tner下载 - tner源代码下载

T-NER：一个全面的Python库，用于基于变压器的命名实体识别

T-NER是通过PIP上实施的Python工具，用于在Pytorch中实现的命名 - 实体识别（NER）。它具有轻松的界面，可用于芬太纳模型并在跨域和多语言数据集上进行测试。 T-NER当前集成了公开可用的NER数据集的高度覆盖范围，并可以简单地集成自定义数据集。所有使用T-NER固定的模型都可以在我们的Web应用程序上部署，以进行可视化。我们的论文证明了T-NER已被EACL 2021接受。所有模型和数据集均通过T-Ner HuggingFace组共享。

New（2022年9月）：我们根据Twitter tweetner7发布了新的NER数据集，该论文被AACL 2022主要会议所接受！我们将数据集与微调模型一起发布，并在论文，存储库和数据集页面上找到更多详细信息。 Twitter NER模型也已集成到TweetNLP中，并在此处提供演示。

资源： model_card ， dataset_card ， Gradio在线演示
拥抱面： https：//huggingface.co/tner
github： https：//github.com/asahi417/tner
文件
- T-NER（EACL2021）： ACL选集， ARXIV
- Tweetner7（AACL 2022）： arxiv

通过PIP安装tner即可开始！

pip install tner

描述	关联
模型登录和评估
模型预测
多语言NER工作流程

数据集

NER数据集包含每个分型的令牌和标签序列（通常是train / validation / test ），

{
    'train' : {
        'tokens' : [
            [ '@paulwalk' , 'It' , "'s" , 'the' , 'view' , 'from' , 'where' , 'I' , "'m" , 'living' , 'for' , 'two' , 'weeks' , '.' , 'Empire' , 'State' , 'Building' , '=' , 'ESB' , '.' , 'Pretty' , 'bad' , 'storm' , 'here' , 'last' , 'evening' , '.' ],
            [ 'From' , 'Green' , 'Newsfeed' , ':' , 'AHFA' , 'extends' , 'deadline' , 'for' , 'Sage' , 'Award' , 'to' , 'Nov' , '.' , '5' , 'http://tinyurl.com/24agj38' ], ...
        ],
        'tags' : [
            [ 0 , 0 , 0 , 0 , 0 , 0 , 0 , 0 , 0 , 0 , 0 , 0 , 0 , 0 , 1 , 2 , 2 , 0 , 1 , 0 , 0 , 0 , 0 , 0 , 0 , 0 , 0 ],
            [ 0 , 0 , 0 , 0 , 3 , 0 , 0 , 0 , 0 , 0 , 0 , 0 , 0 , 0 , 0 ], ...
        ]
    },
    'validation' : ...,
    'test' : ...,
}

用字典将标签映射到其索引（ label2id ）如下。

{ "O" : 0 , "B-ORG" : 1 , "B-MISC" : 2 , "B-PER" : 3 , "I-PER" : 4 , "B-LOC" : 5 , "I-ORG" : 6 , "I-MISC" : 7 , "I-LOC" : 8 }

预设数据集

我们的HuggingFace组可用各种公共数据集，可以按照下面的方式使用（有关完整数据集列表，请参见数据集卡）。

 from tner import get_dataset
data , label2id = get_dataset ( dataset = "tner/wnut2017" )

用户可以指定多个数据集以获取串联数据集。

 data , label2id = get_dataset ( dataset = [ "tner/conll2003" , "tner/ontonotes5" ])

在串联数据集中，我们使用统一标签集来统一实体标签。这个想法是以统一格式在拥抱面上共享所有可用的NER数据集，因此，如果您希望在此处添加任何NER数据集！

自定义数据集

为了超越公共数据集，用户可以通过将其格式化为Conll 2003 NER共享任务文件中所述的IOB格式来使用自己的数据集，其中所有数据文件都包含一个单词一个单词，带有代表句子边界的空行。在每行的末尾，都有一个标签，该标签说明当前单词是否在命名实体内部。标签还编码命名实体的类型。这是一个示例句子：

