利用预训练模型skep进行情感分析的方法是_模型训练平台

利用预训练模型skep进行情感分析的方法是_模型训练平台本项目将详细全面介绍情感分析任务的两种子任务,句子级情感分析和目标级情感分析

 项目地址:文本情感分析 – 飞桨AI Studio星河社区 (baidu.com)

 baidu/Senta: Baidu’s open-source Sentiment Analysis System. (github.com) 

本项目将详细全面介绍情感分析任务的两种子任务,句子级情感分析和目标级情感分析。

同时演示如何使用情感分析预训练模型SKEP完成以上两种任务,详细介绍预训练模型SKEP及其在 PaddleNLP 的使用方式。

本项目主要包括“任务介绍”、“情感分析预训练模型SKEP”、“句子级情感分析”、“目标级情感分析”等四个部分。

!pip install --upgrade paddlenlp -i https://pypi.org/simple 

1、情感分析任务

众所周知,人类自然语言中包含了丰富的情感色彩:表达人的情绪(如悲伤、快乐)、表达人的心情(如倦怠、忧郁)、表达人的喜好(如喜欢、讨厌)、表达人的个性特征和表达人的立场等等。情感分析在商品喜好、消费决策、舆情分析等场景中均有应用。利用机器自动分析这些情感倾向,不但有助于帮助企业了解消费者对其产品的感受,为产品改进提供依据;同时还有助于企业分析商业伙伴们的态度,以便更好地进行商业决策。

被人们所熟知的情感分析任务是将一段文本分类,如分为情感极性为正向负向其他的三分类问题:

利用预训练模型skep进行情感分析的方法是_模型训练平台

  • 正向: 表示正面积极的情感,如高兴,幸福,惊喜,期待等。
  • 负向: 表示负面消极的情感,如难过,伤心,愤怒,惊恐等。
  • 其他: 其他类型的情感。

实际上,以上熟悉的情感分析任务是句子级情感分析任务

情感分析任务还可以进一步分为句子级情感分析目标级情感分析等任务。在下面章节将会详细介绍两种任务及其应用场景。

2、情感分析预训练模型SKEP

近年来,大量的研究表明基于大型语料库的预训练模型(Pretrained Models, PTM)可以学习通用的语言表示,有利于下游NLP任务,同时能够避免从零开始训练模型。随着计算能力的发展,深度模型的出现(即 Transformer)和训练技巧的增强使得 PTM 不断发展,由浅变深。

情感预训练模型SKEP(Sentiment Knowledge Enhanced Pre-training for Sentiment Analysis)。SKEP利用情感知识增强预训练模型, 在14项中英情感分析典型任务上全面超越SOTA,此工作已经被ACL 2020录用。SKEP是百度研究团队提出的基于情感知识增强的情感预训练算法,此算法采用无监督方法自动挖掘情感知识,然后利用情感知识构建预训练目标,从而让机器学会理解情感语义。SKEP为各类情感分析任务提供统一且强大的情感语义表示。

论文地址:[2005.05635] SKEP: Sentiment Knowledge Enhanced Pre-training for Sentiment Analysis (arxiv.org)

利用预训练模型skep进行情感分析的方法是_模型训练平台

百度研究团队在三个典型情感分析任务,句子级情感分类(Sentence-level Sentiment Classification),评价目标级情感分类(Aspect-level Sentiment Classification)、观点抽取(Opinion Role Labeling),共计14个中英文数据上进一步验证了情感预训练模型SKEP的效果。

 具体实验效果参考:baidu/Senta: Baidu’s open-source Sentiment Analysis System. (github.com) 

3、句子级情感分析 & 目标级情感分析

3.1 句子级情感分析

对给定的一段文本进行情感极性分类,常用于影评分析、网络论坛舆情分析等场景。如:

选择珠江花园的原因就是方便,有电动扶梯直接到达海边,周围餐馆、食廊、商场、超市、摊位一应俱全。酒店装修一般,但还算整洁。 泳池在大堂的屋顶,因此很小,不过女儿倒是喜欢。 包的早餐是西式的,还算丰富。 服务吗,一般    1
15.4寸笔记本的键盘确实爽,基本跟台式机差不多了,蛮喜欢数字小键盘,输数字特方便,样子也很美观,做工也相当不错    1
房间太小。其他的都一般… … … …    0

