基于Python深度学习实现主动学习的关系抽取方法

矫情吗；* 2024-02-25 03:24 16阅读 0赞

目 录  
第一章 引言 1  
1.1研究背景 1  
1.2国内外研究现状 2  
1.2.1关系抽取研究现状 2  
1.2.2主动学习研究现状 3  
1.3研究目标与内容 4  
1.4本文组织结构 4  
1.5本章小结 5  
第二章 相关工作 6  
2.1关系抽取 6  
2.2主动学习 9  
2.2.1主动学习算法模型 9  
2.2.2主流主动学习介绍 10  
2.3本章小结 11  
第三章 基于主动学习的关系抽取 12  
3.1抽取模型 12  
3.2基本采样方法 14  
3.2.1基于不确定性的采样方法 14  
3.2.2基于多样性的采样方法 15  
3.2.3基于代表性的采样方法 16  
3.3集成采样策略 16  
3.3.1基于多标准的赋权采样策略 17  
3.3.2基于多标准的逐层采样策略 18  
3.4本章小结 19  
第四章 实验及分析 20  
4.1实验环境与数据 20  
4.2实验评价指标 21  
4.3实验结果与分析 21  
4.4本章小结 25  
第五章 总结与展望 26  
致谢 27  
参考文献 27  
1.3研究目标与内容  
本工作的主要目标是设计一种基于主动学习的关系抽取方法，得到一个具有高普适性和可迁移性的样本查询策略，在不同抽取模型、不同数据集上与随机采样对比，验证其有效性。  
本文的主要研究内容包括：

1.实现多种主动学习算法，包括基于不确定性、代表性、多样性的基本采样方法，验证主动学习算法在关系抽取任务中的有效性。  
2.在基本采样方法的基础上，设计融合多种标准的采样策略，最后在关系抽取任务上进行对比实验，得到最优采样策略。  
3.实现多种关系抽取模型，包括 CNN、BLSTM、R-BERT 等，验证设计的主动学习算法具有普适性。  
1.4本文组织结构  
针对当前研究现状存在的问题以及文本的研究目标、研究内容，拟定本文的基本结构如下：  
一、引言。本章首先指出研究基于主动学习的关系抽取方法的背景和意义，概括性地

介绍了国内外当前关系抽取和主动学习的研究现状，在此基础上，确定了本文研究目标和内容，最后给出全文的组织结构。  
二、相关工作。介绍关系抽取的任务中常用的分类模型和基本算法，以及描述主动学习的算法流程、分类情况。  
三、基于主动学习的关系抽取。具体介绍本文提出的基于主动学习的关系抽取方法， 在分析了基于不确定性、多样性、代表性三类基本的样本采样策略的基础上，提出多标准融合采样策略，进一步提高主动学习的提升性能。  
四、实验及分析。验证提出的基于主动学习的关系抽取方法的有效性、健壮性、可迁移性。

import hydra
    import torch
    import logging
    import random
    import torch.nn as nn
    from torch import optim
    from hydra import utils
    from torch.utils.data import DataLoader
    from torch.utils.tensorboard import SummaryWriter
    from numpy import *
    import models
    import data_select
    
    from utils.preprocess import preprocess
    from utils.dataset import CustomDataset, collate_fn
    from trainer import train, validate
    from utils.util import manual_seed, load_pkl
    
    logger = logging.getLogger(__name__)
    
    
    @hydra.main(config_path='conf/config.yaml')
    def main(cfg):
        cwd = utils.get_original_cwd()
        cfg.cwd = cwd
        cfg.lm_file = cwd + '/pretrained/data'
        cfg.pos_size = 2 * cfg.pos_limit + 2
        logger.info(f'\n{
        cfg.pretty()}')
    
        __Model__ = {
        
            'cnn': models.PCNN,
            'rnn': models.BiLSTM,
            'transformer': models.Transformer,
            'gcn': models.GCN,
            'capsule': models.Capsule,
            'lm': models.LM,
        }
    
