用于问答的 BERT 中长文本的滑动窗口

Posted 2023-02-15

【中文标题】用于问答的 BERT 中长文本的滑动窗口

【英文标题】：Sliding window for long text in BERT for Question Answering

【发布时间】：2020-11-08 17:38:09

【问题描述】：

我已经阅读了解释滑动窗口如何工作的帖子，但我找不到任何关于它是如何实际实现的信息。

据我了解，如果输入太长，可以使用滑动窗口来处理文本。

如果我错了，请纠正我。 假设我有一个文本“2017 年 6 月，Kaggle 宣布注册用户超过 100 万”。

给定一些stride 和max_len，可以将输入拆分为重叠单词的块（不考虑填充）。

In June 2017 Kaggle announced that # chunk 1
announced that it passed 1 million # chunk 2
1 million registered users # chunk 3

如果我的问题是“Kaggle 何时发布公告” 和 “有多少注册用户”我可以在模型中使用 chunk 1 和 chunk 3 和 not use chunk 2 at all。不确定我是否还应该使用chunk 2 来训练模型

所以输入将是： [CLS]when did Kaggle make the announcement[SEP]In June 2017 Kaggle announced that[SEP] 和 [CLS]how many registered users[SEP]1 million registered users[SEP]

---

然后，如果我有一个没有答案的问题，我是否将它与所有块一起输入到模型中，并将开始和结束索引指示为 -1？例如“猪会飞吗？”

[CLS]can pigs fly[SEP]In June 2017 Kaggle announced that[SEP]

[CLS]can pigs fly[SEP]announced that it passed 1 million[SEP]

[CLS]can pigs fly[SEP]1 million registered users[SEP]

---

按照 cmets 中的建议，我尝试运行 squad_convert_example_to_features (source code) 来调查我上面遇到的问题，但它似乎不起作用，也没有任何文档。似乎来自拥抱脸的run_squad.py 使用squad_convert_example_to_features 和example 中的s。

from transformers.data.processors.squad import SquadResult, SquadV1Processor, SquadV2Processor, squad_convert_example_to_features
from transformers import AutoTokenizer, AutoConfig, squad_convert_examples_to_features

FILE_DIR = "."

tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
processor = SquadV2Processor()
examples = processor.get_train_examples(FILE_DIR)

features = squad_convert_example_to_features(
    example=examples[0],
    max_seq_length=384,
    doc_stride=128,
    max_query_length=64,
    is_training=True,
)

我得到了错误。

100%|██████████| 1/1 [00:00<00:00, 159.95it/s]
Traceback (most recent call last):
  File "<input>", line 25, in <module>
    sub_tokens = tokenizer.tokenize(token)
NameError: name 'tokenizer' is not defined

错误表明没有tokenizers，但它不允许我们传递tokenizer。虽然如果我在调试模式下在函数内部添加标记器，它确实有效。那么我究竟该如何使用squad_convert_example_to_features函数呢？

【问题讨论】：

run_squad 已经涵盖了所有这些内容。

好像没有做滑窗，我试了一下huggingface的squad_convert_examples_to_features的长文。

也许您可以添加您尝试过的内容，并指出为什么它不是直接针对您的问题的滑动窗口方法。相关代码为here

对不起，我跑了squad_convert_examples_to_features 的错误，它不处理带有滑动窗口的文本，它与你链接到的squad_convert_example_to_features 不同。我现在更新了我的问题。

【参考方案1】：

我认为您选择的示例存在问题。 squad_convert_examples_to_features 和 squad_convert_example_to_features 都实现了滑动窗口方法，因为 squad_convert_examples_to_features 只是 squad_convert_example_to_features 的并行化包装器。但是让我们看一下单个示例函数。首先，您需要调用 squad_convert_example_to_features_init 来使标记器成为全局的（这会在 squad_convert_examples_to_features 中自动为您完成）：

from transformers.data.processors.squad import SquadResult, SquadV1Processor, SquadV2Processor, squad_convert_examples_to_features, squad_convert_example_to_features_init
from transformers import AutoTokenizer, AutoConfig, squad_convert_examples_to_features

FILE_DIR = "."

tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
squad_convert_example_to_features_init(tokenizer)

processor = SquadV2Processor()
examples = processor.get_train_examples(FILE_DIR)

features = squad_convert_example_to_features(
    example=examples[0],
    max_seq_length=384,
    doc_stride=128,
    max_query_length=64,
    is_training=True,
)
print(len(features))

输出：

1

你可能会说这个函数没有使用滑动窗口的方法，但这是错误的，因为你的示例不需要拆分：

print(len(examples[0].question_text.split()) + len(examples[0].doc_tokens))

