随着数字化经济的进一步发展,越来越多的企业开始将现代信息网络作为数据资源的主要载体,并通过网络通信技术进行数据传输。同时,疫情也促使越来越多行业逐步将互联网作为主要的信息交流和分享的方式。以往的研究表明,会议记录的口语语言处理(SLP)技术如关键词提取和摘要,对于信息的提取、组织和排序至关重要,可以显著提高用户对重要信息的掌握效率。
本项目源自于ICASSP2023信号处理大挑战的通用会议理解及生成挑战赛(MUG challenge),赛事构建并发布了目前为止规模最大的中文会议数据集,并基于会议人工转写结果进行了多项SLP任务的标注;目标是推动SLP在会议文本处理场景的研究并应对其中的多项关键挑战,包括 人人交互场景下多样化的口语现象、会议场景下的长篇章文档建模 等。
赛事报名页面:
https://modelscope.cn/competition/12/summary - 话题分割
https://modelscope.cn/competition/13/summary - 抽取式摘要
https://modelscope.cn/competition/14/summary - 话题子标题生成
https://modelscope.cn/competition/17/summary - 行动项抽取
https://modelscope.cn/competition/18/summary - 关键词抽取
基线模型训练及推理:
https://github.com/alibaba-damo-academy/SpokenNLP
PALM预训练语言生成模型是针对实际场景中常见的文本生成需求所设计的一个模型。模型利用大量无监督数据,通过结合自编码和自回归任务进行预训练,更贴合下游生成任务所同时需要的理解和生成能力。本模型是PALM通用预训练生成模型,在竞赛子标题生成数据集上进行finetune得到的会议子标题生成模型。PALM模型介绍,详见:PALM:Pre-training an Autoencoding&Autoregressive Language Model for Context-conditioned Generation
本模型主要用于给输入文档生成摘要内容。用户可以自行尝试各种输入文档。具体调用方式请参考代码示例。
在安装完成ModelScope-library之后即可使用text-generation的能力
from modelscope.pipelines import pipeline
from modelscope.utils.constant import Tasks
from modelscope.outputs import OutputKeys
input = '哎大家好啊欢迎大家准时来参加我们的会议,今天我们会议的主题呢是如何提升我们这个洗发水儿的品牌影响力啊。我们现在是主要的产品是这个霸王防脱洗发水儿,现在请大家发表自己的意见啊欢迎大家努力的发表,请先从这位男士开始。是这个。'
text_summary = pipeline(Tasks.text_generation, model='damo/nlp_palm2.0_text-generation_meeting_title_chinese-base')
result = text_summary(input)
print('输入文本:\n' + input + '\n')
print('文本生成结果:\n' + result[OutputKeys.TEXT])
模型在新闻相关数据集上训练,在新闻等类似文章上摘要生成性能较好,其他垂直领域效果可能会有所下降。
本模型是由竞赛子标题数据集训练得到的, 具体数据可以参考右侧标签栏
用户可以基于这个PALM预训练底座模型进一步优化,训练代码如下,更详细代码可参考alimeeting4mug
import tempfile
import os
from modelscope.msdatasets import MsDataset
from modelscope.metainfo import Trainers
from modelscope.trainers import build_trainer
from modelscope.utils.hub import read_config
from modelscope.metainfo import Metrics, Trainers
from datasets import Dataset
# ---------------------------- Train ---------------------------------
# 用户自己数据集构造
tmp_dir = "./sub_title_generation_models/"
data_pth = "./ttg_transformer"
def load_local_data(data_pth):
train_dataset_dict = {"src_txt": [], "tgt_txt": []}
with open(os.path.join(data_pth, "train.source"), "r") as f:
for line in f:
train_dataset_dict["src_txt"].append(line.strip())
with open(os.path.join(data_pth, "train.target"), "r") as f:
for line in f:
train_dataset_dict["tgt_txt"].append(line.strip())
eval_dataset_dict = {"src_txt": [], "tgt_txt": []}
with open(os.path.join(data_pth, "val.source"), "r") as f:
for line in f:
eval_dataset_dict["src_txt"].append(line.strip())
with open(os.path.join(data_pth, "val.target"), "r") as f:
for line in f:
eval_dataset_dict["tgt_txt"].append(line.strip())
return train_dataset_dict, eval_dataset_dict
train_dataset_dict, eval_dataset_dict = load_local_data(data_pth)
