您或许知道,作者后续分享网络安全的文章会越来越少。但如果您想学习人工智能和安全结合的应用,您就有福利了,作者将重新打造一个《当人工智能遇上安全》系列博客,详细介绍人工智能与安全相关的论文、实践,并分享各种案例,涉及恶意代码检测、恶意请求识别、入侵检测、对抗样本等等。只想更好地帮助初学者,更加成体系的分享新知识。该系列文章会更加聚焦,更加学术,更加深入,也是作者的慢慢成长史。换专业确实挺难的,系统安全也是块硬骨头,但我也试试,看看自己未来四年究竟能将它学到什么程度,漫漫长征路,偏向虎山行。享受过程,一起加油~
前文讲解如何实现威胁情报实体识别,利用BiLSTM-CRF算法实现对ATT&CK相关的技战术实体进行提取,是安全知识图谱构建的重要支撑。这篇文章将以中文语料为主,介绍中文命名实体识别研究,并构建BiGRU-CRF模型实现。基础性文章,希望对您有帮助,如果存在错误或不足之处,还请海涵。且看且珍惜!
由于上一篇文章详细讲解ATT&CK威胁情报采集、预处理、BiLSTM-CRF实体识别内容,这篇文章不再详细介绍,本文将在上一篇文章基础上补充:
- 中文命名实体识别如何实现,以字符为主
- 以中文CSV文件为语料,介绍其处理过程,中文威胁情报类似
- 构建BiGRU-CRF模型实现中文实体识别
版本信息:
- keras-contrib V2.0.8
- keras V2.3.1
- tensorflow V2.2.0
常见框架如下图所示:
- https://aclanthology.org/2021.acl-short.4/
文章目录
- 一.ATT&CK数据采集
- 二.数据预处理
- 三.基于BiLSTM-CRF的实体识别
- 1.安装keras-contrib
- 2.安装Keras
- 3.中文实体识别
- 四.基于BiGRU-CRF的实体识别
- 五.总结
作者作为网络安全的小白,分享一些自学基础教程给大家,主要是在线笔记,希望您们喜欢。同时,更希望您能与我一起操作和进步,后续将深入学习AI安全和系统安全知识并分享相关实验。总之,希望该系列文章对博友有所帮助,写文不易,大神们不喜勿喷,谢谢!如果文章对您有帮助,将是我创作的最大动力,点赞、评论、私聊均可,一起加油喔!
前文推荐:
- [当人工智能遇上安全] 1.人工智能真的安全吗?浙大团队外滩大会分享AI对抗样本技术
- [当人工智能遇上安全] 2.清华张超老师 – GreyOne: Discover Vulnerabilities with Data Flow Sensitive Fuzzing
- [当人工智能遇上安全] 3.安全领域中的机器学习及机器学习恶意请求识别案例分享
- [当人工智能遇上安全] 4.基于机器学习的恶意代码检测技术详解
- [当人工智能遇上安全] 5.基于机器学习算法的主机恶意代码识别研究
- [当人工智能遇上安全] 6.基于机器学习的入侵检测和攻击识别——以KDD CUP99数据集为例
- [当人工智能遇上安全] 7.基于机器学习的安全数据集总结
- [当人工智能遇上安全] 8.基于API序列和机器学习的恶意家族分类实例详解
- [当人工智能遇上安全] 9.基于API序列和深度学习的恶意家族分类实例详解
- [当人工智能遇上安全] 10.威胁情报实体识别之基于BiLSTM-CRF的实体识别万字详解
- [当人工智能遇上安全] 11.威胁情报实体识别 (2)基于BiGRU-CRF的中文实体识别万字详解
作者的github资源:
- https://github.com/eastmountyxz/When-AI-meet-Security
- https://github.com/eastmountyxz/AI-Security-Paper
一.ATT&CK数据采集
了解威胁情报的同学,应该都熟悉Mitre的ATT&CK网站,前文已介绍如何采集该网站APT组织的攻击技战术数据。网址如下:
- http://attack.mitre.org
第一步,通过ATT&CK网站源码分析定位APT组织名称,并进行系统采集。
安装BeautifulSoup扩展包,该部分代码如下所示:
01-get-aptentity.py
#encoding:utf-8#By:Eastmount CSDNimport reimport requestsfrom lxml import etreefrom bs4 import BeautifulSoupimport urllib.request#-------------------------------------------------------------------------------------------#获取APT组织名称及链接#设置浏览器代理,它是一个字典headers = {'User-Agent':'Mozilla/5.0 (Windows NT 10.0; Win64; x64) \AppleWebKit/537.36 (KHTML, like Gecko) Chrome/80.0.3987.149 Safari/537.36'}url = 'https://attack.mitre.org/groups/'#向服务器发出请求r = requests.get(url = url, headers = headers).text#解析DOM树结构html_etree = etree.HTML(r)names = html_etree.xpath('//*[@class="table table-bordered table-alternate mt-2"]/tbody/tr/td[2]/a/text()')print (names)print(len(names),names[0])filename = []for name in names:filename.append(name.strip())print(filename)#链接urls = html_etree.xpath('//*[@class="table table-bordered table-alternate mt-2"]/tbody/tr/td[2]/a/@href')print(urls)print(len(urls), urls[0])print("\n")
