4 个月前

DeepStruct:用于结构预测的语言模型预训练

DeepStruct:用于结构预测的语言模型预训练

摘要

我们提出了一种提高语言模型结构理解能力的方法。与以往通过任务特定增强来微调模型的方法不同,我们在一组任务无关的语料库上预训练语言模型,以从文本中生成结构。我们的结构预训练使模型所学到的结构知识能够实现零样本迁移。我们在这种方法上研究了其在28个数据集上的性能,这些数据集涵盖了10项结构预测任务,包括开放信息抽取、联合实体和关系抽取、命名实体识别、关系分类、语义角色标注、事件抽取、共指消解、事实探测、意图检测和对话状态跟踪。此外,我们还通过任务特定的训练集进一步增强了预训练过程。实验结果表明,一个参数量为100亿的语言模型能够在大多数任务上实现非平凡的迁移,并在我们评估的28个数据集中有21个达到了当前最佳性能。

代码仓库

cgraywang/deepstruct
官方
pytorch
GitHub 中提及

基准测试

基准方法指标
coreference-resolution-on-conll12DeepStruct multi-task
Average F1: 60.6
B3: 57.7
CEAFϕ4: 60.2
MUC: 63.9
coreference-resolution-on-conll12DeepStruct multi-task w/ finetune
Average F1: 73.1
B3: 71.3
CEAFϕ4: 73.1
MUC: 74.9
dialogue-state-tracking-on-multiwoz-2-1DeepStruct multi-task w/ finetune
Joint Acc: 54.2
dialogue-state-tracking-on-multiwoz-2-1DeepStruct multi-task
Joint Acc: 53.5
event-extraction-on-ace2005DeepStruct multi-task
Argument Cl: 63.9
Argument Id: 67.5
Trigger Cl: 69.2
Trigger Id: 72.7
event-extraction-on-ace2005DeepStruct multi-task w/ finetune
Argument Cl: 56.2
Argument Id: 59.4
Trigger Cl: 69.8
Trigger Id: 73.5
joint-entity-and-relation-extraction-on-2DeepStruct multi-task w/ finetune
Entity F1: 90.7
Relation F1: 78.3
joint-entity-and-relation-extraction-on-2Deepstruct zero-shot
Entity F1: 48.3
Relation F1: 25.8
joint-entity-and-relation-extraction-on-2DeepStruct multi-task
Entity F1: 88.4
Relation F1: 72.8
joint-entity-and-relation-extraction-on-7DeepStruct multi-task
Entity F1: 90.2
Relation F1: 58.9
joint-entity-and-relation-extraction-on-7DeepStruct multi-task w/ finetune
Entity F1: 90.0
Relation F1: 66.8
joint-entity-and-relation-extraction-on-7Deepstruct zero-shot
Entity F1: 31.8
Relation F1: 5.3
joint-entity-and-relation-extraction-on-ade-1Deepstruct zero-shot
Entity F1: 60.7
Relation F1: 10.6
joint-entity-and-relation-extraction-on-ade-1DeepStruct multi-task
Entity F1: 90.5
Relation F1: 83.6
joint-entity-and-relation-extraction-on-ade-1DeepStruct multi-task w/ finetune
Entity F1: 91.1
Relation F1: 83.8
joint-entity-and-relation-extraction-on-nytDeepStruct multi-task w/ finetune
Entity F1: 95.9
Relation F1: 93.3
joint-entity-and-relation-extraction-on-nytDeepStruct multi-task
Entity F1: 95.4
Relation F1: 93.7
joint-entity-and-relation-extraction-on-nytDeepstruct zero-shot
Entity F1: 60.5
Relation F1: 28.6
named-entity-recognition-on-ace2005Deepstruct zero-shot
F1: 28.1
named-entity-recognition-on-ace2005DeepStruct multi-task w/ finetune
F1: 86.9
named-entity-recognition-on-conll03DeepStruct multi-task
F1: 93.1
named-entity-recognition-on-conll03Deepstruct zero-shot
F1: 44.4
named-entity-recognition-on-conll03DeepStruct multi-task w/ finetune
F1: 93.0
