FastTextClassification Save
FastTextClassification is all you need for text classification! Fast text classification for everyone, enjoy your NLP journey! 这可能是目前为止最全面的开源文本分类项目，支持中英双语、多种模型、多种任务。
FastTextClassification
功能免费，代码开源，大家放心使用，欢迎贡献！
最新讯息
- 2024/05/06:
- ernie支持多卡、slurm等分布式训练
- 支持配置文件改变mlp的层数、预训练模型等
- 支持二分类、多分类、多标签分类
- 2024/04/30: clone原项目进行再次开发
- 2023/03/23：FastTextClassification V0.0.1版正式开源，版本特性：
- 支持中英双语的文本分类
- 支持多种文本分类模型：传统机器学习浅层模型、深度学习模型和transformers类模型
- 支持多标签文本分类
- 支持多种embedding方式：inner/outer/random
开发计划
本项目的开发宗旨，打造全网最全面和最实用的文本分类项目和教程。如果有机会，未来希望可以做成开箱即用的文本分类工具，文本分类任务非常特殊，大部分情况下被认为是简单且基础的，然而却很难找到比较通用的文本分类工具，往往都是针对具体任务进行训练和部署。在NLP逐渐趋于大一统的今天，这一点非常不优雅，而且浪费资源。：*Fast text classification for you, Start your NLP journey!*
简要的开发计划：
- 【P3】支持中英双语的文本分类：100%，也欢迎支持其他语种
- 【P0】支持多种文本分类模型：基本完成，欢迎补充
- 浅层文本分类模型：done
- 【P1】DNN类模型：已支持常见模型
- 【P0】transformer类模型：Bert/ERNIE等
- 【P0】prompt learning for Text Classification：TODO
- 【P0】ChatGPT for Text Classification：TODO
- 【P1】支持多标签文本分类：
- 多种多标签分类loss：done，如有遗漏，欢迎补充
- 复杂的多标签分类：比如层次化等，TODO
- 【P0】支持不同的文本分类数据集/任务：文本分类任务又多又散，这是好事儿也是坏事儿。欢迎基于本项目报告各种数据集上的效果
- 【P4】支持简明易用的文本分类API：终极目标为实现一个足够通用和强大的文本分类模型，并实现自然语言交互的文本分类接口text_cls(text, candidate_labels)->label，给定文本和候选类别(有默认值)，输出文本所属的类别；同时支持可无成本或尽可能小的成本向特定领域泛化
- 多GPU和集群训练: TODO
使用步骤
1.克隆本项目
git clone https://github.com/fast-llm/FastTextClassification.git
2.数据集下载和预处理
请自行下载数据集，将其放到data目录下，数据统一处理成text+label格式，以\t或逗号分隔。有空我再来补一个自动化脚本，暂时请自行处理或者参考preprocessing.py。
最好将数据统一放到data目录下，比如data/dbpedia，然后分3个子目录，input存放原始数据集（你下载的数据集），data存放预处理后的格式化的数据集（text-label格式），saved_dict存放训练结果（模型和日志等）。
3.运行示例
经过测试的开发环境如下，仅供参考，差不多的环境应该都可以运行。
- python：3.10
- torch：2.3.0
- transformers：4.39.1
conda create -n fasttext python=3.10
conda activate fasttext
pip install poetry
poetry install
根据自己的需要选择模块运行，详见下一节。
python run.py
运行示例
1.运行DNN/transformers类模型做文本分类
python run.py
2.运行传统浅层机器学习模型做文本分类
python run_shallow.py
3.运行DNN/transformers类模型做多标签文本分类
python run_multi_label.py
下表是直接运行demo的参考结果：
运行环境：python3.6 + T4
| demo | 数据集 | 示例模型 | Acc | 耗时 | 备注 |
|---|---|---|---|---|---|
| run.py | THUCNews/cn | TextCNN | 89.94% | ~2mins | |
| run_multi_label.py | rcv1/en | bert | 61.04% | ~40mins | 其他指标见运行结果 |
| run_shallow.py | THUCNews/cn | NB | 89.44% | 105.34 ms |
结果展示：持续更新中
笔者提供了从浅到深再到多标签的详细实验结果，可供大家参考。但受限于时间和算力，很多实验可能未达到最优，望知悉！因此，非常欢迎大家积极贡献，补充相关实验、代码和新的模型等等，一起建设FastTextClassification。
暂时只提供部分汇总的结果，详细的实验结果及参数等我有空再补，比较多，需要一些时间整理。
1.传统浅层文本分类模型
| Data | Model | tokenizer | 最小词长 | Min_df | ngram | binary | Use_idf | Test acc | 备注 |
|---|---|---|---|---|---|---|---|---|---|
| THUCNews/cn | LR | lcut | 1 | 2 | (1,1) | False | True | 90.61% | C=1.0, max_iter=1000 词表61549； train score: 94.22% valid score: 89.84% test score: 90.61% training time: 175070.97 ms |
