can find here, or open in Colab here.
I - Processing & Exploratory Data Analysis
- Understanding the data
- Text Preprocessing
II - Sentiment classification models
- Bert Model
- LSTM recurrent model
- Baseline method : textblob
III - Document Embeddings
- Training doc2vec
- Doc2vec sentiment classifier
IV - Model performance visualisation
- Bert model
- LSTM model
- Logreg model
- Textblob
You can also find .pdfreport with code here.
This was tested on Ubuntu 20.04 with Python 3.7, but should run on any device and any python 3 version.
Before running it, make sure to install dependencies, by running in terminal :
pip install -r requirements.txt
On Google colab, you would need to import the requirements.txt file, and the tweets.csv dataset to your colab session.