roberta_tokenization.py

from transformers import RobertaTokenizer, TFRobertaForMultipleChoice
tokenizer = RobertaTokenizer.from_pretrained('roberta-base', do_lower_case=True)
def arc_preprocessor(dataset, tokenizer):
    '''
    This function will convert a given article, question, choices in a format:

    <s> article </s> </s> question </s> </s> choices[0] </s>
    <s> article </s> </s> question </s> </s> choices[1] </s>
    <s> article </s> </s> question </s> </s> choices[2] </s>
    <s> article </s> </s> question </s> </s> choices[3] </s>

    After converting in this format the data will be tokenized using a given tokenizer.
    This function will return 4 arrays namely, input_ids, attention_mask, token_type_ids and labels.

    individual input_ids, token_type_ids, attention_mask shape will be as: [num_choices, max_seq_length]
    '''
    global MAX_LEN
    all_input_ids = []
    all_attention_mask = []
    all_token_type_ids = []

    for i in range(len(dataset)):
        
        context = ' '.join(dataset['context'].iloc[i].split()[:400]) #Limiting the maximum length of context to 400 words.
        question = dataset['only_question'].iloc[i]
        options = dataset['options_list'].iloc[i]
        choice_features = []

        for j in range(len(options)):

            option = options[j]
            input_string = '<s>' + ' ' + context + ' ' + '</s>' + ' ' + '</s>' + ' ' + question + ' ' + '</s>' + ' ' + '</s>' + ' ' + option + ' ' + '</s>'
            input_string = re.sub(r'\s+', ' ', input_string)

            input_ids = tokenizer(input_string,
                               max_length=MAX_LEN,
                               add_special_tokens=False, truncation='only_first')['input_ids']
            
            attention_mask = [1] * len(input_ids)

            padding_id = tokenizer.pad_token_id
            padding_length = MAX_LEN - len(input_ids)

            input_ids = input_ids + [padding_id]*padding_length
            attention_mask = attention_mask + [0]*padding_length

            assert len(input_ids) == MAX_LEN
            assert len(attention_mask) == MAX_LEN

            choice_features.append({'input_ids':input_ids,
                                    'attention_mask':attention_mask})

        all_input_ids.append(np.asarray([cf['input_ids'] for cf in choice_features], dtype='int32'))
        all_attention_mask.append(np.asarray([cf['attention_mask'] for cf in choice_features], dtype='int32'))

    return all_input_ids, all_attention_mask

ok