arian-askari / bem_score_pytorch

Answer Equivalence BEM score example in PyTorch using Huggingface Tokenizer

Answer Equivalence BEM Score

Overview

In the following you will find and example of running the Answer Equivalence BEM score using the Huggingface Tokenizer and Pytorch.

This requires the following libraries: torch, tensorflow_hub and transformers. For the installation use pip:

pip install -r requirements.py

BEM score example:

import torch
from torch.nn import functional as F
import tensorflow_hub as hub
from transformers import BertTokenizer

tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')

def bertify_example(example, max_length=512):
    question = tokenizer.tokenize(example['question'])
    reference = tokenizer.tokenize(example['reference'])
    candidate = tokenizer.tokenize(example['candidate'])
    
    tokens = ['[CLS]'] + candidate + ['[SEP]'] + reference + ['[SEP]'] + question + ['[SEP]']
    
    input_ids = torch.tensor(tokenizer.convert_tokens_to_ids(tokens))
    segment_ids = torch.tensor([0] * (len(candidate) + 2) + [1] * (len(reference) + 1) + [2] * (len(question) + 1))

    input_ids = F.pad(torch.tensor(input_ids), (0, max_length - len(input_ids)), value=0)
    segment_ids = F.pad(torch.tensor(segment_ids), (0, max_length - len(segment_ids)), value=0)
    
    return {'input_ids': input_ids, 'segment_ids': segment_ids}


def bertify_examples(examples, max_length=512):
    input_ids = []
    segment_ids = []
    for example in examples:
        example_inputs = bertify_example(example, max_length=max_length)
        input_ids.append(example_inputs['input_ids'])
        segment_ids.append(example_inputs['segment_ids'])

    return {'input_ids': torch.stack(input_ids), 'segment_ids': torch.stack(segment_ids)}

# Loading the TensorFlow Hub model
bem = hub.load('https://tfhub.dev/google/answer_equivalence/bem/1')

# Example
examples = [{
    'question': 'why is the sky blue',
    'reference': 'light scattering',
    'candidate': 'scattering of light'
}]

# Prepare inputs in PyTorch format
inputs = bertify_examples(examples, max_length=tokenizer.model_max_length)

# The outputs are raw logits.
raw_outputs = bem(inputs)
raw_outputs_torch = torch.from_numpy(raw_outputs.numpy())
# They can be transformed into a classification 'probability' like so:
bem_score = float(F.softmax(raw_outputs_torch, dim=1)[:, 1].item())

print(f'BEM score: {bem_score}')
del bem

BEM score: 0.9891803860664368

Credits

Please see original paper Tomayto, Tomahto. Beyond Token-level Answer Equivalence for Question Answering Evaluation

Original implementation in Tensorflow:

https://www.kaggle.com/models/google/bert/frameworks/TensorFlow2/variations/answer-equivalence-bem/versions/1