Skip to content

CodeBERTScore: an automatic metric for code generation, based on BERTScore

License

Notifications You must be signed in to change notification settings

neulab/code-bert-score

Repository files navigation

CodeBERTScore

This is the official implementation of the paper:

Shuyan Zhou, Uri Alon, Sumit Agarwal, Graham Neubig, CodeBERTScore: Evaluating Code Generation with Pretrained Models of Code

CodeBERTScore is an Automatic Evaluation Metric for Code, based on BERTScore. This repository is based on the code of BERTScore, and we are grateful to the authors for releasing their code.

April 2023 - CodeBERTScore is now available on pypi, which means that you can simply pip install code-bert-score!


Example:

Figure (a) shows a reference code snippet in Java. Figures (b) and (c) show two generated predictions. Among these two candidates and given the reference, BLEU prefers (scores higher) the code in (b), which is not functionally equivalent to the reference, while CodeBERTScore prefers the code in (c), which is functionaly equivalent to the reference.

How does it work?

As BERTScore, CodeBERTScore leverages the pre-trained contextual embeddings from a model such as CodeBERT and matches words in candidate and reference sentences by cosine similarity. Differently from BERTScore, CodeBERTScore also encodes natural language input or other context along with the generated code, but does not use that context to compute cosine similarities.

This example shows how CodeBERTScore can compute the similarity between the Python expressions x ** 0.5 and math.sqrt(x), which are functionally equivalent, even though they have very few overlapping tokens.

Usage

import code_bert_score
pred_results = code_bert_score.score(cands=predictions, refs=refs, lang='python')

Where pred_results is a 4-tuple of (precision, recall, F1, F3), where each is a 1-D tensor of scores for each prediction-reference pair. F3 is similar to the well-known F1 score, that considers recall 3 times as important as precision. See the definition on Wikipedia.

See our example.py script. Additional details are shown in the original BERTScore demo notebook.

Huggingface 🤗 Models

We fine-tuned the microsoft/codebert-base-mlm model for 1,000,000 steps (with batch_size=32) on several languages separately.

We released the following models to the Huggingface hub:

  • neulab/codebert-python (the default model for lang='python')
  • neulab/codebert-javascript (the default model for lang='javascript' or 'js')
  • neulab/codebert-c (the default model for lang='c')
  • neulab/codebert-cpp (the default model for lang='cpp' or 'c++')
  • neulab/codebert-java (the default model for lang='java')

The appropriate model will be loaded automatically when passing the lang argument to the score(..) function, for example: lang='python'. For other uses, these models can be loaded using (for example):

from transformers import AutoTokenizer, AutoModelForMaskedLM

tokenizer = AutoTokenizer.from_pretrained("neulab/codebert-python")
model = AutoModelForMaskedLM.from_pretrained("neulab/codebert-python")

Additional Features

  • We found that in NL->Code problems, more accurate results are achieved by encoding the NL sources with the code prediction, but then measuring similarity only for the encoded code:
pred_results = code_bert_score.score(cands=predictions, refs=refs, lang='python', sources=sources)
  • We also found that using Inverse Document Frequencies improve the results, similarly to the original BERTScore. We included an example script that shows how to precompute them here compute_idf.py. Then, the resulting dictionary can be used with the argument idf=idf_dict. Our IDF dicts can be found in ./idf_dicts/.

  • Tuning the layer that the similarity is computed from is also helpful, using num_layers=N where N is between 5-10:

  • We found that more accurate results are achieved by encoding the entire inputs, but measures the similarity only between non-punctuation and non-whitespace tokens. To disable the removal of punctuation tokens, use no_punc=False.

See also our example.py script. Additional details are shown in the original BERTScore demo notebook.

Training

The run_mlm.py script can be used to fine-tune the base model microsoft/codebert-base-mlm on specific languages.

Evaluation

The code to reproduce the results in the paper can be found in the evaluation.

Human Evaluation

We find that CodeBERTScore is more correlated with human preference compared to a variety of common metrics. See more details in the paper.

Functional Correctness

We find that CodeBERTScore is more correlated with functional correctness compared to a variety of common metrics. See more details in the paper.

Citation

@article{zhou2023codebertscore,
  url = {https://arxiv.org/abs/2302.05527},
  author = {Zhou, Shuyan and Alon, Uri and Agarwal, Sumit and Neubig, Graham},
  title = {CodeBERTScore: Evaluating Code Generation with Pretrained Models of Code},  
  publisher = {arXiv},
  year = {2023},
}