train_dataset = MsDataset(Dataset.from_dict(train_dataset_dict))
eval_dataset = MsDataset(Dataset.from_dict(eval_dataset_dict))
num_warmup_steps = 500
def noam_lambda(current_step: int):
current_step += 1
return min(current_step**(-0.5),
current_step * num_warmup_steps**(-1.5))
# 可以在代码修改 configuration 的配置
def cfg_modify_fn(cfg):
cfg.train.lr_scheduler = {
'type': 'LambdaLR',
'lr_lambda': noam_lambda,
'options': {
'by_epoch': False
}
}
cfg.preprocessor = {
"type": "text-gen-tokenizer",
"sequence_length": 512
}
cfg.train.optimizer = {
"type": "AdamW",
"lr": 5e-4,
"options": {}
}
cfg.train.hooks[-1] = {
"type": "EvaluationHook",
"by_epoch": True,
"interval": 1
}
cfg.train.max_epochs = 10
cfg.train.dataloader = {
"batch_size_per_gpu": 16,
"workers_per_gpu": 1
}
return cfg
kwargs = dict(
model='damo/nlp_palm2.0_pretrained_chinese-base',
train_dataset=train_dataset,
eval_dataset=eval_dataset,
work_dir=tmp_dir,
cfg_modify_fn=cfg_modify_fn)
trainer = build_trainer(
name=Trainers.text_generation_trainer, default_args=kwargs)
trainer.train()
"""
# ---------------------------- Evaluation ---------------------------------
from modelscope.pipelines import pipeline
from modelscope.utils.constant import Tasks
from modelscope.outputs import OutputKeys
from modelscope.metrics import TextGenerationMetric
text_generation = pipeline(Tasks.text_generation,
model=os.path.join(tmp_dir, "output"),
sequence_length=512)
m = TextGenerationMetric()
outputs = {'preds': []}
inputs = {'tgts': []}
with open(os.path.join(data_pth, "val.source"), "r") as f:
for line in f:
output = text_generation(line.strip())['text']
outputs['preds'].append(output)
with open(os.path.join(data_pth, "val.target"), "r") as f:
for line in f:
inputs['tgts'].append(line.strip())
m.add(outputs, inputs)
print(f'{m.preds[:10]}\n{m.tgts[:10]}')
print(m.evaluate())
# ---------------------------- Inference ---------------------------------
output_list = []
with open(os.path.join(data_pth, "test.source"), "r") as f:
for line in f:
output = text_generation(line.strip())['text']
output_list.append(line.strip().replace("\t", " ") + "\t" + output)
with open(os.path.join(tmp_dir, "test_predict_result.txt"), "w") as f:
f.write("\n".join(output_list))
"""
模型在竞赛dev集上评估结果
Rouge-1 | Rouge-L |
---|---|
31.28 | 29.43 |
如果我们的模型对您有帮助,请您引用我们的文章:
@inproceedings{bi-etal-2020-palm,
title = "{PALM}: Pre-training an Autoencoding & Autoregressive Language Model for Context-conditioned Generation",
author = "Bi, Bin and
Li, Chenliang and
Wu, Chen and
Yan, Ming and
Wang, Wei and
Huang, Songfang and
Huang, Fei and
Si, Luo",
booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
month = nov,
year = "2020",
address = "Online",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2020.emnlp-main.700",
doi = "10.18653/v1/2020.emnlp-main.700",
pages = "8681--8691"}