| MultinomialNB(alpha=0.3) | lcut | 1 | 2 | (1,1) | False | True | 89.86% | 词表61549； training time: 94.18ms | |
| ComplementNB(alpha=0.8) | lcut | 1 | 2 | (1,1) | False | True | 89.88% | 词表61549； training time: 98.31ms | |
| SVC(C=1.0) | lcut | 1 | 2 | (1,1) | False | True | 81.49% | 词表61549； 维度200 training time: 7351155.59 ms train score: 85.95% valid score: 80.07% test score: 81.49% | |
| DT | lcut | 1 | 2 | (1,1) | False | True | 71.19% | max_depth=None training time: 149216.53 ms train score: 99.97% valid score: 70.57% test score: 71.19% | |
| xgboost | lcut | 1 | 2 | (1,1) | False | True | 90.08% | XGBClassifier(n_estimators=2000,eta=0.3,gamma=0.1,max_depth=6,subsample=1,colsample_bytree=0.8, nthread=10) training time: 1551260.28 ms train score: 99.00% valid score: 89.34% test score: 90.08% | |
| KNN | lcut | 1 | 2 | (1,1) | False | True | 85.17% | k=10 training time: 21.24 ms train score: 89.05% valid score: 84.53% test score: 85.17% | |
| dbpedia/en | LR | None | 2 | 2 | (1,1) | False | True | 98.26% | C=1.0, max_iter=100 词表237777 training time: 220177.59 ms train score: 98.85% valid score: 98.19% test score: 98.26% |
| MultinomialNB(alpha=1.0) | None | 2 | 2 | (1,1) | False | True | 95.35% | training time: 786.24 ms train score: 96.36% valid score: 95.34% test score: 95.35% | |
| ComplementNB(alpha=1.0) | None | 2 | 2 | (1,1) | False | True | 93.73% | training time: 805.69 ms train score: 95.30% valid score: 93.79% test score: 93.73% | |
| SVC(C=1.0) | None | 2 | 2 | (1,1) | False | True | 94.67% | 维度200； max_iter=100 training time: 144163.81 ms train score: 94.75% valid score: 94.59% test score: 94.67% 注意：SVM的计算和存储成本正比于样本数的平方； | |
| DT | None | 2 | 2 | (1,1) | False | True | 92.41% | max_depth=100, min_samples_leaf=5 training time: 639744.56 ms train score: 95.79% valid score: 92.43% test score: 92.41% | |
| xgboost | None | 2 | 2 | (1,1) | False | True | 97.99% | XGBClassifier(n_estimators=200,eta=0.3,gamma=0.1,max_depth=6,subsample=1,colsample_bytree=0.8, nthread=10,reg_alpha=0,reg_lambda=1) training time: 1838434.42 ms train score: 99.35% valid score: 97.96% test score: 97.99% | |
| KNN | None | 2 | 2 | (1,1) | False | True | 80.05% | k=10 training time: 137.72 ms train score: 84.66% valid score: 80.20% test score: 80.05% | |
2.深度学习文本分类模型
| Data | Model | Embed | Bz | Lr | epochs | acc | 备注 |
|---|---|---|---|---|---|---|---|
| THUCNews/cn | TextCNN | outer | 128 | 1e-3 | 3/20 | 90.45% | |
| TextRNN | - | - | 1e-3 | 5/10 | 90.38% | ||
| TextRNN_Att | 1e-3 | 2/10 | 90.55% | ||||
| TextRCNN | 1e-3 | 3/10 | 91.01% | ||||
| DPCNN | 1e-3 | 3/20 | 90.12% | ||||
| FastText | 1e-3 | 5/20 | 90.48% | ||||
| bert | inner | 5e-5 | 2/3 | 94.10% | bert-base-chinese | ||
| ERNIE | inner | 5e-5 | 3/3 | 94.58% | ernie-3.0-base-zh | ||