        __Select__ = {
        
            'random': data_select.random_sample,
            'uncertainty': data_select.uncertainty_sample,
            'diversity': data_select.diversity_sample,
            'iter_diff': data_select.iter_diff_sample,
            'similarity': data_select.similarity_sample,
            'ensemble_sample_by_weight': data_select.ensemble_sample_by_weight,
            'uncertain_diverse': data_select.ensemble_sample_by_uncertain_similar,
            'uncertain_similar': data_select.ensemble_sample_by_uncertain_diverse,
            'multi_criterion_sampling': data_select.multi_criterion_sampling
        }
    
        # device
        if cfg.use_gpu and torch.cuda.is_available():
            device = torch.device('cuda', cfg.gpu_id)
        else:
            device = torch.device('cpu')
    
        # 如果不修改预处理的过程，这一步最好注释掉，不用每次运行都预处理数据一次
        if cfg.preprocess:
            preprocess(cfg)
    
        train_data_path = os.path.join(cfg.cwd, cfg.out_path, 'train.pkl')
        valid_data_path = os.path.join(cfg.cwd, cfg.out_path, 'valid.pkl')
        test_data_path = os.path.join(cfg.cwd, cfg.out_path, 'test.pkl')
        vocab_path = os.path.join(cfg.cwd, cfg.out_path, 'vocab.pkl')
    
        if cfg.model_name == 'lm':
            vocab_size = None
        else:
            vocab = load_pkl(vocab_path)
            vocab_size = vocab.count
        cfg.vocab_size = vocab_size
    
        valid_ds = CustomDataset(valid_data_path)
        test_ds = CustomDataset(test_data_path)
        valid_dataloader = DataLoader(valid_ds, batch_size=cfg.batch_size, shuffle=True, collate_fn=collate_fn(cfg))
        test_dataloader = DataLoader(test_ds, batch_size=cfg.batch_size, shuffle=True, collate_fn=collate_fn(cfg))
    
        all_train_ds = load_pkl(train_data_path)
        random.shuffle(all_train_ds)
    
        start_size, all_size, size, per_log_num = 4000, 9800, 200, 400
        select_method = __Select__[cfg.select_method]
        pre_labels = None
        print(len(all_train_ds))
        cur_labeled_ds = all_train_ds[:start_size]
        unlabeled_ds = all_train_ds[start_size:]
    
        writer = SummaryWriter('tensorboard')
    
        logger.info('=' * 10 + ' Start training ' + '=' * 10)
        test_f1_scores, test_losses = [], []
        while len(cur_labeled_ds) <= all_size:
            model = __Model__[cfg.model_name](cfg)
            if len(cur_labeled_ds) == start_size:
                logger.info(f'\n {
        model}')
            model.to(device)
            optimizer = optim.Adam(model.parameters(), lr=cfg.learning_rate, weight_decay=cfg.weight_decay)
            scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, factor=cfg.lr_factor, patience=cfg.lr_patience)
            criterion = nn.CrossEntropyLoss()
    
            train_dataloader = DataLoader(cur_labeled_ds, batch_size=cfg.batch_size, shuffle=True,
                                          collate_fn=collate_fn(cfg))
            train_losses, valid_losses, one_f1_scores = [], [], []
            for epoch in range(1, cfg.epoch + 1):
                # 保证随机种子每一轮不一样
                manual_seed(cfg.seed + epoch)
                train_loss = train(epoch, model, train_dataloader, optimizer, criterion, device, writer, cfg)
                # 验证集
                valid_f1, valid_loss = validate(epoch, model, valid_dataloader, criterion, device, cfg)
                # 根据valid_loss来调整学习率
                scheduler.step(valid_loss)
                # model_path = model.save(epoch, cfg)
    
                train_losses.append(train_loss)
                valid_losses.append(valid_loss)
                one_f1_scores.append(valid_f1)
    
            if cfg.show_plot and cfg.plot_utils == 'tensorboard' and (len(cur_labeled_ds) - start_size) % per_log_num == 0:
                # logger.info(f'one_f1_scores:{one_f1_scores}')
                for i in range(len(train_losses)):
                    writer.add_scalars(f'valid_copy/valid_loss_{
        len(cur_labeled_ds)}', {
        
