yuzhoumao / autocomplete

neural network (vanilla, RNN, LSTM) based auto sentence completion (suggestion) system on yelp user reviews

Review Auto Completion

The complete documentation of our code is at the wiki page: https://github.com/liuxinglian/autocomplete/wiki.

Introduction

Inspired by Smart Compose, a new feature recently launched by Google for its email service that provides automatic sentence completion using machine learning and artificial intelligence techniques, we aim to build a system that could do auto sentence completion on yelp user reviews.

Dataset

• The dataset we use is the Yelp Open Dataset. The Yelp Open Dataset consists of a few JSON data files (business.json, review.json, user.json, etc). We only use one of the JSON data files in the dataset, namely review.json.

• In review.json, there is one JSON object per line. One sample json object in review.json is provided as follows.

{
    // string, 22 character unique review id
    "review_id": "zdSx_SD6obEhz9VrW9uAWA",

    // string, 22 character unique user id, maps to the user in user.json
    "user_id": "Ha3iJu77CxlrFm-vQRs_8g",

    // string, 22 character business id, maps to business in business.json
    "business_id": "tnhfDv5Il8EaGSXZGiuQGg",

    // integer, star rating
    "stars": 4,

    // string, date formatted YYYY-MM-DD
    "date": "2016-03-09",

    // string, the review itself
    "text": "Great place to hang out after work: the prices are decent, and the ambience is fun. It's a bit loud, but very lively. The staff is friendly, and the food is good. They have a good selection of drinks.",

    // integer, number of useful votes received
    "useful": 0,

    // integer, number of funny votes received
    "funny": 0,

    // integer, number of cool votes received
    "cool": 0
}

• The small_dataset_1200.json consists of randomly selected 1200 reviews from Yelp dataset, mainly used for hyperparameter tuning. The large_dataset_12000.json consists of randomly selected 12000 reviews from Yelp dataset, which is used for training our model with 10000 reviews and test on 2000 reviews to get the performance of our models.

Requirements

• python (3.5.x, 3.6.x)

  • tensorflow==1.12.0

  • gensim==3.6.0

  • nltk==3.4

  • pygtrie==2.3

Usage

• Usage of the models that we built is as follows:

  • Model 1

    python model1/model1.py

    To change the hyperparameter of model 1, please change the variables in model1/model1_config.py

  • Model 2

    python model2/model2.py

    To change the hyperparameter of model 2, please change the variables in model2/model2_config.py

  • Model 3

    python model3/model3.py

    To change the hyperparameter of model 3, please see the documentation at https://github.com/liuxinglian/autocomplete/wiki/model3.py