此时输出结果如下图所示,包括APT组织名称及对应的URL网址。
第二步,访问APT组织对应的URL,采集详细信息(正文描述)。
第三步,采集对应的技战术TTPs信息,其源码定位如下图所示。
第四步,编写代码完成威胁情报数据采集。01-spider-mitre.py 完整代码如下:
#encoding:utf-8#By:Eastmount CSDNimport reimport requestsfrom lxml import etreefrom bs4 import BeautifulSoupimport urllib.request#-------------------------------------------------------------------------------------------#获取APT组织名称及链接#设置浏览器代理,它是一个字典headers = {'User-Agent':'Mozilla/5.0 (Windows NT 10.0; Win64; x64) \AppleWebKit/537.36 (KHTML, like Gecko) Chrome/80.0.3987.149 Safari/537.36'}url = 'https://attack.mitre.org/groups/'#向服务器发出请求r = requests.get(url = url, headers = headers).text#解析DOM树结构html_etree = etree.HTML(r)names = html_etree.xpath('//*[@class="table table-bordered table-alternate mt-2"]/tbody/tr/td[2]/a/text()')print (names)print(len(names),names[0])#链接urls = html_etree.xpath('//*[@class="table table-bordered table-alternate mt-2"]/tbody/tr/td[2]/a/@href')print(urls)print(len(urls), urls[0])print("\n")#-------------------------------------------------------------------------------------------#获取详细信息k = 0while k<len(names):filename = str(names[k]).strip() + ".txt"url = "https://attack.mitre.org" + urls[k]print(url)#获取正文信息page = urllib.request.Request(url, headers=headers)page = urllib.request.urlopen(page)contents = page.read()soup = BeautifulSoup(contents, "html.parser")#获取正文摘要信息content = ""for tag in soup.find_all(attrs={"class":"description-body"}):#contents = tag.find("p").get_text()contents = tag.find_all("p")for con in contents:content += con.get_text().strip() + "###\n"#标记句子结束(第二部分分句用)#print(content)#获取表格中的技术信息for tag in soup.find_all(attrs={"class":"table techniques-used table-bordered mt-2"}):contents = tag.find("tbody").find_all("tr")for con in contents:value = con.find("p").get_text() #存在4列或5列 故获取p值#print(value)content += value.strip() + "###\n" #标记句子结束(第二部分分句用)#删除内容中的参考文献括号 [n]result = re.sub(u"\\[.*" />, "", content)print(result)#文件写入filename = "Mitre//" + filenameprint(filename)f = open(filename, "w", encoding="utf-8")f.write(result)f.close()k += 1
输出结果如下图所示,共整理100个组织信息。
每个文件显示内容如下图所示:
数据标注采用暴力的方式进行,即定义不同类型的实体名称并利用BIO的方式进行标注。通过ATT&CK技战术方式进行标注,后续可以结合人工校正,同时可以定义更多类型的实体。
- BIO标注
实体名称 | 实体数量 | 示例 |
---|---|---|
APT攻击组织 | 128 | APT32、Lazarus Group |
攻击漏洞 | 56 | CVE-2009-0927 |
区域位置 | 72 | America、Europe |
攻击行业 | 34 | companies、finance |
攻击手法 | 65 | C&C、RAT、DDoS |
利用软件 | 48 | 7-Zip、Microsoft |
操作系统 | 10 | Linux、Windows |
更多标注和预处理请查看上一篇文章。
- [当人工智能遇上安全] 10.威胁情报实体识别之基于BiLSTM-CRF的实体识别万字详解
常见的数据标注工具:
- 图像标注:labelme,LabelImg,Labelbox,RectLabel,CVAT,VIA
- 半自动ocr标注:PPOCRLabel
- NLP标注工具:labelstudio
温馨提示:
由于网站的布局会不断变化和优化,因此读者需要掌握数据采集及语法树定位的基本方法,以不变应万变。此外,读者可以尝试采集所有锻炼甚至是URL跳转链接内容,请读者自行尝试和拓展!