 EU B-ORG
rejects O
German B-MISC
call O
to O
boycott O
British B-MISC
lamb O
. O

用o标记的单词在命名实体之外，i-xxx标签用于XXX类型的命名实体内部的单词。每当XXX类型的两个实体彼此隔壁时，第二个实体的第一个单词将被标记为B-XXX，以表明它启动了另一个实体。请查看示例自定义数据。这些自定义文件的加载方式与下面的拥抱面数据集相同。

 from tner import get_dataset
data , label2id = get_dataset ( local_dataset = {
    "train" : "examples/local_dataset_sample/train.txt" ,
    "valid" : "examples/local_dataset_sample/train.txt" ,
    "test" : "examples/local_dataset_sample/test.txt"
})

与HuggingFace数据集相同，可以连接数据集。

 data , label2id = get_dataset ( local_dataset = [
   { "train" : "..." , "valid" : "..." , "test" : "..." },
   { "train" : "..." , "valid" : "..." , "test" : "..." }
   ]
)

模型

如上表所示，T-NER当前在HuggingFace组上共享了100多个NER模型，该表仅报告主要模型，并查看Model_card以获取完整模型列表。所有模型都可以与tner一起使用如下。

 from tner import TransformersNER
model = TransformersNER ( "tner/roberta-large-wnut2017" )  # provide model alias on huggingface
output = model . predict ([ "Jacob Collier is a Grammy awarded English artist from London" ])  # give a list of sentences (or tokenized sentence) 
print ( output )
{
   'prediction' : [[ 'B-person' , 'I-person' , 'O' , 'O' , 'O' , 'O' , 'O' , 'O' , 'O' , 'B-location' ]],
   'probability' : [[ 0.9967652559280396 , 0.9994561076164246 , 0.9986955523490906 , 0.9947081804275513 , 0.6129112243652344 , 0.9984312653541565 , 0.9868122935295105 , 0.9983410835266113 , 0.9995284080505371 , 0.9838910698890686 ]],
   'input' : [[ 'Jacob' , 'Collier' , 'is' , 'a' , 'Grammy' , 'awarded' , 'English' , 'artist' , 'from' , 'London' ]],
   'entity_prediction' : [[
       { 'type' : 'person' , 'entity' : [ 'Jacob' , 'Collier' ], 'position' : [ 0 , 1 ], 'probability' : [ 0.9967652559280396 , 0.9994561076164246 ]},
       { 'type' : 'location' , 'entity' : [ 'London' ], 'position' : [ 9 ], 'probability' : [ 0.9838910698890686 ]}
    ]]
}

model.predict选出句子列表和批处理大小batch_size ，并通过半空间或separator指定的符号将句子归为句子，该句子在其输出对象中返回为input 。可选地，用户可以通过任何令牌（Spacy，nltk等）事先进行输入，并且预测将遵循令牌化。

 output = model . predict ([[ "Jacob Collier" , "is" , "a" , "Grammy awarded" , "English artist" , "from" , "London" ]])
print ( output )
{
    'prediction' : [[ 'B-person' , 'O' , 'O' , 'O' , 'O' , 'O' , 'B-location' ]],
    'probability' : [[ 0.9967652559280396 , 0.9986955523490906 , 0.9947081804275513 , 0.6129112243652344 , 0.9868122935295105 , 0.9995284080505371 , 0.9838910698890686 ]],
    'input' : [[ 'Jacob Collier' , 'is' , 'a' , 'Grammy awarded' , 'English artist' , 'from' , 'London' ]],
    'entity_prediction' : [[
        { 'type' : 'person' , 'entity' : [ 'Jacob Collier' ], 'position' : [ 0 ], 'probability' : [ 0.9967652559280396 ]},
        { 'type' : 'location' , 'entity' : [ 'London' ], 'position' : [ 6 ], 'probability' : [ 0.9838910698890686 ]}
    ]]
}

可以指定本地模型检查点，而不是模型别名TransformersNER("path-to-checkpoint") 。重新生产这些已发布模型的脚本在这里。

命令行工具

以下命令行工具可用于模型预测。

tner-predict [-h] -m MODEL

command line tool to test finetuned NER model

optional arguments:
  -h, --help            show this help message and exit
  -m MODEL, --model MODEL
                        model alias of huggingface or local checkpoint