其中1表示正向情感,0表示负向情感。

利用预训练模型skep进行情感分析的方法是_模型训练平台

句子级情感分析任务

3.1.1 常用数据集

ChnSenticorp数据集是公开中文情感分析常用数据集, 其为2分类数据集。PaddleNLP已经内置该数据集,一键即可加载。

from paddlenlp.datasets import load_dataset

train_ds, dev_ds, test_ds = load_dataset("chnsenticorp", splits=["train", "dev", "test"])

print(train_ds[0])
print(train_ds[1])
print(train_ds[2])

{‘text’: ‘选择珠江花园的原因就是方便,有电动扶梯直接到达海边,周围餐馆、食廊、商场、超市、摊位一应俱全。酒店装修一般,但还算整洁。 泳池在大堂的屋顶,因此很小,不过女儿倒是喜欢。 包的早餐是西式的,还算丰富。 服务吗,一般’, ‘label’: 1, ‘qid’: ”}
{‘text’: ‘15.4寸笔记本的键盘确实爽,基本跟台式机差不多了,蛮喜欢数字小键盘,输数字特方便,样子也很美观,做工也相当不错’, ‘label’: 1, ‘qid’: ”}
{‘text’: ‘房间太小。其他的都一般。。。。。。。。。’, ‘label’: 0, ‘qid’: ”}

3.1.2 SKEP模型加载

PaddleNLP已经实现了SKEP预训练模型,可以通过一行代码实现SKEP加载。

句子级情感分析模型是SKEP fine-tune 文本分类常用模型SkepForSequenceClassification。其首先通过SKEP提取句子语义特征,之后将语义特征进行分类。

利用预训练模型skep进行情感分析的方法是_模型训练平台

from paddlenlp.transformers import SkepForSequenceClassification, SkepTokenizer

# 指定模型名称,一键加载模型
model = SkepForSequenceClassification.from_pretrained(pretrained_model_name_or_path="skep_ernie_1.0_large_ch", num_classes=len(train_ds.label_list))
# 同样地,通过指定模型名称一键加载对应的Tokenizer,用于处理文本数据,如切分token,转token_id等。
tokenizer = SkepTokenizer.from_pretrained(pretrained_model_name_or_path="skep_ernie_1.0_large_ch")

kepForSequenceClassification可用于句子级情感分析和目标级情感分析任务。其通过预训练模型SKEP获取输入文本的表示,之后将文本表示进行分类。

  • pretrained_model_name_or_path:模型名称。支持”skep_ernie_1.0_large_ch”,“skep_ernie_2.0_large_en”。

    • “skep_ernie_1.0_large_ch”:是SKEP模型在预训练ernie_1.0_large_ch基础之上在海量中文数据上继续预训练得到的中文预训练模型;
    • “skep_ernie_2.0_large_en”:是SKEP模型在预训练ernie_2.0_large_en基础之上在海量英文数据上继续预训练得到的英文预训练模型;
  • num_classes: 数据集分类类别数。

关于SKEP模型实现详细信息参考:PaddleNLP/paddlenlp/transformers/skep at develop · PaddlePaddle/PaddleNLP (github.com)

3.1.3 数据处理

同样地,我们需要将原始ChnSentiCorp数据处理成模型可以读入的数据格式。

SKEP模型对中文文本处理按照字粒度进行处理,我们可以使用PaddleNLP内置的SkepTokenizer完成一键式处理。

import os
from functools import partial

import numpy as np
import paddle
import paddle.nn.functional as F
from paddlenlp.data import Stack, Tuple, Pad

from utils import create_dataloader

def convert_example(example,
                    tokenizer,
                    max_seq_length=512,
                    is_test=False):
    """
    Builds model inputs from a sequence or a pair of sequence for sequence classification tasks
    by concatenating and adding special tokens. And creates a mask from the two sequences passed 
    to be used in a sequence-pair classification task.
        
    A skep_ernie_1.0_large_ch/skep_ernie_2.0_large_en sequence has the following format:
    ::
        - single sequence: ``[CLS] X [SEP]``
        - pair of sequences: ``[CLS] A [SEP] B [SEP]``

    A skep_ernie_1.0_large_ch/skep_ernie_2.0_large_en sequence pair mask has the following format:
    ::

        0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1
        | first sequence    | second sequence |

    If `token_ids_1` is `None`, this method only returns the first portion of the mask (0s).


    Args:
        example(obj:`list[str]`): List of input data, containing text and label if it have label.
        tokenizer(obj:`PretrainedTokenizer`): This tokenizer inherits from :class:`~paddlenlp.transformers.PretrainedTokenizer` 
            which contains most of the methods. Users should refer to the superclass for more information regarding methods.
        max_seq_len(obj:`int`): The maximum total input sequence length after tokenization. 
            Sequences longer than this will be truncated, sequences shorter will be padded.
        is_test(obj:`False`, defaults to `False`): Whether the example contains label or not.

