如何从Hugging Face Transformer问答示例代码中获取答案置信度得分

现有代码完全可以实现置信度得分返回，你只需要在原有逻辑基础上增加少量计算即可：

实现原理 #

你代码中已经获取到的answer_start_scores和answer_end_scores是问答模型输出的原始logits值，分别对应输入序列中每个token作为答案起始、结束位置的打分，只需要将这两个打分通过softmax转换为概率分布，取答案对应起止位置的概率相乘，即可得到该答案的置信度，这个计算逻辑和Hugging Face官方pipeline的置信度计算规则一致。

修改后的核心代码 #

from transformers import AutoTokenizer, TFAutoModelForQuestionAnswering
import tensorflow as tf

tokenizer = AutoTokenizer.from_pretrained("bert-large-uncased-whole-word-masking-finetuned-squad")
model = TFAutoModelForQuestionAnswering.from_pretrained("bert-large-uncased-whole-word-masking-finetuned-squad")

text = r"""
🤗 Transformers (前称pytorch-transformers和pytorch-pretrained-bert)提供了面向自然语言理解(NLU)和自然语言生成(NLG)的通用架构（包括BERT、GPT-2、RoBERTa、XLM、DistilBert、XLNet等），支持100多种语言的超过32个预训练模型，同时可在TensorFlow 2.0和PyTorch之间无缝切换使用。
"""

questions = [
    "Transformers库中提供了多少个预训练模型？",
    "Transformers库提供了什么能力？",
    "Transformers库支持哪些框架之间的互操作？",
]

for question in questions:
    inputs = tokenizer.encode_plus(question, text, add_special_tokens=True, return_tensors="tf")
    input_ids = inputs["input_ids"].numpy()[0]

    text_tokens = tokenizer.convert_ids_to_tokens(input_ids)
    answer_start_scores, answer_end_scores = model(inputs)

    answer_start = tf.argmax(
        answer_start_scores, axis=1
    ).numpy()[0]  # 取打分最高的位置作为答案起始下标
    answer_end = (
        tf.argmax(answer_end_scores, axis=1) + 1
    ).numpy()[0]  # 取打分最高的位置作为答案结束下标
    answer = tokenizer.convert_tokens_to_string(tokenizer.convert_ids_to_tokens(input_ids[answer_start:answer_end]))
    
    # 新增置信度计算逻辑
    start_probs = tf.nn.softmax(answer_start_scores, axis=1)
    end_probs = tf.nn.softmax(answer_end_scores, axis=1)
    start_conf = tf.reduce_max(start_probs).numpy()
    end_conf = tf.reduce_max(end_probs).numpy()
    total_confidence = start_conf * end_conf

    print(f"问题: {question}")
    print(f"答案: {answer}")
    print(f"置信度: {total_confidence:.4f}\n")

内容的提问来源于stack exchange，提问作者Abhishek K