Skip to content

Latest commit

 

History

History
72 lines (52 loc) · 4.4 KB

README.md

File metadata and controls

72 lines (52 loc) · 4.4 KB

Yelp Me Rate This: Using Deep Learning to Predict Yelp Review Ratings

A yelp rating predictor built using NLP models to predict the star rating of a written review.

FINAL REPORT: https://github.com/aatifjiwani/yelp-rating-predictor/blob/master/project_report.pdf

Instructions on running project/testing models

1. Clone the repository using

git clone https://github.com/aatifjiwani/yelp-rating-predictor.git

2. Download the model checkpoints from the following link:

https://drive.google.com/drive/folders/1hiW7-aJJoUsPAjYbuK8ymuz2YupcExYE?usp=sharing

Place all checkpoints within model_checkpoints/ for ease of running the scripts without FileNotFound errors.

The model_checks.zip file contains the model checkpoints from two BiLSTMs, 7 Transformers, and a RoBERTa model. If you've read the project report, the following models correspond to these checkpoints:

  • BiLSTM-Concat --> torch_bilstm_v1.pt (Requires Base Tokenizer)
  • BiLSTM-Sum --> torch_bilstm_v3_1e3.pt (Requires Base Tokenizer)
  • Transformer-256 --> torch_transformer_v1.pt (Requires Base Tokenizer)
  • Transformer-360-WCE-BPE --> torch_transformer_v4_weight.pt (Requires BPE Tokenizer)
  • Transformer-5Layer-WCE-BPE --> torch_transformer_v7_weight.pt(Requires BPE Tokenizer)
  • RoBERTa --> checkpoint_3000/(Tokenizer already included within model checkpoint) [2nd BEST MODEL]

The XLNet.zip file contains the model checkpoints for the XLNet. The zip file in its entirety is a smaller duplicate of the entire codebase but only utilize models/xlnet-checkpoint-30000/. Move this directory into the same directory as the above model checkpoints.

3. Download the Embeddings from the following link

https://drive.google.com/file/d/1FikCTiuYW_s3DOz8NpEc3pVN1n0Ocyj2/view?usp=sharing

Place embeddingsV1.txt inside embedders.

4. Install required packages from requirements.txt using

pip install -r requirements.txt

NOTE: If you're using a Mac OS/Windows, ensure you have Python 3.6 as simpletransformers requires it. If you're using Linux, ensure you have Python 3.7.

5. Testing/Running Models

IMPORTANT: All of non pre-trained Transformer models utilize YelpDataset in datasets/YelpDataset.py. For initial usage, you must pass in a tokenizer object (see datasets/vocab_gen.py).

An example of instantiating a Base Tokenizer and BPE Tokenizer is as follows:

base_tokenizer = Tokenizer("global", "datasets/vocabulary.txt")
bpe_tokenizer = ByteBPETokenizer("datasets/yelp_bpe/yelp-bpe-vocab.json", "datasets/yelp_bpe/yelp-bpe-merges.txt", max_length=250)
    
base_yelp = YelpDataset("datasets/yelp_challenge_3.jsonl", tokenizer=base_tokenizer, max_len=1000, is_from_partition=False)
bpe_yelp = YelpDataset("datasets/yelp_challenge_3.jsonl", tokenizer=bpe_tokenizer, max_len=250, is_from_partition=False)

base_loader = torch.utils.data.DataLoader(base_yelp, batch_size=32, num_workers=4, shuffle=False)    
bpe_loader = torch.utils.data.DataLoader(bpe_yelp, batch_size=32, num_workers=4, shuffle=False)

For more information on how to load the LSTMs and Transformers, refer to test_submission.py and test_torchModel.py. Please remember that the models ending in BPE require a Byte-Level BPE Tokenizer. Refer to test_torchModel.py on how to specify the tokenizer in the YelpDataset.

For information on how to load RoBERTa or XLNet, please refer to test_submission_v2.py. Loading these models are much easier than the custom PyTorch models. Also, note that it is much faster to test these pre-trained transformers when all the reviews are cleaned and in a 2D array. These models will automatically tokenize these sentences and pad if needed to the max length.

Testing datasets MUST be in a .jsonl format and obey the following convention:

{"review_id": "0", "text": "...", "stars": "1.0"}
{"review_id": "1", "text": "...", "stars": "5.0"}
...
{"review_id": "XYZ", "text": "...", "stars": "3.0"}

To test an entire dataset all at once, use

python3 test_submission<_v2>.py dataset_name.jsonl

Output predictions will be in output.jsonl. Please remember to rename before testing to another dataset otherwise predictions will be overwritten. If you want to test a different model that is not defaulted in test_submission.py, simply copy and paste the correct lines from test_torchModel.py.