    Returns:
        input_ids(obj:`list[int]`): The list of token ids.
        token_type_ids(obj: `list[int]`): List of sequence pair mask.
        label(obj:`int`, optional): The input label if not is_test.
    """
    # 将原数据处理成model可读入的格式,enocded_inputs是一个dict,包含input_ids、token_type_ids等字段
    encoded_inputs = tokenizer(
        text=example["text"], max_seq_len=max_seq_length)

    # input_ids:对文本切分token后,在词汇表中对应的token id
    input_ids = encoded_inputs["input_ids"]
    # token_type_ids:当前token属于句子1还是句子2,即上述图中表达的segment ids
    token_type_ids = encoded_inputs["token_type_ids"]

    if not is_test:
        # label:情感极性类别
        label = np.array([example["label"]], dtype="int64")
        return input_ids, token_type_ids, label
    else:
        # qid:每条数据的编号
        qid = np.array([example["qid"]], dtype="int64")
        return input_ids, token_type_ids, qid
# 批量数据大小
batch_size = 32
# 文本序列最大长度
max_seq_length = 128

# 将数据处理成模型可读入的数据格式
trans_func = partial(
    convert_example,
    tokenizer=tokenizer,
    max_seq_length=max_seq_length)

# 将数据组成批量式数据,如
# 将不同长度的文本序列padding到批量式数据中最大长度
# 将每条数据label堆叠在一起
batchify_fn = lambda samples, fn=Tuple(
    Pad(axis=0, pad_val=tokenizer.pad_token_id),  # input_ids
    Pad(axis=0, pad_val=tokenizer.pad_token_type_id),  # token_type_ids
    Stack()  # labels
): [data for data in fn(samples)]
train_data_loader = create_dataloader(
    train_ds,
    mode='train',
    batch_size=batch_size,
    batchify_fn=batchify_fn,
    trans_fn=trans_func)
dev_data_loader = create_dataloader(
    dev_ds,
    mode='dev',
    batch_size=batch_size,
    batchify_fn=batchify_fn,
    trans_fn=trans_func)
3.1.4 模型训练和评估

定义损失函数、优化器以及评价指标后,即可开始训练。

推荐超参设置:

  • max_seq_length=256
  • batch_size=48
  • learning_rate=2e-5
  • epochs=10

实际运行时可以根据显存大小调整batch_size和max_seq_length大小。

utils.py文件如下(放在项目同级目录中)

# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import numpy as np
import paddle


def create_dataloader(dataset,
                      trans_fn=None,
                      mode='train',
                      batch_size=1,
                      batchify_fn=None):
    """
    Creats dataloader.

    Args:
        dataset(obj:`paddle.io.Dataset`): Dataset instance.
        trans_fn(obj:`callable`, optional, defaults to `None`): function to convert a data sample to input ids, etc.
        mode(obj:`str`, optional, defaults to obj:`train`): If mode is 'train', it will shuffle the dataset randomly.
        batch_size(obj:`int`, optional, defaults to 1): The sample number of a mini-batch.
        batchify_fn(obj:`callable`, optional, defaults to `None`): function to generate mini-batch data by merging
            the sample list, None for only stack each fields of sample in axis
            0(same as :attr::`np.stack(..., axis=0)`).

    Returns:
        dataloader(obj:`paddle.io.DataLoader`): The dataloader which generates batches.
    """
    if trans_fn:
        dataset = dataset.map(trans_fn)

    shuffle = True if mode == 'train' else False
    if mode == "train":
        sampler = paddle.io.DistributedBatchSampler(
            dataset=dataset, batch_size=batch_size, shuffle=shuffle)
    else:
        sampler = paddle.io.BatchSampler(
            dataset=dataset, batch_size=batch_size, shuffle=shuffle)
    dataloader = paddle.io.DataLoader(
        dataset, batch_sampler=sampler, collate_fn=batchify_fn)
    return dataloader


def convert_example(example, tokenizer, is_test=False):
    """
    Builds model inputs from a sequence for sequence classification tasks. 
    It use `jieba.cut` to tokenize text.

    Args:
        example(obj:`list[str]`): List of input data, containing text and label if it have label.
        tokenizer(obj: paddlenlp.data.JiebaTokenizer): It use jieba to cut the chinese string.
        is_test(obj:`False`, defaults to `False`): Whether the example contains label or not.