二.数据预处理
假设存在已经采集和标注好的中文数据集,通常采用按字(Char)分隔,如下图所示,古籍为数据集,当然中文威胁情报也类似。
数据集划分为训练集和测试集。
接下来,我们需要读取CSV数据集,并构建汉字词典。关键函数:
- read_csv(filename):读取语料CSV文件
- count_vocab(words,labels):统计不重复词典
- build_vocab():构造词典
完整代码如下:
#encoding:utf-8# By: Eastmount WuShuai 2024-02-05import reimport osimport csvimport systrain_data_path = "data/train.csv"test_data_path = "data/test.csv"char_vocab_path = "char_vocabs.txt"#字典文件special_words = ['', ''] #特殊词表示final_words = [] #统计词典(不重复出现)final_labels = []#统计标记(不重复出现)#语料文件读取函数def read_csv(filename):words = []labels = []with open(filename,encoding='utf-8') as csvfile:reader = csv.reader(csvfile)for row in reader:if len(row)>0: #存在空行报错越界word,label = row[0],row[1]words.append(word)labels.append(label)return words,labels#统计不重复词典def count_vocab(words,labels):fp = open(char_vocab_path, 'a') #注意a为叠加(文件只能运行一次)k = 0while k<len(words):word = words[k]label = labels[k]if word not in final_words:final_words.append(word)fp.writelines(word + "\n")if label not in final_labels:final_labels.append(label)k += 1fp.close() #读取数据并构造原文字典(第一列)def build_vocab():words,labels = read_csv(train_data_path)print(len(words),len(labels),words[:8],labels[:8])count_vocab(words,labels)print(len(final_words),len(final_labels))#测试集words,labels = read_csv(test_data_path)print(len(words),len(labels))count_vocab(words,labels)print(len(final_words),len(final_labels))print(final_labels)#labels生成字典label_dict = {}k = 0for value in final_labels:label_dict[value] = kk += 1print(label_dict)return label_dictif __name__ == '__main__':build_vocab()
输出结果如下,包括训练集数量,并输出前8行文字及标注,以及不重复的汉字个数,以及实体类别14个。
['晉', '樂', '王', '鮒', '曰', ':', '', '小'] ['S-LOC', 'B-PER', 'I-PER', 'E-PER', 'O', 'O', '', 'O']xxx 14
输出类别如下。
['S-LOC', 'B-PER', 'I-PER', 'E-PER', 'O', '', 'B-LOC','E-LOC', 'S-PER', 'S-TIM', 'B-TIM', 'E-TIM', 'I-TIM', 'I-LOC']
接着实体类别进行编码处理,输出结果如下:
{'S-LOC': 0, 'B-PER': 1, 'I-PER': 2, 'E-PER': 3, 'O': 4, '': 5, 'B-LOC': 6,'E-LOC': 7, 'S-PER': 8, 'S-TIM': 9, 'B-TIM': 10, 'E-TIM': 11, 'I-TIM': 12, 'I-LOC': 13}
需要注意:在实体识别中,我们可以通过调用该函数获取识别的实体类别,关键代码如下。然而,由于真实分析中“O”通常建议编码为0,因此建议重新定义字典编码,更方便我们撰写代码,尤其是中文本遇到换句处理时,上述编码会乱序。
#原计划from get_data import build_vocab #调取第一阶段函数label2idx = build_vocab()#实际情况label2idx = {'O': 0, 'S-LOC': 1, 'B-LOC': 2,'I-LOC': 3,'E-LOC': 4, 'S-PER': 5, 'B-PER': 6,'I-PER': 7,'E-PER': 8, 'S-TIM': 9, 'B-TIM': 10, 'E-TIM': 11, 'I-TIM': 12 }....sent_ids = [vocab2idx[char] if char in vocab2idx else vocab2idx[''] for char in sent_]tag_ids = [label2idx[label] if label in label2idx else 0 for label in tag_]
最终生成词典char_vocabs.txt。
三.基于BiLSTM-CRF的实体识别
1.安装keras-contrib
CRF模型作者安装的是 keras-contrib
。
第一步,如果读者直接使用“pip install keras-contrib”可能会报错,远程下载也报错。
- pip install git+https://www.github.com/keras-team/keras-contrib.git
甚至会报错 ModuleNotFoundError: No module named ‘keras_contrib’。
第二步,作者从github中下载该资源,并在本地安装。
- https://github.com/keras-team/keras-contrib
- keras-contrib 版本:2.0.8
git clone https://www.github.com/keras-team/keras-contrib.gitcd keras-contribpython setup.py install