输出：

115

它小于您设置为 384 的 max_seq_length。现在让我们尝试不同的：

print(len(examples[129603].question_text.split()) + len(examples[129603].doc_tokens))

features = squad_convert_example_to_features(
    example=examples[129603],
    max_seq_length=384,
    doc_stride=128,
    max_query_length=64,
    is_training=True,
)
print(len(features))

输出：

454
3

您现在可以与原始样本进行比较：

print('[CLS]' + examples[129603].question_text + '[SEP]' + ' '.join(examples[129603].doc_tokens) + '[SEP]')

for idx, f in enumerate(features):
    print('Split '.format(idx))
    print(' '.join(f.tokens))

输出：

[CLS]How often is hunting occurring in Delaware each year?[SEP]There is a very active tradition of hunting of small to medium-sized wild game in Trinidad and Tobago. Hunting is carried out with firearms, and aided by the use of hounds, with the illegal use of trap guns, trap cages and snare nets. With approximately 12,000 sport hunters applying for hunting licences in recent years (in a very small country of about the size of the state of Delaware at about 5128 square kilometers and 1.3 million inhabitants), there is some concern that the practice might not be sustainable. In addition there are at present no bag limits and the open season is comparatively very long (5 months - October to February inclusive). As such hunting pressure from legal hunters is very high. Added to that, there is a thriving and very lucrative black market for poached wild game (sold and enthusiastically purchased as expensive luxury delicacies) and the numbers of commercial poachers in operation is unknown but presumed to be fairly high. As a result, the populations of the five major mammalian game species (red-rumped agouti, lowland paca, nine-banded armadillo, collared peccary, and red brocket deer) are thought to be quite low (although scientifically conducted population studies are only just recently being conducted as of 2013). It appears that the red brocket deer population has been extirpated on Tobago as a result of over-hunting. Various herons, ducks, doves, the green iguana, the gold tegu, the spectacled caiman and the common opossum are also commonly hunted and poached. There is also some poaching of 'fully protected species', including red howler monkeys and capuchin monkeys, southern tamanduas, Brazilian porcupines, yellow-footed tortoises, Trinidad piping guans and even one of the national birds, the scarlet ibis. Legal hunters pay very small fees to obtain hunting licences and undergo no official basic conservation biology or hunting-ethics training. There is presumed to be relatively very little subsistence hunting in the country (with most hunting for either sport or commercial profit). The local wildlife management authority is under-staffed and under-funded, and as such very little in the way of enforcement is done to uphold existing wildlife management laws, with hunting occurring both in and out of season, and even in wildlife sanctuaries. There is some indication that the government is beginning to take the issue of wildlife management more seriously, with well drafted legislation being brought before Parliament in 2015. It remains to be seen if the drafted legislation will be fully adopted and financially supported by the current and future governments, and if the general populace will move towards a greater awareness of the importance of wildlife conservation and change the culture of wanton consumption to one of sustainable management.[SEP]
Split 0
[CLS] how often is hunting occurring in delaware each year ? [SEP] there is a very active tradition of hunting of small to medium - sized wild game in trinidad and tobago . hunting is carried out with firearms , and aided by the use of hounds , with the illegal use of trap guns , trap cages and s ##nare nets . with approximately 12 , 000 sport hunters applying for hunting licence ##s in recent years ( in a very small country of about the size of the state of delaware at about 512 ##8 square kilometers and 1 . 3 million inhabitants ) , there is some concern that the practice might not be sustainable . in addition there are at present no bag limits and the open season is comparatively very long ( 5 months - october to february inclusive ) . as such hunting pressure from legal hunters is very high . added to that , there is a thriving and very lucrative black market for po ##ache ##d wild game ( sold and enthusiastically purchased as expensive luxury del ##ica ##cies ) and the numbers of commercial po ##ache ##rs in operation is unknown but presumed to be fairly high . as a result , the populations of the five major mammalian game species ( red - rum ##ped ago ##uti , lowland pac ##a , nine - banded arm ##adi ##llo , collar ##ed pe ##cca ##ry , and red brock ##et deer ) are thought to be quite low ( although scientific ##ally conducted population studies are only just recently being conducted as of 2013 ) . it appears that the red brock ##et deer population has been ex ##ti ##rp ##ated on tobago as a result of over - hunting . various heron ##s , ducks , dove ##s , the green i ##gua ##na , the gold te ##gu , the spectacle ##d cai ##man and the common op ##oss ##um are also commonly hunted and po ##ache ##d . there is also some po ##achi ##ng of ' fully protected species ' , including red howl ##er monkeys and cap ##uchi ##n monkeys , southern tam ##and ##ua ##s , brazilian por ##cup ##ines , yellow - footed tor ##to ##ises , [SEP]