    Returns:
        input_ids(obj:`list[int]`): The list of token ids.
        valid_length(obj:`int`): The input sequence valid length.
        label(obj:`numpy.array`, data type1 of int64, optional): The input label if not is_test.
    """

    input_ids = tokenizer.encode(example["text"])
    input_ids = np.array(input_ids, dtype='int64')

    if not is_test:
        label = np.array(example["label"], dtype="int64")
        return input_ids, label
    else:
        return input_ids


@paddle.no_grad()
def evaluate(model, criterion, metric, data_loader):
    """
    Given a dataset, it evals model and computes the metric.

    Args:
        model(obj:`paddle.nn.Layer`): A model to classify texts.
        criterion(obj:`paddle.nn.Layer`): It can compute the loss.
        metric(obj:`paddle.metric.Metric`): The evaluation metric.
        data_loader(obj:`paddle.io.DataLoader`): The dataset loader which generates batches.
    """
    model.eval()
    metric.reset()
    losses = []
    for batch in data_loader:
        input_ids, token_type_ids, labels = batch
        logits = model(input_ids, token_type_ids)
        loss = criterion(logits, labels)
        losses.append(loss.numpy())
        correct = metric.compute(logits, labels)
        metric.update(correct)
        accu = metric.accumulate()
    print("eval loss: %.5f, accu: %.5f" % (np.mean(losses), accu))
    model.train()
    metric.reset()

import time

from utils import evaluate

# 训练轮次
epochs = 1
# 训练过程中保存模型参数的文件夹
ckpt_dir = "skep_ckpt"
# len(train_data_loader)一轮训练所需要的step数
num_training_steps = len(train_data_loader) * epochs

# Adam优化器
optimizer = paddle.optimizer.AdamW(
    learning_rate=2e-5,
    parameters=model.parameters())
# 交叉熵损失函数
criterion = paddle.nn.loss.CrossEntropyLoss()
# accuracy评价指标
metric = paddle.metric.Accuracy()
# 开启训练
global_step = 0
tic_train = time.time()
for epoch in range(1, epochs + 1):
    for step, batch in enumerate(train_data_loader, start=1):
        input_ids, token_type_ids, labels = batch
        # 喂数据给model
        logits = model(input_ids, token_type_ids)
        # 计算损失函数值
        loss = criterion(logits, labels)
        # 预测分类概率值
        probs = F.softmax(logits, axis=1)
        # 计算acc
        correct = metric.compute(probs, labels)
        metric.update(correct)
        acc = metric.accumulate()

        global_step += 1
        if global_step % 10 == 0:
            print(
                "global step %d, epoch: %d, batch: %d, loss: %.5f, accu: %.5f, speed: %.2f step/s"
                % (global_step, epoch, step, loss, acc,
                    10 / (time.time() - tic_train)))
            tic_train = time.time()
        
        # 反向梯度回传,更新参数
        loss.backward()
        optimizer.step()
        optimizer.clear_grad()

        if global_step % 100 == 0:
            save_dir = os.path.join(ckpt_dir, "model_%d" % global_step)
            if not os.path.exists(save_dir):
                os.makedirs(save_dir)
            # 评估当前训练的模型
            evaluate(model, criterion, metric, dev_data_loader)
            # 保存当前模型参数等
            model.save_pretrained(save_dir)
            # 保存tokenizer的词表等
            tokenizer.save_pretrained(save_dir)