安装成功如下图所示:
读者可以从我的资源中下载代码和扩展包。
- https://github.com/eastmountyxz/When-AI-meet-Security
2.安装Keras
同样需要安装keras和TensorFlow扩展包。
如果TensorFlow下载太慢,可以设置清华大学镜像,实际安装2.2版本。
pip config set global.index-url https://pypi.tuna.tsinghua.edu.cn/simplepip install tensorflow==2.2
3.中文实体识别
第一步,数据预处理,包括BIO标记及词典转换。
#encoding:utf-8# By: Eastmount WuShuai 2024-02-05# 参考:https://github.com/huanghao128/zh-nlp-demoimport reimport osimport csvimport sysfrom get_data import build_vocab #调取第一阶段函数#------------------------------------------------------------------------#第一步 数据预处理#------------------------------------------------------------------------train_data_path = "data/train.csv"test_data_path = "data/test.csv"val_data_path = "data/val.csv"char_vocab_path = "char_vocabs.txt" #字典文件(防止多次写入仅读首次生成文件)special_words = ['', ''] #特殊词表示final_words = [] #统计词典(不重复出现)final_labels = []#统计标记(不重复出现)#BIO标记的标签 字母O初始标记为0#label2idx = build_vocab()label2idx = {'O': 0, 'S-LOC': 1, 'B-LOC': 2,'I-LOC': 3,'E-LOC': 4, 'S-PER': 5, 'B-PER': 6,'I-PER': 7,'E-PER': 8, 'S-TIM': 9, 'B-TIM': 10, 'E-TIM': 11, 'I-TIM': 12 }print(label2idx)#索引和BIO标签对应idx2label = {idx: label for label, idx in label2idx.items()}print(idx2label)#读取字符词典文件with open(char_vocab_path, "r") as fo:char_vocabs = [line.strip() for line in fo]char_vocabs = special_words + char_vocabsprint(char_vocabs)#字符和索引编号对应idx2vocab = {idx: char for idx, char in enumerate(char_vocabs)}vocab2idx = {char: idx for idx, char in idx2vocab.items()}print(idx2vocab)print(vocab2idx)
输出结果如下所示:
{'O': 0, 'S-LOC': 1, 'B-LOC': 2, 'I-LOC': 3, 'E-LOC': 4, 'S-PER': 5, 'B-PER': 6,'I-PER': 7, 'E-PER': 8, 'S-TIM': 9, 'B-TIM': 10, 'E-TIM': 11, 'I-TIM': 12}{0: 'O', 1: 'S-LOC', 2: 'B-LOC', 3: 'I-LOC', 4: 'E-LOC', 5: 'S-PER', 6: 'B-PER',7: 'I-PER', 8: 'E-PER', 9: 'S-TIM', 10: 'B-TIM', 11: 'E-TIM', 12: 'I-TIM'}['', '', '晉', '樂', '王', '鮒', '曰', ':', '', '小', '旻', ...]{0: '', 1: '', 2: '晉', 3: '樂', 4: '王', 5: '鮒', 6: '曰', 7: ':', 8: '', 9: '小', 10: '旻', ... ]{'': 0, '': 1, '晉': 2, '樂': 3, '王': 4, '鮒': 5, '曰': 6, ':': 7, '': 8, '小': 9, '旻': 10, ... ]
第二步,读取CSV数据,并获取汉字、标记对应的下标,以下标存储。
#------------------------------------------------------------------------#第二步 数据读取#------------------------------------------------------------------------def read_corpus(corpus_path, vocab2idx, label2idx):datas, labels = [], []with open(corpus_path, encoding='utf-8') as csvfile:reader = csv.reader(csvfile)sent_, tag_ = [], []for row in reader:word,label = row[0],row[1]if word!="" and label!="": #断句sent_.append(word)tag_.append(label)"""print(sent_) #['晉', '樂', '王', '鮒', '曰', ':']print(tag_)#['S-LOC', 'B-PER', 'I-PER', 'E-PER', 'O', 'O']"""else:#vocab2idx[0] => sent_ids = [vocab2idx[char] if char in vocab2idx else vocab2idx[''] for char in sent_]tag_ids = [label2idx[label] if label in label2idx else 0 for label in tag_]"""print(sent_ids,tag_ids)for idx,idy in zip(sent_ids,tag_ids):print(idx2vocab[idx],idx2label[idy])#[2, 3, 4, 5, 6, 7] [1, 6, 7, 8, 0, 0]#晉 S-LOC 樂 B-PER 王 I-PER 鮒 E-PER 曰 O : O"""datas.append(sent_ids) #按句插入列表labels.append(tag_ids)sent_, tag_ = [], []return datas, labels#原始数据train_datas_, train_labels_ = read_corpus(train_data_path, vocab2idx, label2idx)test_datas_, test_labels_ = read_corpus(test_data_path, vocab2idx, label2idx)#输出测试结果 (第五句语料)print(len(train_datas_),len(train_labels_),len(test_datas_),len(test_labels_))print(train_datas_[5])print([idx2vocab[idx] for idx in train_datas_[5]])print(train_labels_[5])print([idx2label[idx] for idx in train_labels_[5]])
输出结果如下,获取汉字和BIO标记的下标。
[2, 3, 4, 5, 6, 7] [1, 6, 7, 8, 0, 0]晉 S-LOC 樂 B-PER 王 I-PER 鮒 E-PER 曰 O : O
其中,第5行数据示例如下:
[46, 47, 48, 47, 49, 50, 51, 52, 53, 54, 55, 56]['齊', '、', '衛', '、', '陳', '大', '夫', '其', '不', '免', '乎', '!'][1, 0, 1, 0, 1, 0, 0, 0, 0, 0, 0, 0]['S-LOC', 'O', 'S-LOC', 'O', 'S-LOC', 'O', 'O', 'O', 'O', 'O', 'O', 'O']
对应语料如下:
第三步,数据填充和one-hot编码。
#------------------------------------------------------------------------#第三步 数据填充 one-hot编码#------------------------------------------------------------------------import kerasfrom keras.preprocessing import sequenceMAX_LEN = 100VOCAB_SIZE = len(vocab2idx)CLASS_NUMS = len(label2idx)#padding dataprint('padding sequences')train_datas = sequence.pad_sequences(train_datas_, maxlen=MAX_LEN)train_labels = sequence.pad_sequences(train_labels_, maxlen=MAX_LEN)test_datas = sequence.pad_sequences(test_datas_, maxlen=MAX_LEN)test_labels = sequence.pad_sequences(test_labels_, maxlen=MAX_LEN)print('x_train shape:', train_datas.shape)print('x_test shape:', test_datas.shape)#encoder one-hottrain_labels = keras.utils.to_categorical(train_labels, CLASS_NUMS)test_labels = keras.utils.to_categorical(test_labels, CLASS_NUMS)print('trainlabels shape:', train_labels.shape)print('testlabels shape:', test_labels.shape)
输出结果如下所示:
padding sequencesx_train shape: (xxx, 100)x_test shape: (xxx, 100)trainlabels shape: (xxx, 100, 13)testlabels shape: (xxx, 100, 13)
编码示例如下:
[ 00000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000 2163410294980 18]
第四步,构建BiLSTM+CRF模型。
#------------------------------------------------------------------------#第四步 构建BiLSTM+CRF模型# pip install git+https://www.github.com/keras-team/keras-contrib.git# 安装过程详见文件夹截图# ModuleNotFoundError: No module named ‘keras_contrib’#------------------------------------------------------------------------import numpy as npfrom keras.models import Sequentialfrom keras.models import Modelfrom keras.layers import Masking, Embedding, Bidirectional, LSTM, \ Dense, Input, TimeDistributed, Activationfrom keras_contrib.layers import CRFfrom keras_contrib.losses import crf_lossfrom keras_contrib.metrics import crf_viterbi_accuracyfrom keras import backend as Kfrom keras.models import load_modelfrom sklearn import metricsEPOCHS = 2EMBED_DIM = 128HIDDEN_SIZE = 64MAX_LEN = 100VOCAB_SIZE = len(vocab2idx)CLASS_NUMS = len(label2idx)K.clear_session()print(VOCAB_SIZE, CLASS_NUMS) #3319 13#模型构建 BiLSTM-CRFinputs = Input(shape=(MAX_LEN,), dtype='int32')x = Masking(mask_value=0)(inputs)x = Embedding(VOCAB_SIZE, EMBED_DIM, mask_zero=False)(x) #修改掩码Falsex = Bidirectional(LSTM(HIDDEN_SIZE, return_sequences=True))(x)x = TimeDistributed(Dense(CLASS_NUMS))(x)outputs = CRF(CLASS_NUMS)(x)model = Model(inputs=inputs, outputs=outputs)model.summary()