Split 1
[CLS] how often is hunting occurring in delaware each year ? [SEP] october to february inclusive ) . as such hunting pressure from legal hunters is very high . added to that , there is a thriving and very lucrative black market for po ##ache ##d wild game ( sold and enthusiastically purchased as expensive luxury del ##ica ##cies ) and the numbers of commercial po ##ache ##rs in operation is unknown but presumed to be fairly high . as a result , the populations of the five major mammalian game species ( red - rum ##ped ago ##uti , lowland pac ##a , nine - banded arm ##adi ##llo , collar ##ed pe ##cca ##ry , and red brock ##et deer ) are thought to be quite low ( although scientific ##ally conducted population studies are only just recently being conducted as of 2013 ) . it appears that the red brock ##et deer population has been ex ##ti ##rp ##ated on tobago as a result of over - hunting . various heron ##s , ducks , dove ##s , the green i ##gua ##na , the gold te ##gu , the spectacle ##d cai ##man and the common op ##oss ##um are also commonly hunted and po ##ache ##d . there is also some po ##achi ##ng of ' fully protected species ' , including red howl ##er monkeys and cap ##uchi ##n monkeys , southern tam ##and ##ua ##s , brazilian por ##cup ##ines , yellow - footed tor ##to ##ises , trinidad pip ##ing gu ##ans and even one of the national birds , the scarlet ib ##is . legal hunters pay very small fees to obtain hunting licence ##s and undergo no official basic conservation biology or hunting - ethics training . there is presumed to be relatively very little subsistence hunting in the country ( with most hunting for either sport or commercial profit ) . the local wildlife management authority is under - staffed and under - funded , and as such very little in the way of enforcement is done to uphold existing wildlife management laws , with hunting occurring both in and out of season , and even in wildlife san ##ct ##uaries . there is some indication that the government is beginning to [SEP]
Split 2
[CLS] how often is hunting occurring in delaware each year ? [SEP] being conducted as of 2013 ) . it appears that the red brock ##et deer population has been ex ##ti ##rp ##ated on tobago as a result of over - hunting . various heron ##s , ducks , dove ##s , the green i ##gua ##na , the gold te ##gu , the spectacle ##d cai ##man and the common op ##oss ##um are also commonly hunted and po ##ache ##d . there is also some po ##achi ##ng of ' fully protected species ' , including red howl ##er monkeys and cap ##uchi ##n monkeys , southern tam ##and ##ua ##s , brazilian por ##cup ##ines , yellow - footed tor ##to ##ises , trinidad pip ##ing gu ##ans and even one of the national birds , the scarlet ib ##is . legal hunters pay very small fees to obtain hunting licence ##s and undergo no official basic conservation biology or hunting - ethics training . there is presumed to be relatively very little subsistence hunting in the country ( with most hunting for either sport or commercial profit ) . the local wildlife management authority is under - staffed and under - funded , and as such very little in the way of enforcement is done to uphold existing wildlife management laws , with hunting occurring both in and out of season , and even in wildlife san ##ct ##uaries . there is some indication that the government is beginning to take the issue of wildlife management more seriously , with well drafted legislation being brought before parliament in 2015 . it remains to be seen if the drafted legislation will be fully adopted and financially supported by the current and future governments , and if the general populace will move towards a greater awareness of the importance of wildlife conservation and change the culture of want ##on consumption to one of sustainable management . [SEP]

---

如果我的问题是“Kaggle 何时发布公告”和“如何 许多注册用户”我可以使用块 1 和块 3 而不能使用块 2 在模型中。不确定我是否还应该使用块 2 训练模型

是的，您还应该使用块 2 来训练您的模型，因为当您尝试预测相同的序列时，您希望您的模型预测 0:0 作为块 2 的答案跨度（即您可以轻松选择包含回答）。

【讨论】：

回复chunk2 问题。如果我要使用 chunk2 来训练答案跨度为 0:0 的模型，我应该将哪个问题与 chunk2 上下文配对？我应该同时使用chunk1 和chunk3 中使用的问题吗？感谢您的帮助。

您给定的示例将导致 6 个序列（每个块的问题 1 和每个块的问题 2）。通常，当您考虑模型在预测案例中应该做什么时，您可以回答这些问题。例如，问题一在第一块中只有一个答案，因此您的模型应该预测第二块和第三块没有答案。