例子

tner-predict -m " tner/roberta-large-wnut2017 "

Web应用程序

要安装依赖项以运行Web应用程序，请在安装时添加选项。

pip install tner[app]

然后，克隆存储库

git clone https://github.com/asahi417/tner
cd tner

并启动服务器。

uvicorn app:app --reload --log-level debug --host 0.0.0.0 --port 8000

准备就绪后打开浏览器http://0.0.0.0:8000。您可以指定通过环境变量NER_MODEL部署的模型，该模型默认设置为tner/roberta-large-wnut2017 。 NER_MODEL可以是您本地模型检查点目录的路径，也可以是变形金刚模型中心上的模型名称。

确认应用程序接口受此存储库的启发。

NER上的微调语言模型

T-NER提供了一个简单的API，可以通过上述有效的参数搜索在NER上进行语言模型进行微调。它由2个阶段组成：（i）对所有模型的验证集对小时期的所有可能配置进行微调，并计算评估度量度（默认为默认值），以及（ii）拾取顶部K型号以继续进行微调直至L Epoch。第二阶段的最佳模型将继续进行微调，直到验证度量减少为止。

使用tner的几行可以在几行中实现使用两阶段参数搜索的微调。

 from tner import GridSearcher
searcher = GridSearcher (
   checkpoint_dir = './ckpt_tner' ,
   dataset = "tner/wnut2017" ,  # either of `dataset` (huggingface dataset) or `local_dataset` (custom dataset) should be given
   model = "roberta-large" ,  # language model to fine-tune
   epoch = 10 ,  # the total epoch (`L` in the figure)
   epoch_partial = 5 ,  # the number of epoch at 1st stage (`M` in the figure)
   n_max_config = 3 ,  # the number of models to pass to 2nd stage (`K` in the figure)
   batch_size = 16 ,
   gradient_accumulation_steps = [ 4 , 8 ],
   crf = [ True , False ],
   lr = [ 1e-4 , 1e-5 ],
   weight_decay = [ 1e-7 ],
   random_seed = [ 42 ],
   lr_warmup_step_ratio = [ 0.1 ],
   max_grad_norm = [ 10 ]  
)
searcher . train ()

以下参数目前是可调的。

gradient_accumulation_steps ：梯度积累的数量
crf ：在输出嵌入式上使用CRF
lr ：学习率
weight_decay ：重量衰减系数
random_seed ：随机种子
lr_warmup_step_ratio ：学习率的线性热身比），例如，如果是0.3，则学习率将线性热身到总步骤的30％（毕竟没有衰减）
max_grad_norm ：梯度剪辑的规范

有关每个参数的更多信息，请参见来源。

命令行工具

以下命令行工具可用于微调。

tner-train-search [-h] -c CHECKPOINT_DIR [-d DATASET [DATASET ...]] [-l LOCAL_DATASET [LOCAL_DATASET ...]]
                         [--dataset-name DATASET_NAME [DATASET_NAME ...]] [-m MODEL] [-b BATCH_SIZE] [-e EPOCH] [--max-length MAX_LENGTH] [--use-auth-token]
                         [--dataset-split-train DATASET_SPLIT_TRAIN] [--dataset-split-valid DATASET_SPLIT_VALID] [--lr LR [LR ...]]
                         [--random-seed RANDOM_SEED [RANDOM_SEED ...]] [-g GRADIENT_ACCUMULATION_STEPS [GRADIENT_ACCUMULATION_STEPS ...]]
                         [--weight-decay WEIGHT_DECAY [WEIGHT_DECAY ...]] [--lr-warmup-step-ratio LR_WARMUP_STEP_RATIO [LR_WARMUP_STEP_RATIO ...]]
                         [--max-grad-norm MAX_GRAD_NORM [MAX_GRAD_NORM ...]] [--crf CRF [CRF ...]] [--optimizer-on-cpu] [--n-max-config N_MAX_CONFIG]
                         [--epoch-partial EPOCH_PARTIAL] [--max-length-eval MAX_LENGTH_EVAL]