输出结果如下图所示,显示该模型的结构。
第五步,模型训练和测试。flag标记变量分别设置为“train”和“test”。
flag = "train"if flag=="train":#模型训练model.compile(loss=crf_loss, optimizer='adam', metrics=[crf_viterbi_accuracy])model.fit(train_datas, train_labels, epochs=EPOCHS, verbose=1, validation_split=0.1)score = model.evaluate(test_datas, test_labels, batch_size=256)print(model.metrics_names)print(score)model.save("bilstm_ner_model.h5")elif flag=="test":#训练模型char_vocab_path = "char_vocabs_.txt"#字典文件model_path = "bilstm_ner_model.h5"#模型文件ner_labels = label2idxspecial_words = ['', '']MAX_LEN = 100#预测结果model = load_model(model_path, custom_objects={'CRF': CRF}, compile=False)y_pred = model.predict(test_datas)y_labels = np.argmax(y_pred, axis=2) #取最大值z_labels = np.argmax(test_labels, axis=2)#真实值word_labels = test_datas #真实值k = 0final_y = [] #预测结果对应的标签final_z = [] #真实结果对应的标签final_word = []#对应的特征单词while k<len(y_labels):y = y_labels[k]for idx in y:final_y.append(idx2label[idx])#print("预测结果:", [idx2label[idx] for idx in y])z = z_labels[k]for idx in z:final_z.append(idx2label[idx])#print("真实结果:", [idx2label[idx] for idx in z])word = word_labels[k]for idx in word:final_word.append(idx2vocab[idx])k += 1print("最终结果大小:", len(final_y),len(final_z))n = 0numError = 0numRight = 0while n<len(final_y):if final_y[n]!=final_z[n] and final_z[n]!='O':numError += 1if final_y[n]==final_z[n] and final_z[n]!='O':numRight += 1n += 1print("预测错误数量:", numError)print("预测正确数量:", numRight)print("Acc:", numRight*1.0/(numError+numRight))print("预测单词:", [idx2vocab[idx] for idx in test_datas_[5]])print("真实结果:", [idx2label[idx] for idx in test_labels_[5]])print("预测结果:", [idx2label[idx] for idx in y_labels[5]][-len(test_datas_[5]):])
训练结果如下所示:
Epoch 1/232/8439 [..............................] - ETA: 6:51 - loss: 2.5549 - crf_viterbi_accuracy: 3.1250e-0464/8439 [..............................] - ETA: 3:45 - loss: 2.5242 - crf_viterbi_accuracy: 0.11428439/8439 [==============================] - 118s 14ms/step - loss: 0.1833 - crf_viterbi_accuracy: 0.9591 - val_loss: 0.0688 - val_crf_viterbi_accuracy: 0.9820Epoch 2/1032/8439 [..............................] - ETA: 19s - loss: 0.0644 - crf_viterbi_accuracy: 0.982564/8439 [..............................] - ETA: 42s - loss: 0.0592 - crf_viterbi_accuracy: 0.9845...['loss', 'crf_viterbi_accuracy'][0.043232945389307574, 0.9868513941764832]
最终测试结果如下所示,由于作者数据集仅放了少量数据,且未进行调参比较,真实数据更多且效果会更好。
预测错误数量: 2183预测正确数量: 2209Acc: 0.5029599271402551预测单词: ['冬', ',', '楚', '公', '子', '罷', '如', '晉', '聘', ',', '且', '涖', '盟', '。']真实结果: ['O', 'O', 'B-PER', 'I-PER', 'I-PER', 'E-PER', 'O', 'S-LOC', 'O', 'O', 'O', 'O', 'O', 'O']预测结果: ['O', 'O', 'B-PER', 'E-PER', 'E-PER', 'E-PER', 'O', 'S-LOC', 'O', 'O', 'O', 'O', 'O', 'O']
四.基于BiGRU-CRF的实体识别
接下来构建BiGRU-CRF代码,以完整代码为例,并将预测结果存储在CSV文件上。
