Improved BERTScore for image captioning evaluation

Implementation of paper: Improving Image Captioning Evaluation by Considering Inter References Variance (ACL2020)

Usage:

Recently, this repo provides two metrics ('with BERT' and 'simple')

• python3 run_metric.py

• python3 run_metric_simple.py

example data:

example/example.json (you can modify this file for your own datasets)

Fields explanation:

• "refs": reference captions (each sample 5 references)
• "cand": candidate caption (each sample 1 candidate)
• "refs_hid": contextual embeddings of reference captions
• "cand_hid": contextual embeddings of cand captions
• "mismatch": mismatches marks computed from all of reference captions
• "metric_result": scores on our metric

NOTE:
we also provide Flickr 8K Expert Annotation file with our format 'example/flickr.json'
you can easily reproduce our result following run_metric.py lines 223-235.

Dependencies:

pytorch-pretrained-bert==0.6.2 (old version of huggingface/transformers)
torch==0.4.1
bert_score==0.1.2 (already in this repo)