| bert_CNN | - | 3/3 | 94.14% | ||||
| bert_RNN | - | 3/3 | 93.92% | ||||
| bert_RNN | - | 3/3 | 94.45% | ||||
| bert_RCNN | - | 3/3 | 94.32% | ||||
| bert_DPCNN | - | 3/3 | 94.17% | ||||
| dbpedia/en | TextCNN | outer | 128 | 5e-5 | 9/20 | 98.35% | glove |
| TextRNN | - | - | - | 6/10 | 97.97% | ||
| TextRNN_Att | - | 4/10 | 97.80% | ||||
| TextRCNN | - | 3/10 | 97.71% | ||||
| DPCNN | - | 3/20 | 97.86% | ||||
| FastText | - | 10/20 | 97.84% | ||||
| bert | inner | 5e-5 | 2/3 | 97.78% | bert-base-uncased | ||
| ERNIE | 2/10 | 97.75% | ernie-2.0-base-en | ||||
| bert_CNN | - | 2/3 | 97.91% | ||||
| bert_RNN | - | 2/3 | 97.87% | ||||
| bert_RCNN | - | 2/3 | 98.04% | ||||
| bert_DPCNN | - | 2/3 | 97.95% | ||||
| gpt | 3/3 | 97.03 | |||||
| gpt2 | 3/3 | 97.00 | |||||
| T5 | 3/3 | 96.57 | |||||
3.多标签文本分类
| Data | Model | 分层 | 样本数 | Embed | loss | Bz | Lr | epochs | Test acc (绝对匹配率） | Micro-F1 | Macro-F1 | 备注 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Rcv1/en | TextCNN | - | all | outer | multi_label_circle_loss | 128 | 1e-3 | 9/20 | 51.02% | 0.7904 | 0.4515 | eval_activate = None cls_threshold = 0 |
| TextRNN | - | - | - | 13/20 | 54.00% | 0.7950 | 0.4358 | |||||
| TextRNN_Att | - | 11/20 | 53.97% | 0.8011 | 0.4538 | |||||||
| TextRCNN | - | 10/20 | 53.62% | 0.8111 | 0.4900 | |||||||
| DPCNN | - | 10/20 | 51.66% | 0.7890 | 0.4111 | |||||||
| FastText | - | 12/20 | 51.31% | 0.7936 | 0.4728 | |||||||
| bert | all | inner | - | 128 | 2e-5 | 20/20 | 61.04% | 0.8454 | 0.5729 | bert-base-cased | ||
| ERNIE | all | inner | - | 128 | 2e-5 | 20/20 | 61.67% | 0.8486 | 0.5861 | ernie-2.0-base-en | ||
| Bert_CNN | all | inner | - | 128 | 2e-5 | 12/20 | 58.31% | 0.8364 | 0.5736 | 同bert配置 | ||
| Bert_RNN | all | inner | - | 128 | 2e-5 | 17/20 | 60.48% | 0.8371 | 0.5640 | |||
| Bert_RCNN | all | inner | - | 128 | 2e-5 | 15/20 | 60.54% | 0.8457 | 0.5969 | |||
| Bert_DPCNN | all | inner | - | 128 | 2e-5 | 13/20 | 56.52% | 0.8082 | 0.4273 | |||
常见报错
参考资料&致谢
A Survey on Text Classification: From Shallow to Deep Learning：https://arxiv.org/pdf/2008.00364.pdf?utm_source=summari
Deep Learning--based Text Classification: A Comprehensive Review：https://arxiv.org/pdf/2004.03705.pdf
https://github.com/649453932/Chinese-Text-Classification-Pytorch
https://github.com/649453932/Bert-Chinese-Text-Classification-Pytorch
https://github.com/facebookresearch/fastText
https://github.com/brightmart/text_classification
https://github.com/kk7nc/Text_Classification
https://github.com/Tencent/NeuralNLP-NeuralClassifier
https://github.com/vandit15/Class-balanced-loss-pytorch
https://scikit-learn.org/stable/modules/model_evaluation.html#classification-metrics
赞助我们
Starchart
贡献者
Open Source Agenda Badge