                        'train': train_losses[i],
                        'valid': valid_losses[i]
                    }, i)
                    writer.add_scalars(f'valid/valid_f1_score_{
        len(cur_labeled_ds)}', {
        
                        'valid_f1_score': one_f1_scores[i]
                    }, i)
    
            test_f1, test_loss = validate(-1, model, test_dataloader, criterion, device, cfg)
            test_f1_scores.append(test_f1)
            test_losses.append(test_loss)
            # 最后5次迭代的平均值作为当前样本数下的f1_score表现
            # f1_scores.append(mean(one_f1_scores[-5:]))
            if len(cur_labeled_ds) == all_size:
                break
    
            cur_labeled_ds, unlabeled_ds = select_method(cur_labeled_ds, unlabeled_ds, size, model, device, cfg)
    
        if cfg.show_plot and cfg.plot_utils == 'tensorboard':
            for j in range(len(test_f1_scores)):
                writer.add_scalars('test/test_losses', {
        
                    'test_losses': test_losses[j]
                }, j)
                writer.add_scalars('test/test_f1_scores', {
        
                    'test_f1_scores': test_f1_scores[j],
                }, j)
            writer.close()
    
        # 测试集
        validate(-1, model, test_dataloader, criterion, device, cfg)
    
    
    if __name__ == '__main__':
        import time
    
        # 查看显卡使用情况：nvidia-smi
        cur = time.time()
        main()
        logger.info(f'run time:{
        time.time() - cur}')

![在这里插入图片描述][6a8800cf023440c2844640c3068bcc3d.png_pic_center]  
![在这里插入图片描述][63a19e8d8516479c82f7d5b31bca0834.png_pic_center]  
![在这里插入图片描述][228bd9cfef89433f8ebbdf677a8ae373.png_pic_center]  
![在这里插入图片描述][be8f86996f9e4756bdbd0cc4e3e5e234.png_pic_center]  
![在这里插入图片描述][d983cf7866584fcf8e8076c096ba7e40.png_pic_center]  
![在这里插入图片描述][ee594a3de169494a8cda129a516067e7.png_pic_center]  
![在这里插入图片描述][8575e3b3e87b44feb23c39861c512ba3.png_pic_center]  
![在这里插入图片描述][2e1b85a6e1b3480ba48b3eb9bd8cc966.png_pic_center]  
![在这里插入图片描述][4b7024eaa93b48eea844514416a33472.png_pic_center]  
![在这里插入图片描述][009ed1bf8472463db42b9c7e32d50f6f.png_pic_center]  
![在这里插入图片描述][23d69de539b94769aee49311aee509eb.png_pic_center]  
![在这里插入图片描述][f63881199510456fa04455f9ec9bb0dc.png_pic_center]  
![在这里插入图片描述][08bf1ae03cbe4619988d784ed8a84779.png_pic_center]  
![在这里插入图片描述][209fcf074db14a719138cd0c2d5474c7.png_pic_center]  
![在这里插入图片描述][9b503537b9764f6aa9d37b2ff0a97619.png_pic_center]  
![在这里插入图片描述][d44184ff8c2140c29a078d5b8e183baf.png_pic_center]  
![在这里插入图片描述][bdbab1cddd5f4bf3865696993025ec39.png_pic_center]  
![在这里插入图片描述][1f6d0afc81f34141b031ca9b8e045aa9.png_pic_center]  
![在这里插入图片描述][1e45d555875a49ff840d2332b768f893.png_pic_center]  
![在这里插入图片描述][0fdba746a1504134bef8764846b6e519.png_pic_center]  
![在这里插入图片描述][434cc884cd0b48fbb283395436fe2fad.png_pic_center]  
![在这里插入图片描述][9f2e93fdf95b4a13a61556db598cf9f5.png_pic_center]  
![在这里插入图片描述][075bec559d44473a98d6053bf1a92076.png_pic_center]  
![在这里插入图片描述][45207d6a6ace452eb4a884edecba460f.png_pic_center]  
![在这里插入图片描述][e80b2f1e1cd6412eaf601c31c3871c0f.png_pic_center]  
![在这里插入图片描述][751a0c8f0d9c4902b52c9923a42084a4.png_pic_center]  
![在这里插入图片描述][72a081ab74a7445383d7fcf287101489.png_pic_center]