global step 10, epoch: 1, batch: 10, loss: 0.53868, accu: 0.66250, speed: 1.45 step/s
global step 20, epoch: 1, batch: 20, loss: 0.38239, accu: 0.76562, speed: 1.40 step/s
global step 30, epoch: 1, batch: 30, loss: 0.14145, accu: 0.81667, speed: 1.39 step/s
global step 40, epoch: 1, batch: 40, loss: 0.19523, accu: 0.84219, speed: 1.40 step/s
global step 50, epoch: 1, batch: 50, loss: 0.17806, accu: 0.85688, speed: 1.40 step/s
global step 60, epoch: 1, batch: 60, loss: 0.34572, accu: 0.86771, speed: 1.40 step/s
global step 70, epoch: 1, batch: 70, loss: 0.28901, accu: 0.87634, speed: 1.40 step/s
global step 80, epoch: 1, batch: 80, loss: 0.30491, accu: 0.87891, speed: 1.40 step/s
global step 90, epoch: 1, batch: 90, loss: 0.21844, accu: 0.88403, speed: 1.40 step/s
global step 100, epoch: 1, batch: 100, loss: 0.08482, accu: 0.88687, speed: 1.40 step/s
eval loss: 0.24119, accu: 0.91083
global step 110, epoch: 1, batch: 110, loss: 0.23338, accu: 0.89375, speed: 0.49 step/s
global step 120, epoch: 1, batch: 120, loss: 0.11810, accu: 0.89375, speed: 1.40 step/s
global step 130, epoch: 1, batch: 130, loss: 0.15867, accu: 0.90208, speed: 1.40 step/s
global step 140, epoch: 1, batch: 140, loss: 0.09246, accu: 0.90391, speed: 1.40 step/s
global step 150, epoch: 1, batch: 150, loss: 0.17813, accu: 0.90750, speed: 1.40 step/s
global step 160, epoch: 1, batch: 160, loss: 0.30430, accu: 0.90885, speed: 1.41 step/s
global step 170, epoch: 1, batch: 170, loss: 0.09656, accu: 0.90893, speed: 1.40 step/s
global step 180, epoch: 1, batch: 180, loss: 0.03513, accu: 0.91016, speed: 1.40 step/s
global step 190, epoch: 1, batch: 190, loss: 0.21260, accu: 0.90938, speed: 1.40 step/s
global step 200, epoch: 1, batch: 200, loss: 0.43565, accu: 0.90906, speed: 1.40 step/s
eval loss: 0.20330, accu: 0.93083
global step 210, epoch: 1, batch: 210, loss: 0.25406, accu: 0.93750, speed: 0.49 step/s
global step 220, epoch: 1, batch: 220, loss: 0.24473, accu: 0.93750, speed: 1.39 step/s
global step 230, epoch: 1, batch: 230, loss: 0.30164, accu: 0.94271, speed: 1.40 step/s
global step 240, epoch: 1, batch: 240, loss: 0.30389, accu: 0.93516, speed: 1.39 step/s
global step 250, epoch: 1, batch: 250, loss: 0.26582, accu: 0.93063, speed: 1.40 step/s
global step 260, epoch: 1, batch: 260, loss: 0.17866, accu: 0.93073, speed: 1.40 step/s
global step 270, epoch: 1, batch: 270, loss: 0.11397, accu: 0.93304, speed: 1.40 step/s
global step 280, epoch: 1, batch: 280, loss: 0.13630, accu: 0.93281, speed: 1.40 step/s
global step 290, epoch: 1, batch: 290, loss: 0.13803, accu: 0.93229, speed: 1.40 step/s
global step 300, epoch: 1, batch: 300, loss: 0.06872, accu: 0.93312, speed: 1.43 step/s
eval loss: 0.17526, accu: 0.94083

3.1.5 预测提交结果

使用训练得到的模型还可以对文本进行情感预测。

import numpy as np
import paddle

# 处理测试集数据
trans_func = partial(
    convert_example,
    tokenizer=tokenizer,
    max_seq_length=max_seq_length,
    is_test=True)
batchify_fn = lambda samples, fn=Tuple(
    Pad(axis=0, pad_val=tokenizer.pad_token_id),  # input
    Pad(axis=0, pad_val=tokenizer.pad_token_type_id),  # segment
    Stack() # qid
): [data for data in fn(samples)]
test_data_loader = create_dataloader(
    test_ds,
    mode='test',
    batch_size=batch_size,
    batchify_fn=batchify_fn,
    trans_fn=trans_func)
# 根据实际运行情况,更换加载的参数路径
params_path = 'skep_ckp/model_500/model_state.pdparams'
if params_path and os.path.isfile(params_path):
    # 加载模型参数
    state_dict = paddle.load(params_path)
    model.set_dict(state_dict)
    print("Loaded parameters from %s" % params_path)
label_map = {0: '0', 1: '1'}
results = []
# 切换model模型为评估模式,关闭dropout等随机因素
model.eval()
for batch in test_data_loader:
    input_ids, token_type_ids, qids = batch
    # 喂数据给模型
    logits = model(input_ids, token_type_ids)
    # 预测分类
    probs = F.softmax(logits, axis=-1)
    idx = paddle.argmax(probs, axis=1).numpy()
    idx = idx.tolist()
    labels = [label_map[i] for i in idx]
    qids = qids.numpy().tolist()
    results.extend(zip(qids, labels))
res_dir = "./results"
if not os.path.exists(res_dir):
    os.makedirs(res_dir)
# 写入预测结果
with open(os.path.join(res_dir, "ChnSentiCorp.tsv"), 'w', encoding="utf8") as f:
    f.write("index\tprediction\n")
    for qid, label in results:
        f.write(str(qid[0])+"\t"+label+"\n")