Fine-tune transformers on NER dataset with Robust Parameter Search

optional arguments:
  -h , --help            show this help message and exit
  -c CHECKPOINT_DIR, --checkpoint-dir CHECKPOINT_DIR
                        checkpoint directory
  -d DATASET [DATASET ...], --dataset DATASET [DATASET ...]
                        dataset name (or a list of it) on huggingface tner organization eg. ' tner/conll2003 ' [ ' tner/conll2003 ' , ' tner/ontonotes5 ' ]] see
                        https://huggingface.co/datasets ? search = tner for full dataset list
  -l LOCAL_DATASET [LOCAL_DATASET ...], --local-dataset LOCAL_DATASET [LOCAL_DATASET ...]
                        a dictionary (or a list) of paths to local BIO files eg.{ " train " : " examples/local_dataset_sample/train.txt " , " test " :
                        " examples/local_dataset_sample/test.txt " }
  --dataset-name DATASET_NAME [DATASET_NAME ...]
                        [optional] data name of huggingface dataset (should be same length as the ` dataset ` )
  -m MODEL, --model MODEL
                        model name of underlying language model (huggingface model)
  -b BATCH_SIZE, --batch-size BATCH_SIZE
                        batch size
  -e EPOCH, --epoch EPOCH
                        the number of epoch
  --max-length MAX_LENGTH
                        max length of language model
  --use-auth-token      Huggingface transformers argument of ` use_auth_token `
  --dataset-split-train DATASET_SPLIT_TRAIN
                        dataset split to be used for training ( ' train ' as default)
  --dataset-split-valid DATASET_SPLIT_VALID
                        dataset split to be used for validation ( ' validation ' as default)
  --lr LR [LR ...]      learning rate
  --random-seed RANDOM_SEED [RANDOM_SEED ...]
                        random seed
  -g GRADIENT_ACCUMULATION_STEPS [GRADIENT_ACCUMULATION_STEPS ...], --gradient-accumulation-steps GRADIENT_ACCUMULATION_STEPS [GRADIENT_ACCUMULATION_STEPS ...]
                        the number of gradient accumulation
  --weight-decay WEIGHT_DECAY [WEIGHT_DECAY ...]
                        coefficient of weight decay (set 0 for None)
  --lr-warmup-step-ratio LR_WARMUP_STEP_RATIO [LR_WARMUP_STEP_RATIO ...]
                        linear warmup ratio of learning rate (no decay).eg) if it ' s 0.3, the learning rate will warmup linearly till 30% of the total step
                        (set 0 for None)
  --max-grad-norm MAX_GRAD_NORM [MAX_GRAD_NORM ...]
                        norm for gradient clipping (set 0 for None)
  --crf CRF [CRF ...]   use CRF on top of output embedding (0 or 1)
  --optimizer-on-cpu    put optimizer on CPU to save memory of GPU
  --n-max-config N_MAX_CONFIG
                        the number of configs to run 2nd phase search
  --epoch-partial EPOCH_PARTIAL
                        the number of epoch for 1st phase search
  --max-length-eval MAX_LENGTH_EVAL
                        max length of language model at evaluation

例子

tner-train-search -m " roberta-large " -c " ckpt " -d " tner/wnut2017 " -e 15 --epoch-partial 5 --n-max-config 3 -b 64 -g 1 2 --lr 1e-6 1e-5 --crf 0 1 --max-grad-norm 0 10 --weight-decay 0 1e-7

评估NER模型

NER模型的评估是通过model.evaluate完成的，该功能将dataset或local_dataset作为用于评估的数据集。

 from tner import TransformersNER
model = TransformersNER ( "tner/roberta-large-wnut2017" )  # provide model alias on huggingface
# huggingface dataset
metric = model . evaluate ( 'tner/wnut2017' , dataset_split = 'test' )
# local dataset
metric = model . evaluate ( local_dataset = { "test" : "examples/local_dataset_sample/test.txt" }, dataset_split = 'test' )

