Train a Sentiment Classifier on Amazon Reviews

The goal of this project is to train several sentiment classifiers on a subset of the Amazon Reviews dataset in order to predict if a given text has a positive or negative sentiment: the best model on efficiency and efficacy is then tested on a corpus made of several labelled and unlabelled tweets.

In order to correctly execute this project, the following steps should be done:

1. Download Amazon Reviews dataset

For downloading the different datasets it is sufficient to execute script.sh as follows:

./script.sh dataset_links.txt  

If the script doesn't work you have to change read, write and execute permissions as follows:

chmod 700 script.sh

A huge thanks should be made to Jianmo Ni and his colleagues for making the dataset available (citation on the references section) on this website.

2. Install required libraries

In order to install the required libraries for running this project the following instruction should be executed:

pip install -r requirements.txt

Windows users may have trouble in installing the pystemer library. In order to do proceed with the installation it is required to download the Build Tools for Visual Studio 2019: please select the tools related to C++ and proceed with the download (it should be around 4-5 GB).

3. Execute notebooks

The project consists of different notebooks:

1. 00_Initial_Data_Cleaning.ipynb : in this notebook normalization techniques such as lowercasing, emoticon and slang replacements are performed
2. 01_Preprocessing_Representation_Classification.ipynb: in this notebook preprocessing techniques such as tokenization, spell checking and stemming are performed. The preprocessed dataset is then used for computing the TF-IDF Matrix and training several classification algorithms
3. 02_Download_tweets.ipynb: in this notebook tweets related to CES2020 and replies to a given Samsung tweet are downloaded for later evaluation
4. 03_Tweets_Sentiment_Analysis.ipynb: the best predictive model built in the notebook 01_Preprocessing_Representation_Classification.ipynb is used for performing sentiment analysis on downloaded tweets

Each notebook, apart from 01_Preprocessing_Representation_Classification.ipynb can be executed independently thanks to:

• intermediate datasets in the "dataset/" folder
• vectorizer and predictive model in the "joblib_data/" folder. The reason why 01_Preprocessing_Representation_Classification.ipynb can't be executed independently is that the starting dataset, called preprocessed_dataset.csv, is larger than GH maximum size, even in a compressed format.

4. Have fun !

References

Justifying recommendations using distantly-labeled reviews and fined-grained aspects

Jianmo Ni, Jiacheng Li, Julian McAuley

Empirical Methods in Natural Language Processing (EMNLP), 2019

pdf