[6a8800cf023440c2844640c3068bcc3d.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/02/25/9cd1b57f76204bedae310c6f4ab2dec7.png
[63a19e8d8516479c82f7d5b31bca0834.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/02/25/d2756309420543c792ad2301b9444444.png
[228bd9cfef89433f8ebbdf677a8ae373.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/02/25/258ca33bd8f3474e8236f01355095beb.png
[be8f86996f9e4756bdbd0cc4e3e5e234.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/02/25/e0d660ff889c4b13827a08144e62faec.png
[d983cf7866584fcf8e8076c096ba7e40.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/02/25/f181476385a84c7798762122b2dd5a42.png
[ee594a3de169494a8cda129a516067e7.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/02/25/9d6c6cfbef984acc93dc1e791e85334f.png
[8575e3b3e87b44feb23c39861c512ba3.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/02/25/9b19346548bc48cf9ea3275fad37e65c.png
[2e1b85a6e1b3480ba48b3eb9bd8cc966.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/02/25/90722659000942da805c6a98a800e084.png
[4b7024eaa93b48eea844514416a33472.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/02/25/ffd682f77240426faeb3b6ae02d1143a.png
[009ed1bf8472463db42b9c7e32d50f6f.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/02/25/7dc12a715ddb42b49ad891ae4d15933a.png
[23d69de539b94769aee49311aee509eb.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/02/25/25749f51aca14df2ba04366ef481eda4.png
[f63881199510456fa04455f9ec9bb0dc.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/02/25/01ca9cdaed7c4f6ea24a96493ef5ce38.png
[08bf1ae03cbe4619988d784ed8a84779.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/02/25/0997059b4da64fad88749946127e635d.png
[209fcf074db14a719138cd0c2d5474c7.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/02/25/90ee0c60504644569e4af216ba121860.png
[9b503537b9764f6aa9d37b2ff0a97619.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/02/25/ef25e6ad0a20474f847b87cf1d2ce994.png
[d44184ff8c2140c29a078d5b8e183baf.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/02/25/65d1bc7c8ea342a7b820db214497fd79.png
[bdbab1cddd5f4bf3865696993025ec39.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/02/25/a2d8824e32274e2aabe9c1d44af4d3e9.png
[1f6d0afc81f34141b031ca9b8e045aa9.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/02/25/49708d1a220d47f9a9775df08bf910f2.png
[1e45d555875a49ff840d2332b768f893.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/02/25/1c80d6c95a6d4c2b93b299066781c7eb.png
[0fdba746a1504134bef8764846b6e519.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/02/25/5d711f2bfedc469f98aa515b28080235.png
[434cc884cd0b48fbb283395436fe2fad.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/02/25/199c702df0584a62bcd6536d4a38d61e.png
[9f2e93fdf95b4a13a61556db598cf9f5.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/02/25/2452e665e5cf412488020f6886c303a1.png
[075bec559d44473a98d6053bf1a92076.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/02/25/6eed11eb68d341d78b81b2ce82907c71.png
[45207d6a6ace452eb4a884edecba460f.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/02/25/daa1273fc9f54778a980bd5925ebb23b.png
[e80b2f1e1cd6412eaf601c31c3871c0f.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/02/25/ed678b16e83041229c0d787cb193dff9.png
[751a0c8f0d9c4902b52c9923a42084a4.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/02/25/df5009ad870549209788cb4bb178d672.png
[72a081ab74a7445383d7fcf287101489.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/02/25/800a4281af304cc9b1e81536c318d777.png