可以在此处找到输出对象metric的示例。

实体跨度预测

为了更好地理解室外准确性，我们提供了实体跨度预测管道，该管道忽略了实体类型和计算指标，仅在IOB实体位置（二进制序列标记）上。

 metric = model . evaluate ( datasets = 'tner/wnut2017' , dataset_split = 'test' , span_detection_mode = True )

命令行工具

以下命令行工具可用于模型预测。

tner-evaluate [-h] -m MODEL -e EXPORT [-d DATASET [DATASET ...]] [-l LOCAL_DATASET [LOCAL_DATASET ...]]
                     [--dataset-name DATASET_NAME [DATASET_NAME ...]] [--dataset-split DATASET_SPLIT] [--span-detection-mode] [--return-ci] [-b BATCH_SIZE]

Evaluate NER model

optional arguments:
  -h , --help            show this help message and exit
  -m MODEL, --model MODEL
                        model alias of huggingface or local checkpoint
  -e EXPORT, --export EXPORT
                        file to export the result
  -d DATASET [DATASET ...], --dataset DATASET [DATASET ...]
                        dataset name (or a list of it) on huggingface tner organization eg. ' tner/conll2003 ' [ ' tner/conll2003 ' , ' tner/ontonotes5 ' ]] see
                        https://huggingface.co/datasets ? search = tner for full dataset list
  -l LOCAL_DATASET [LOCAL_DATASET ...], --local-dataset LOCAL_DATASET [LOCAL_DATASET ...]
                        a dictionary (or a list) of paths to local BIO files eg.{ " train " : " examples/local_dataset_sample/train.txt " , " test " :
                        " examples/local_dataset_sample/test.txt " }
  --dataset-name DATASET_NAME [DATASET_NAME ...]
                        [optional] data name of huggingface dataset (should be same length as the ` dataset ` )
  --dataset-split DATASET_SPLIT
                        dataset split to be used for test ( ' test ' as default)
  --span-detection-mode
                        return F1 of entity span detection (ignoring entity type error and cast as binary sequence classification as below)- NER : [ " O " ,
                        " B-PER " , " I-PER " , " O " , " B-LOC " , " O " , " B-ORG " ]- Entity-span detection: [ " O " , " B-ENT " , " I-ENT " , " O " , " B-ENT " , " O " , " B-ENT " ]
  --return-ci           return confidence interval by bootstrap
  -b BATCH_SIZE, --batch-size BATCH_SIZE
                        batch size

例子

tner-evaluate -m " tner/roberta-large-wnut2017 " -e " metric.json " -d " tner/conll2003 " -b " 32 "

参考

如果您使用这些资源中的任何一个，请引用以下论文：

 @inproceedings{ushio-camacho-collados-2021-ner,
    title = "{T}-{NER}: An All-Round Python Library for Transformer-based Named Entity Recognition",
    author = "Ushio, Asahi  and
      Camacho-Collados, Jose",
    booktitle = "Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: System Demonstrations",
    month = apr,
    year = "2021",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2021.eacl-demos.7",
    pages = "53--62",
    abstract = "Language model (LM) pretraining has led to consistent improvements in many NLP downstream tasks, including named entity recognition (NER). In this paper, we present T-NER (Transformer-based Named Entity Recognition), a Python library for NER LM finetuning. In addition to its practical utility, T-NER facilitates the study and investigation of the cross-domain and cross-lingual generalization ability of LMs finetuned on NER. Our library also provides a web app where users can get model predictions interactively for arbitrary text, which facilitates qualitative model evaluation for non-expert programmers. We show the potential of the library by compiling nine public NER datasets into a unified format and evaluating the cross-domain and cross- lingual performance across the datasets. The results from our initial experiments show that in-domain performance is generally competitive across datasets. However, cross-domain generalization is challenging even with a large pretrained LM, which has nevertheless capacity to learn domain-specific features if fine- tuned on a combined dataset. To facilitate future research, we also release all our LM checkpoints via the Hugging Face model hub.",
}

展开