named-entity-recognition-on-geniaDeepStruct multi-task
F1: 80.2
named-entity-recognition-on-geniaDeepStruct multi-task w/ finetune
F1: 80.8
named-entity-recognition-on-geniaDeepstruct zero-shot
F1: 47.2
named-entity-recognition-on-ontonotesDeepstruct zero-shot
F1: 2.5
named-entity-recognition-on-ontonotesDeepStruct multi-task
F1: 87.6
named-entity-recognition-on-ontonotesDeepStruct multi-task w/ finetune
F1: 87.8
open-information-extraction-on-nytDeepStruct multi-task
F1: 43.6
open-information-extraction-on-nytDeepstruct zero-shot
F1: 28.9
open-information-extraction-on-nytDeepStruct multi-task w/ finetune
F1: 45.0
open-information-extraction-on-oie2016Deepstruct zero-shot
F1: 28.1
open-information-extraction-on-oie2016DeepStruct multi-task w/ finetune
F1: 71.3
open-information-extraction-on-oie2016Deepstruct multi-task
F1: 71.2
open-information-extraction-on-penn-treebankDeepStruct multi-task w/ finetune
F1: 45,1
open-information-extraction-on-penn-treebankDeepStruct multi-task
F1: 54.5
open-information-extraction-on-penn-treebankDeepstruct zero-shot
F1: 51
open-information-extraction-on-webDeepStruct multi-task
F1: 50.8
open-information-extraction-on-webDeepStruct multi-task w/ finetune
F1: 49.1
open-information-extraction-on-webDeepstruct zero-shot
F1: 43.8
relation-classification-on-fewrel-1Deepstruct zero-shot
F1 (10-way 1-shot): 67.6
F1 (10-way 5-shot): 66.4
F1 (5-way 1-shot): 72.4
F1 (5-way 5-shot: 70.8
relation-classification-on-fewrel-1DeepStruct multi-task w/ finetune
F1 (10-way 1-shot): 97.8
F1 (10-way 5-shot): 99.8
F1 (5-way 1-shot): 98.4
F1 (5-way 5-shot: 100
relation-classification-on-fewrel-1DeepStruct multi-task
F1 (10-way 1-shot): 92.2
F1 (10-way 5-shot): 94.6
F1 (5-way 1-shot): 93.6
F1 (5-way 5-shot: 96.4
relation-classification-on-tacred-1Deepstruct zero-shot
F1: 36.1
relation-classification-on-tacred-1DeepStruct multi-task w/ finetune
F1: 76.8
relation-classification-on-tacred-1DeepStruct multi-task
F1: 74.9
relation-extraction-on-tacredDeepStruct multi-task w/ finetune
F1: 76.8
semantic-role-labeling-on-conll05-brownDeepStruct multi-task w/ finetune
F1: 92.1
semantic-role-labeling-on-conll05-brownDeepStruct multi-task
F1: 92.0
semantic-role-labeling-on-conll05-wsjDeepStruct multi-task w/ finetune
F1: 95.2
semantic-role-labeling-on-conll05-wsjDeepStruct multi-task
F1: 95.5
semantic-role-labeling-on-conll12DeepStruct multi-task
F1: 97.2
semantic-role-labeling-on-conll12DeepStruct multi-task w/ finetune
F1: 96.0

用 AI 构建 AI

从想法到上线——通过免费 AI 协同编程、开箱即用的环境和市场最优价格的 GPU 加速您的 AI 开发

AI 协同编程
即用型 GPU
最优价格
立即开始

Hyper Newsletters

订阅我们的最新资讯
我们会在北京时间 每周一的上午九点 向您的邮箱投递本周内的最新更新
邮件发送服务由 MailChimp 提供
DeepStruct:用于结构预测的语言模型预训练 | 论文 | HyperAI超神经