3.2 目标级情感分析

在电商产品分析场景下,除了分析整体商品的情感极性外,还细化到以商品具体的“方面”为分析主体进行情感分析(aspect-level),如下、:

  • 这个薯片口味有点咸,太辣了,不过口感很脆。

关于薯片的口味方面是一个负向评价(咸,太辣),然而对于口感方面却是一个正向评价(很脆)。

  • 我很喜欢夏威夷,就是这边的海鲜太贵了。

关于夏威夷是一个正向评价(喜欢),然而对于夏威夷的海鲜却是一个负向评价(价格太贵)。

利用预训练模型skep进行情感分析的方法是_模型训练平台

目标级情感分析任务

常用数据集

千言数据集已提供了许多任务常用数据集。

其中情感分析数据集下载链接:千言数据集:情感分析_千言数据集评测-飞桨AI Studio星河社区 (baidu.com)

SE-ABSA16_PHNS数据集是关于手机的目标级情感分析数据集。PaddleNLP已经内置了该数据集,加载方式,如下:

train_ds, test_ds = load_dataset("seabsa16", "phns", splits=["train", "test"])

print(train_ds[0])
print(train_ds[1])
print(train_ds[2])
SKEP模型加载

目标级情感分析模型同样使用SkepForSequenceClassification模型,但目标级情感分析模型的输入不单单是一个句子,而是句对。一个句子描述“评价对象方面(aspect)”,另一个句子描述”对该方面的评论”。如下图所示。

利用预训练模型skep进行情感分析的方法是_模型训练平台

# 指定模型名称一键加载模型
model = SkepForSequenceClassification.from_pretrained(
    'skep_ernie_1.0_large_ch', num_classes=len(train_ds.label_list))
# 指定模型名称一键加载tokenizer
tokenizer = SkepTokenizer.from_pretrained('skep_ernie_1.0_large_ch')

数据处理

同样地,我们需要将原始SE_ABSA16_PHNS数据处理成模型可以读入的数据格式。

SKEP模型对中文文本处理按照字粒度进行处理,我们可以使用PaddleNLP内置的SkepTokenizer完成一键式处理。

from functools import partial
import os
import time

import numpy as np
import paddle
import paddle.nn.functional as F
from paddlenlp.data import Stack, Tuple, Pad


def convert_example(example,
                    tokenizer,
                    max_seq_length=512,
                    is_test=False,
                    dataset_name="chnsenticorp"):
    """
    Builds model inputs from a sequence or a pair of sequence for sequence classification tasks
    by concatenating and adding special tokens. And creates a mask from the two sequences passed 
    to be used in a sequence-pair classification task.
        
    A skep_ernie_1.0_large_ch/skep_ernie_2.0_large_en sequence has the following format:
    ::
        - single sequence: ``[CLS] X [SEP]``
        - pair of sequences: ``[CLS] A [SEP] B [SEP]``

    A skep_ernie_1.0_large_ch/skep_ernie_2.0_large_en sequence pair mask has the following format:
    ::

        0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1
        | first sequence    | second sequence |

    If `token_ids_1` is `None`, this method only returns the first portion of the mask (0s).
    
    note: There is no need token type ids for skep_roberta_large_ch model.


    Args:
        example(obj:`list[str]`): List of input data, containing text and label if it have label.
        tokenizer(obj:`PretrainedTokenizer`): This tokenizer inherits from :class:`~paddlenlp.transformers.PretrainedTokenizer` 
            which contains most of the methods. Users should refer to the superclass for more information regarding methods.
        max_seq_len(obj:`int`): The maximum total input sequence length after tokenization. 
            Sequences longer than this will be truncated, sequences shorter will be padded.
        is_test(obj:`False`, defaults to `False`): Whether the example contains label or not.
        dataset_name((obj:`str`, defaults to "chnsenticorp"): The dataset name, "chnsenticorp" or "sst-2".

