Team 01 : Aspect and Opinion Extraction for Amazon Reviews

DESCRIPTION

This is a compilation of all the things required to run this project file. It is arranged in a step-wise manner, where we cover setting up the environment first, explain the different code files and finally how to run the visualisation.

Github link - https://github.com/ishikaarora4/Aspect-Sentiment-Analysis-on-Amazon-Reviews Blog link - http://bit.ly/nlpamazon

The folder structure of the package is mentioned below - /data - Contains all the required data files /src - Contains all the code files requirements.txt - File to setup the python environment for running the model

INSTALLATION

Setting up the environment

  • Setup a base python environment using Miniconda (https://conda.io/miniconda.html)
  • Create a conda environment using the requirements.txt file so that all dependancies are installed conda create --name <env> --file requirements.txt
  • Activate the environment just created using source activate <env_name>
  • Run this to download the language corpus we used - python -m spacy download en_core_web_lg
  • Run this to download the NLTK polarity model inside a python shell - nltk.download('vader_lexicon')

Note : The models take upto 7-8 hours to run for 1M set of reviews. We ran the whole process on an AWS EC2 instance. Moreover, roughly the loading of NLP language corpus itself needs around 2GB RAM, so use higher configurations.

Code Files

/src/dataprep/clean_data.py - Contains function to clean the review text /src/dataprep/aspect_json_encoding.py - An ad-hoc file used to solve an encoding issue when writing the file in json format

/src/models/aspect_extraction.py - Contains all the required functions to fetch the data from S3 buckets and give aspect-aspect modifier in a json format. /src/models/mapper.py - Aggregates all the aspect-aspect modifiers according to product_id which will be passed on to the clustering script

/src/models/aspect_clustering.py - Contains all required functions to run clustering

code/src/UI_stuff - All code related to UI

EXECUTION

Run Model

We are providing with two options to run -

  1. On one whole file (amazon_reviews_us_Electronics_v1) which has around 3M+ reviews
  2. One toy file we have created as a subset of the above mentioned file

Add argument as 1, if you want to run on the toy file.

Note : Run all commands except UI from the code folder

  • python src/models/run_extraction.py <arg = 1 or 0>
  • python src/models/aspect_clustering.py

About the files that are generated -

  • All the files generated in data/interim folder are needed for functioning of the model.
  • Files in data/processed -
    • model_results_encoding.json - the final file that is used to populate the db and then UI used it
    • reviews_aspect_mapping.json - the output file from first step (aspect extraction) used by clustering

Access UI

  1. Navigate to the folder code/src/UI_stuff
  2. Run the command "python app.py"
  3. To avoid getting a cross-origin resource sharing (CORS) error, you will have to install an extension for firefox to allow CORS. https://addons.mozilla.org/en-US/firefox/addon/access-control-allow-origin/
  4. Open "http://127.0.0.1:5000/" in full-sized Firefox window
  5. Search for Amazon product IDs (sample list below) a. It takes ~15-20 seconds to query the database / product review page to load

Sample list of product IDS: B0009YJXMI B000A3WS16 B00006JQ5N B0007PGADE B000068O48 B00000J1QR B00030CHQ2