1. 背景
bert_score是一个用于评估文本生成质量的Python库,基于预训练的BERT模型计算参考文本与生成文本之间的语义相似度。它通过比较两个文本的BERT嵌入向量来评估它们的匹配程度,比传统的n-gram匹配方法(如BLEU、ROUGE)更能捕捉语义层面的相似性。一般bert_score 是仅仅支持指定模型名称然后再代码实现里面从Hugging Face上面拉取模型。但是如果我们在有限制的网络环境中在线拉取模型就会失败,比较合适的解决办法就是先将模型下载到本地,然后给bert_score指定模型路径。
2. 支持的模型和语言
bert_score中通过model_type和lang两个参数来执行模型和对应的文本语言。
具体文档可以参考:https://github.com/Tiiiger/bert_score
from bert_score import score candidates = ["The cat sits on the mat."] # 生成文本 references = ["A cat is sitting on the rug."] # 参考文本 # 计算BERTScore P, R, F1 = score(candidates, references, model_type='roberta-large',lang="en") print(f"Precision: {P.mean():.3f}, Recall: {R.mean():.3f}, F1: {F1.mean():.3f}")3.支持本地模型的方法
3.1 修改源码
1. 在score方法里面可以看到get_model和get_tokenizer两个方法
2.点进去可以看到对应函数的源码
def get_model(model_type, num_layers, all_layers=None): if model_type.startswith("scibert"): model = AutoModel.from_pretrained(cache_scibert(model_type)) elif "t5" in model_type: from transformers import T5EncoderModel model = T5EncoderModel.from_pretrained(model_type) else: model = AutoModel.from_pretrained(model_type) model.eval() if hasattr(model, "decoder") and hasattr(model, "encoder"): model = model.encoder # drop unused layers # ...... return model def get_tokenizer(model_type, use_fast=False): if model_type.startswith("scibert"): model_type = cache_scibert(model_type) if version.parse(trans_version) >= version.parse("4.0.0"): tokenizer = AutoTokenizer.from_pretrained(model_type, use_fast=use_fast) else: assert not use_fast, "Fast tokenizer is not available for version < 4.0.0" tokenizer = AutoTokenizer.from_pretrained(model_type) return tokenizer3. 修改源码直接将我们的model_path硬编码进去,当然为了更好的适配性,可以将model_path做成一个可选参数,然后传递出去给score函数
def get_model(model_type, num_layers, all_layers=None): model_path = 'xxx' if model_type.startswith("scibert"): model = AutoModel.from_pretrained(cache_scibert(model_type)) elif "t5" in model_type: from transformers import T5EncoderModel model = T5EncoderModel.from_pretrained(model_path) else: model = AutoModel.from_pretrained(model_path) model.eval() def get_tokenizer(model_type, use_fast=False): if model_type.startswith("scibert"): model_type = cache_scibert(model_type) model_path = 'xxx' if version.parse(trans_version) >= version.parse("4.0.0"): tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=use_fast) else: assert not use_fast, "Fast tokenizer is not available for version < 4.0.0" tokenizer = AutoTokenizer.from_pretrained(model_path) return tokenizer3.2 添加对应的num_layers参数
这里还有一种方法就是通过给model_type传递给本地模型所在的路径。同时传递num_layers参数来指定。
原因是因为其实Transformers里面的xxx.from_pretrained方法其实也是可以支持你配置本地路径来加载的,这里加载模型也不会有问题。但是在bert_score中他是需要提前知道层数来优化计算过程的。这里才是报错的原因。
这一部分代码是在utils文件中。
def calculate_bert_score(original_text: str, polished_text: str, lang: str = "zh",model_type:str=None) -> float: try: # 使用bert_score计算F1分数 score_params = { "verbose": False, "device": "cuda" if torch.cuda.is_available() else "cpu" } if model_type is None: score_params["model_type"] = BERT_MODEL_PATH #指定模型路径 score_params["num_layers"] = BERT_BASE_CHINESE_NUM_LAYERS#执行模型层数(12) else: score_params["model_type"] = model_type P, R, F1 = score([polished_text], [original_text], **score_params) # 返回F1分数(转换为Python float) f1_score = F1.item() if hasattr(F1, 'item') else float(F1) return f1_score except Exception as e: print(f"计算bert_score时出错: {str(e)}") # 如果计算失败,返回一个默认值或者抛出异常 raise