    Returns:
        input_ids(obj:`list[int]`): The list of token ids.
        token_type_ids(obj: `list[int]`): List of sequence pair mask.
        label(obj:`numpy.array`, data type of int64, optional): The input label if not is_test.
    """
    encoded_inputs = tokenizer(
        text=example["text"],
        text_pair=example["text_pair"],
        max_seq_len=max_seq_length)

    input_ids = encoded_inputs["input_ids"]
    token_type_ids = encoded_inputs["token_type_ids"]

    if not is_test:
        label = np.array([example["label"]], dtype="int64")
        return input_ids, token_type_ids, label
    else:
        return input_ids, token_type_ids
# 处理的最大文本序列长度
max_seq_length=256
# 批量数据大小
batch_size=16

# 将数据处理成model可读入的数据格式
trans_func = partial(
    convert_example,
    tokenizer=tokenizer,
    max_seq_length=max_seq_length)
# 将数据组成批量式数据,如
# 将不同长度的文本序列padding到批量式数据中最大长度
# 将每条数据label堆叠在一起
batchify_fn = lambda samples, fn=Tuple(
    Pad(axis=0, pad_val=tokenizer.pad_token_id),  # input_ids
    Pad(axis=0, pad_val=tokenizer.pad_token_type_id),  # token_type_ids
    Stack(dtype="int64")  # labels
): [data for data in fn(samples)]
train_data_loader = create_dataloader(
    train_ds,
    mode='train',
    batch_size=batch_size,
    batchify_fn=batchify_fn,
    trans_fn=trans_func)

模型训练

定义损失函数、优化器以及评价指标后,即可开始训练。

# 训练轮次
epochs = 3
# 总共需要训练的step数
num_training_steps = len(train_data_loader) * epochs
# 优化器
optimizer = paddle.optimizer.AdamW(
    learning_rate=5e-5,
    parameters=model.parameters())
# 交叉熵损失
criterion = paddle.nn.loss.CrossEntropyLoss()
# Accuracy评价指标
metric = paddle.metric.Accuracy()
# 开启训练
ckpt_dir = "skep_aspect"
global_step = 0
tic_train = time.time()
for epoch in range(1, epochs + 1):
    for step, batch in enumerate(train_data_loader, start=1):
        input_ids, token_type_ids, labels = batch
        # 喂数据给model
        logits = model(input_ids, token_type_ids)
        # 计算损失函数值
        loss = criterion(logits, labels)
        # 预测分类概率
        probs = F.softmax(logits, axis=1)
        # 计算acc
        correct = metric.compute(probs, labels)
        metric.update(correct)
        acc = metric.accumulate()

        global_step += 1
        if global_step % 10 == 0:
            print(
                "global step %d, epoch: %d, batch: %d, loss: %.5f, acc: %.5f, speed: %.2f step/s"
                % (global_step, epoch, step, loss, acc,
                    10 / (time.time() - tic_train)))
            tic_train = time.time()
        
        # 反向梯度回传,更新参数
        loss.backward()
        optimizer.step()
        optimizer.clear_grad()

        if global_step % 100 == 0:
            
            save_dir = os.path.join(ckpt_dir, "model_%d" % global_step)
            if not os.path.exists(save_dir):
                os.makedirs(save_dir)
            # 保存模型参数
            model.save_pretrained(save_dir)
            # 保存tokenizer的词表等
            tokenizer.save_pretrained(save_dir)

global step 10, epoch: 1, batch: 10, loss: 0.65064, acc: 0.53125, speed: 1.27 step/s
global step 20, epoch: 1, batch: 20, loss: 0.52287, acc: 0.55312, speed: 1.26 step/s
global step 30, epoch: 1, batch: 30, loss: 0.71099, acc: 0.57083, speed: 1.27 step/s
global step 40, epoch: 1, batch: 40, loss: 0.70976, acc: 0.59062, speed: 1.27 step/s
global step 50, epoch: 1, batch: 50, loss: 0.62593, acc: 0.59000, speed: 1.26 step/s
global step 60, epoch: 1, batch: 60, loss: 0.70332, acc: 0.58542, speed: 1.26 step/s
global step 70, epoch: 1, batch: 70, loss: 0.52017, acc: 0.59911, speed: 1.25 step/s
global step 80, epoch: 1, batch: 80, loss: 0.64913, acc: 0.60781, speed: 1.27 step/s
global step 90, epoch: 2, batch: 6, loss: 0.56703, acc: 0.60824, speed: 1.30 step/s
global step 100, epoch: 2, batch: 16, loss: 0.59230, acc: 0.61746, speed: 1.26 step/s
global step 110, epoch: 2, batch: 26, loss: 0.74638, acc: 0.61473, speed: 0.84 step/s
global step 120, epoch: 2, batch: 36, loss: 0.67488, acc: 0.62134, speed: 1.25 step/s
global step 130, epoch: 2, batch: 46, loss: 0.60215, acc: 0.62307, speed: 1.27 step/s
global step 140, epoch: 2, batch: 56, loss: 0.47045, acc: 0.63172, speed: 1.26 step/s
global step 150, epoch: 2, batch: 66, loss: 0.53512, acc: 0.63253, speed: 1.27 step/s
global step 160, epoch: 2, batch: 76, loss: 0.59317, acc: 0.63597, speed: 1.26 step/s
global step 170, epoch: 3, batch: 2, loss: 0.50540, acc: 0.63794, speed: 1.31 step/s
global step 180, epoch: 3, batch: 12, loss: 0.69784, acc: 0.63827, speed: 1.25 step/s
global step 190, epoch: 3, batch: 22, loss: 0.57723, acc: 0.64451, speed: 1.26 step/s
global step 200, epoch: 3, batch: 32, loss: 0.53240, acc: 0.64667, speed: 1.26 step/s
global step 210, epoch: 3, batch: 42, loss: 0.87506, acc: 0.64713, speed: 0.86 step/s
global step 220, epoch: 3, batch: 52, loss: 0.60447, acc: 0.64755, speed: 1.26 step/s
global step 230, epoch: 3, batch: 62, loss: 0.51687, acc: 0.64793, speed: 1.26 step/s
global step 240, epoch: 3, batch: 72, loss: 0.57719, acc: 0.65272, speed: 1.25 step/s
global step 250, epoch: 3, batch: 82, loss: 0.43986, acc: 0.65487, speed: 1.29 step/s

