Yelp Dish Ranker

Ranks dishes for Yelp restaurants from reviews using NLP tools (spaCy, TextBlob, Jellyfish)

Built upon data from the Yelp Dataset Challenge.

This repo contains my contributions to my final group project for CS 378, Practical Applications of Natural Language Processing (about the project).

The rank_dishes method...

1. Recognizes dishes from Yelp text reviews using a custom spaCy NER model trained on collected Amazon Turk data
2. Finds the sentiment of sentences containing dishes using TextBlob to collect ratings for dishes
3. Combines similar dishes, e.g. groups of dishes that have a high enough Jaro-Winkler distance
4. Aggregates dish ratings into a single score using Bayesian Average Ratings
5. Returns dishes in an ordered list from best (highest score) to worst (lowest score)

Run

To test the ranking algorithm, simply install the libraries found in requirements.txt and run rank_dishes.py. It will read in and execute on the sample reviews in emerald_reviews.json, which are real reviews for the Emerald Chinese Restaurant in Mississauga, Ontario.

Running rank_dishes.py as is takes <10 seconds and prints the following output:

Emerald Chinese Restaurant - 20 best dishes:
1 - dim sum 0.598
2 - braised beef 0.292
3 - sticky rice 0.286
4 - spring rolls 0.208
5 - lobster 0.208
6 - wonton 0.208
7 - cookies 0.174
8 - bunch our staples 0.167
9 - preserved egg congee 0.167
10 - egg yolk bun 0.167
11 - leaf wrapped sausage 0.167
12 - octopus 0.167
13 - sautéed jumbo shrimp 0.13
14 - grilled vegetables 0.13
15 - fish fried rice 0.13
16 - beef sauté 0.13
17 - pork 0.125
18 - steamed fish 0.12
19 - fried rice 0.111
20 - bubble tea 0.087

Assuming you have the Yelp Dataset downloaded, main.py will run the ranking algorithm for all restaurants in the dataset and save the results to a json file. Note: This takes a really long time as there are over 15000 restaurants in the data.