UniMC 核心思想是将自然语言理解任务转化为 multiple choice 任务，并且使用多个 NLU 任务来进行预训练。我们在英文数据集实验结果表明仅含有 2.35 亿参数的 ALBERT模型的zero-shot性能可以超越众多千亿的模型。并在中文测评基准 FewCLUE 和 ZeroCLUE 两个榜单中，13亿的二郎神获得了第一的成绩。

The core idea of UniMC is to convert natural language understanding tasks into multiple choice tasks and use multiple NLU tasks for pre-training. Our experimental results on the English dataset show that the zero-shot performance of a ALBERT model with only 235 million parameters can surpass that of many hundreds of billions of models. And in the Chinese evaluation benchmarks FewCLUE and ZeroCLUE two lists, 1.3 billion Erlangshen won the first result.

模型分类 Model Taxonomy

需求 Demand	任务 Task	系列 Series	模型 Model	参数 Parameter	额外 Extra
通用 General	自然语言理解 NLU	二郎神 Erlangshen	MegatronBERT	1.3B	Chinese

模型信息 Model Information

我们为零样本学习者提出了一种与输入无关的新范式，从某种意义上说，它与任何格式兼容并适用于一系列语言任务，例如文本分类、常识推理、共指解析、情感分析。我们的方法将零样本学习转化为多项选择任务，避免常用的大型生成模型（如 FLAN）中的问题。它不仅增加了模型的泛化能力，而且显着减少了对参数的需求。我们证明了这种方法可以在通用语言基准上取得最先进的性能，并在自然语言推理和文本分类等任务上产生令人满意的结果。更多详细信息可以参考我们的论文或者GitHub

We propose an new paradigm for zero-shot learners that is input-agnostic, in the sense that it is compatible with any format and applicable to a list of language tasks, such as text classification, commonsense reasoning, coreference resolution, sentiment analysis. Our approach converts zero-shot learning into multiple choice tasks, avoiding problems in commonly used large generative models such as FLAN. It not only adds generalization ability to the models, but also reduces the needs of parameters significantly. We demonstrate that this approach leads to state-of-the-art performance on common language benchmarks, and produces satisfactory results on tasks such as natural language inference and text classification. For more details, please refer to our paper or github

下游效果 Performance

Few-shot

Model	eprstmt	csldcp	tnews	iflytek	ocnli	bustm	chid	csl	wsc	Avg
FineTuning-RoBERTa-110M	65.4	35.5	49	32.8	33	60.7	14.9	50	55.6	44.1
FineTuning-ERNIE1.0-110M	66.5	57	516	42.1	32	60.4	15	60.1	50.3	48.34
PET-ERNIE1.0-110M	84	59.9	56.4	50.3	38.1	58.4	40.6	61.1	58.7	56.39
P-tuning-ERNIE1.0-110M	80.6	56.6	55.9	52.6	35.7	60.8	39.61	51.8	55.7	54.37
EFL-ERNIE1.0-110M	76.7	47.9	56.3	52.1	48.7	54.6	30.3	52.8	52.3	52.7
UniMC-RoBERTa-110M	88.64	54.08	54.32	48.6	66.55	73.76	67.71	52.54	59.92	62.86
UniMC-RoBERTa-330M	89.53	57.3	54.25	50	70.59	77.49	78.09	55.73	65.16	66.46
UniMC-MegatronBERT-1.3B	89.278	60.9	57.46	52.89	76.33	80.37	90.33	61.73	79.15	72.05

Zero-shot

Model	eprstmt	csldcp	tnews	iflytek	ocnli	bustm	chid	csl	wsc	Avg
GPT-110M	57.5	26.2	37	19	34.4	50	65.6	50.1	50.3	43.4
PET-RoBERTa-110M	85.2	12.6	26.1	26.6	40.3	50.6	57.6	52.2	54.7	45.1
NSP-BERT-110M	86.9	47.6	51	41.6	37.4	63.4	52	64.4	59.4	55.96
ZeroPrompt-T5-1.5B	-	-	-	16.14	46.16	-	-	-	47.98	-
Yuan1.0-13B	88.13	38.99	57.47	38.82	48.13	59.38	86.14	50	38.99	56.22
ERNIE3.0-240B	88.75	50.97	57.83	40.42	53.57	64.38	87.13	56.25	53.46	61.41
UniMC-RoBERTa-110M	86.16	31.26	46.61	26.54	66.91	73.34	66.68	50.09	53.66	55.7
UniMC-RoBERTa-330M	87.5	30.4	47.6	31.5	69.9	75.9	78.17	49.5	60.55	59.01
UniMC-MegatronBERT-1.3B	88.79	42.06	55.21	33.93	75.57	79.5	89.4	50.25	66.67	64.53

Full dataset

Model	AFQMC	TNEWS1.1	IFLYTEK	OCNLI	CMNLI	WSC1.1	CSL	CHID	C3
RoBERTa-Base	74.06	57.5	60.36	74.3	79.73	83.48	85.37	-	-
RoBERTa-Large	74.88	58.79	61.52	77.7	81.4	89.14	86	-	-
Erlangshen-MegatronBert-1.3B 「Finetuning」	76.08	59.38	62.34	79.14	81	92.43	87.2	84.65	86.77
Erlangshen-UniMC-MegatronBERT-1.3B-Chinese	77.09	60.4	62.67	83.05	84.76	93.74	87.67	85.93	86.54

使用 Usage

shell

git clone https://github.com/IDEA-CCNL/Fengshenbang-LM.git
cd Fengshenbang-LM
pip install --editable .

python3

import argparse
from fengshen.pipelines.multiplechoice import UniMCPipelines


total_parser = argparse.ArgumentParser("TASK NAME")
total_parser = UniMCPipelines.piplines_args(total_parser)
args = total_parser.parse_args()
pretrained_model_path = 'IDEA-CCNL/Erlangshen-UniMC-MegatronBERT-1.3B-Chinese'
args.learning_rate=2e-5
args.max_length=512
args.max_epochs=3
args.batchsize=8
args.default_root_dir='./'
model = UniMCPipelines(args, pretrained_model_path)

train_data = []
dev_data = []
test_data = [
        {"texta": "放弃了途观L和荣威RX5，果断入手这部车，外观霸气又好开", 
         "textb": "", 
         "question": "下面新闻属于哪一个类别？", 
         "choice": [
            "房产", 
            "汽车", 
            "教育", 
            "科技"
            ], 
         "answer": "汽车", 
         "label": 1, 
         "id": 7759}
    ]

if args.train:
    model.train(train_data, dev_data)
result = model.predict(test_data)
for line in result[:20]:
    print(line)

引用 Citation

如果您在您的工作中使用了我们的模型，可以引用我们的论文：

If you are using the resource for your work, please cite the our paper:

text

@article{unimc,
  author    = {Ping Yang and
               Junjie Wang and
               Ruyi Gan and
               Xinyu Zhu and
               Lin Zhang and
               Ziwei Wu and
               Xinyu Gao and
               Jiaxing Zhang and
               Tetsuya Sakai},
  title     = {Zero-Shot Learners for Natural Language Understanding via a Unified Multiple Choice Perspective},
  journal   = {CoRR},
  volume    = {abs/2210.08590},
  year      = {2022}
}

也可以引用我们的网站:

You can also cite our website:

text

@misc{Fengshenbang-LM,
  title={Fengshenbang-LM},
  author={IDEA-CCNL},
  year={2021},
  howpublished={\url{https://github.com/IDEA-CCNL/Fengshenbang-LM}},
}