What we do

EasyEatOut allows you to quickly find restaurants that accommodate the dietary needs of all the diners in your group.

We believe that the best information you can find about a restaurant comes from other diners. This is why we power our recommendations using customer reviews of restaurants, focusing on dietary needs.

Whether you are dining with your vegan cousin, a gluten-free sister, your best friend in a vegetarian phase, or your meat-loving boss, using EasyEatOut, you’ll be able to find restaurants that leave all of them satisfied.

Product Overview

EasyEatOut allows users to choose from the 6 most common dietary requirements in the US, they can choose as many of these as they want. The dietary requirements available are:

• Vegan
• Vegetarian
• Gluten-free
• Pescatarian
• Keto
• Lactose-free

They can also select a location such as ‘Santa Barbara’ with their desired radius, and they can also input a search term if desired, such as ‘Indian food’

Customers will then be shown the restaurants that best match their input on a map and in a list format, with a score displayed for each restaurant across each dietary requirement, to help them decide where to dine.

The best way to understand our product is to watch the demo video above!

How it works

Customer reviews of restaurants power EasyEatOut. We extract the sentiment of dietary terms in customer reviews and aggregate that sentiment to get a final score for a restaurant for each dietary requirement.

Data

Our source data is publically available Yelp restaurant reviews.

No publicly available labeled data exists for the sentiment of dietary terms in sentences. We hand-labeled 1800 reviews across the 6 dietary types. We believe that if a customer mentions that a dietary requirement could be met but doesn’t give an opinion on the quality of the food, then we should still mark this as positive sentiment. To account for this, we came up with a custom sentiment classification logic outlined in the table.

Model

We used Aspect-Based Sentiment Analysis (ABSA) techniques to classify the sentiment of dietary requirements in the review text.

ABSA usually consists of three sub-tasks:

• Term extraction (e.g. ‘vegan’ or ‘plant-based’)
• Term category classification (e.g. both ‘vegan’ and ‘plant-based’ map to ‘vegan’)
• Category sentiment (positive, negative, or neutral)

Since dietary requirements don’t have many synonyms, we used a rule-based approach for term extraction and term category classification.

For the category sentiment sub-task, we used a BERT-based model. Our best model achieved an accuracy of 86.2% and a macro-F1 score of 81.1%

Code

The source code for our project can be found in our Github repo

Contributors

This project was brought to life by Ruth Ashford, Luis Bochner, Jiayi Hu, Bernardo Montufar, and Ben O’Neil