#encoding:utf-8# By: Eastmount WuShuai 2024-02-05import reimport osimport csvimport sysfrom get_data import build_vocab #调取第一阶段函数#------------------------------------------------------------------------#第一步 数据预处理#------------------------------------------------------------------------train_data_path = "data/train.csv"test_data_path = "data/test.csv"val_data_path = "data/val.csv"char_vocab_path = "char_vocabs.txt"#字典文件(防止多次写入仅读首次生成文件)special_words = ['', ''] #特殊词表示final_words = [] #统计词典(不重复出现)final_labels = []#统计标记(不重复出现)#BIO标记的标签 字母O初始标记为0#label2idx = build_vocab()label2idx = {'O': 0, 'S-LOC': 1, 'B-LOC': 2,'I-LOC': 3,'E-LOC': 4, 'S-PER': 5, 'B-PER': 6,'I-PER': 7,'E-PER': 8, 'S-TIM': 9, 'B-TIM': 10, 'E-TIM': 11, 'I-TIM': 12 }#索引和BIO标签对应idx2label = {idx: label for label, idx in label2idx.items()}#读取字符词典文件with open(char_vocab_path, "r") as fo:char_vocabs = [line.strip() for line in fo]char_vocabs = special_words + char_vocabs#字符和索引编号对应idx2vocab = {idx: char for idx, char in enumerate(char_vocabs)}vocab2idx = {char: idx for idx, char in idx2vocab.items()}#------------------------------------------------------------------------#第二步 数据读取#------------------------------------------------------------------------def read_corpus(corpus_path, vocab2idx, label2idx):datas, labels = [], []with open(corpus_path, encoding='utf-8') as csvfile:reader = csv.reader(csvfile)sent_, tag_ = [], []for row in reader:word,label = row[0],row[1]if word!="" and label!="": #断句sent_.append(word)tag_.append(label)else:#vocab2idx[0] => sent_ids = [vocab2idx[char] if char in vocab2idx else vocab2idx[''] for char in sent_]tag_ids = [label2idx[label] if label in label2idx else 0 for label in tag_]datas.append(sent_ids) #按句插入列表labels.append(tag_ids)sent_, tag_ = [], []return datas, labels#原始数据train_datas_, train_labels_ = read_corpus(train_data_path, vocab2idx, label2idx)test_datas_, test_labels_ = read_corpus(test_data_path, vocab2idx, label2idx)#------------------------------------------------------------------------#第三步 数据填充 one-hot编码#------------------------------------------------------------------------import kerasfrom keras.preprocessing import sequenceMAX_LEN = 100VOCAB_SIZE = len(vocab2idx)CLASS_NUMS = len(label2idx)#padding dataprint('padding sequences')train_datas = sequence.pad_sequences(train_datas_, maxlen=MAX_LEN)train_labels = sequence.pad_sequences(train_labels_, maxlen=MAX_LEN)test_datas = sequence.pad_sequences(test_datas_, maxlen=MAX_LEN)test_labels = sequence.pad_sequences(test_labels_, maxlen=MAX_LEN)#encoder one-hottrain_labels = keras.utils.to_categorical(train_labels, CLASS_NUMS)test_labels = keras.utils.to_categorical(test_labels, CLASS_NUMS)#------------------------------------------------------------------------#第四步 构建BiGRU+CRF模型#------------------------------------------------------------------------import numpy as npfrom keras.models import Sequentialfrom keras.models import Modelfrom keras.layers import Masking, Embedding, Bidirectional, LSTM, GRU, \ Dense, Input, TimeDistributed, Activationfrom keras_contrib.layers import CRFfrom keras_contrib.losses import crf_lossfrom keras_contrib.metrics import crf_viterbi_accuracyfrom keras import backend as Kfrom keras.models import load_modelfrom sklearn import metricsEPOCHS = 2EMBED_DIM = 128HIDDEN_SIZE = 64MAX_LEN = 100VOCAB_SIZE = len(vocab2idx)CLASS_NUMS = len(label2idx)K.clear_session()print(VOCAB_SIZE, CLASS_NUMS)#模型构建 BiGRU-CRFinputs = Input(shape=(MAX_LEN,), dtype='int32')x = Masking(mask_value=0)(inputs)x = Embedding(VOCAB_SIZE, EMBED_DIM, mask_zero=False)(x) #修改掩码Falsex = Bidirectional(GRU(HIDDEN_SIZE, return_sequences=True))(x)x = TimeDistributed(Dense(CLASS_NUMS))(x)outputs = CRF(CLASS_NUMS)(x)model = Model(inputs=inputs, outputs=outputs)model.summary()flag = "test"if flag=="train":#模型训练model.compile(loss=crf_loss, optimizer='adam', metrics=[crf_viterbi_accuracy])model.fit(train_datas, train_labels, epochs=EPOCHS, verbose=1, validation_split=0.1)score = model.evaluate(test_datas, test_labels, batch_size=256)print(model.metrics_names)print(score)model.save("bigru_ner_model.h5")elif flag=="test":#训练模型char_vocab_path = "char_vocabs_.txt"#字典文件model_path = "bigru_ner_model.h5" #模型文件ner_labels = label2idxspecial_words = ['', '']MAX_LEN = 100#预测结果model = load_model(model_path, custom_objects={'CRF': CRF}, compile=False)y_pred = model.predict(test_datas)y_labels = np.argmax(y_pred, axis=2) #取最大值z_labels = np.argmax(test_labels, axis=2)#真实值word_labels = test_datas #真实值k = 0final_y = [] #预测结果对应的标签final_z = [] #真实结果对应的标签final_word = []#对应的特征单词while k<len(y_labels):y = y_labels[k]for idx in y:final_y.append(idx2label[idx])z = z_labels[k]for idx in z:final_z.append(idx2label[idx])word = word_labels[k]for idx in word:final_word.append(idx2vocab[idx])k += 1n = 0numError = 0numRight = 0while n<len(final_y):if final_y[n]!=final_z[n] and final_z[n]!='O':numError += 1if final_y[n]==final_z[n] and final_z[n]!='O':numRight += 1n += 1print("预测错误数量:", numError)print("预测正确数量:", numRight)print("Acc:", numRight*1.0/(numError+numRight))print("预测单词:", [idx2vocab[idx] for idx in test_datas_[5]])print("真实结果:", [idx2label[idx] for idx in test_labels_[5]])print("预测结果:", [idx2label[idx] for idx in y_labels[5]][-len(test_datas_[5]):])#文件存储fw = open("Final_BiGRU_CRF_Result.csv", "w", encoding="utf8", newline='')fwrite = csv.writer(fw)fwrite.writerow(['pre_label','real_label', 'word'])n = 0while n<len(final_y):fwrite.writerow([final_y[n],final_z[n],final_word[n]])n += 1fw.close()
输出结果如下所示:
['loss', 'crf_viterbi_accuracy'][0.03543611364953834, 0.9894005656242371]
生成文件如下图所示:
五.总结
写到这里这篇文章就结束,希望对您有所帮助,后续将结合经典的Bert进行分享。忙碌的2024,真的很忙,项目本子论文毕业工作,等忙完后好好写几篇安全博客,感谢支持和陪伴,尤其是家人的鼓励和支持, 继续加油!
- 一.ATT&CK数据采集
- 二.数据预处理
- 三.基于BiLSTM-CRF的实体识别
1.安装keras-contrib
2.安装Keras
3.中文实体识别 - 四.基于BiGRU-CRF的实体识别
- 五.总结
人生路是一个个十字路口,一次次博弈,一次次纠结和得失组成。得失得失,有得有失,不同的选择,不一样的精彩。虽然累和忙,但看到小珞珞还是挺满足的,感谢家人的陪伴。望小珞能开心健康成长,爱你们喔,继续干活,加油!
(By:Eastmount 2024-02-07 夜于贵阳 http://blog.csdn.net/eastmount/ )