预测提交结果

使用训练得到的模型还可以对评价对象进行情感预测。

@paddle.no_grad()
def predict(model, data_loader, label_map):
    """
    Given a prediction dataset, it gives the prediction results.

    Args:
        model(obj:`paddle.nn.Layer`): A model to classify texts.
        data_loader(obj:`paddle.io.DataLoader`): The dataset loader which generates batches.
        label_map(obj:`dict`): The label id (key) to label str (value) map.
    """
    model.eval()
    results = []
    for batch in data_loader:
        input_ids, token_type_ids = batch
        logits = model(input_ids, token_type_ids)
        probs = F.softmax(logits, axis=1)
        idx = paddle.argmax(probs, axis=1).numpy()
        idx = idx.tolist()
        labels = [label_map[i] for i in idx]
        results.extend(labels)
    return results
# 处理测试集数据
label_map = {0: '0', 1: '1'}
trans_func = partial(
    convert_example,
    tokenizer=tokenizer,
    max_seq_length=max_seq_length,
    is_test=True)
batchify_fn = lambda samples, fn=Tuple(
    Pad(axis=0, pad_val=tokenizer.pad_token_id),  # input_ids
    Pad(axis=0, pad_val=tokenizer.pad_token_type_id),  # token_type_ids
): [data for data in fn(samples)]
test_data_loader = create_dataloader(
    test_ds,
    mode='test',
    batch_size=batch_size,
    batchify_fn=batchify_fn,
    trans_fn=trans_func)
# 根据实际运行情况,更换加载的参数路径
params_path = 'skep_ckpt/model_900/model_state.pdparams'
if params_path and os.path.isfile(params_path):
    # 加载模型参数
    state_dict = paddle.load(params_path)
    model.set_dict(state_dict)
    print("Loaded parameters from %s" % params_path)

results = predict(model, test_data_loader, label_map)
# 写入预测结果
with open(os.path.join("results", "SE-ABSA16_PHNS.tsv"), 'w', encoding="utf8") as f:
    f.write("index\tprediction\n")
    for idx, label in enumerate(results):
        f.write(str(idx)+"\t"+label+"\n")
#将预测文件结果压缩至zip文件,提交
!zip -r results.zip results

wAAACH5BAEKAAAALAAAAAABAAEAAAICRAEAOw==

updating: results/ (stored 0%)
updating: results/ChnSentiCorp.tsv (deflated 63%)
updating: results/SE-ABSA16_PHNS.tsv (deflated 64%)

今天的文章利用预训练模型skep进行情感分析的方法是_模型训练平台分享到此就结束了,感谢您的阅读。

版权声明:本文内容由互联网用户自发贡献,该文观点仅代表作者本人。本站仅提供信息存储空间服务,不拥有所有权,不承担相关法律责任。如发现本站有涉嫌侵权/违法违规的内容, 请发送邮件至 举报,一经查实,本站将立刻删除。
如需转载请保留出处:https://bianchenghao.cn/83927.html

(0)
编程小号编程小号

相关推荐

发表回复

您的电子邮箱地址不会被公开。 必填项已用 * 标注