基于层次化自适应融合机制和类别边界蒸馏的持续少样本事件检测模型

doi:10.11772/j.issn.1001-9081.2025050583

《计算机应用》唯一官方网站 ›› 2026, Vol. 46 ›› Issue (5): 1450-1459.DOI: 10.11772/j.issn.1001-9081.2025050583

• 人工智能 • 上一篇

基于层次化自适应融合机制和类别边界蒸馏的持续少样本事件检测模型

胡婕¹^,²^,³(), 徐彤¹, 张龑¹^,²^,³

^1.湖北大学计算机学院，武汉 430062
^2.大数据智能分析与行业应用湖北省重点实验室（湖北大学），武汉 430062
^3.智能感知系统与安全教育部重点实验室（湖北大学），武汉 430062

收稿日期:2025-05-28 修回日期:2025-08-13 接受日期:2025-08-20 发布日期:2025-09-05 出版日期:2026-05-10
通讯作者: 胡婕
作者简介:徐彤（2001—），女，湖北武汉人，硕士研究生，主要研究方向：自然语言处理
张龑（1974—），男，湖北宜昌人，教授，博士，CCF会员，主要研究方向：软件工程、信息安全。
基金资助:
国家自然科学基金资助项目(61977021)

Continual few-shot event detection model based on hierarchical adaptive fusion mechanism and category boundary distillation

Jie HU¹^,²^,³(), Tong XU¹, Yan ZHANG¹^,²^,³

^1.School of Computer Science，Hubei University，Wuhan Hubei 430062，China
^2.Hubei Key Laboratory of Big Data Intelligent Analysis and Application （Hubei University），Wuhan Hubei 430062，China
^3.Key Laboratory of Intelligent Sensing System and Security，Ministry of Education （Hubei University），Wuhan Hubei 430062，China

Received:2025-05-28 Revised:2025-08-13 Accepted:2025-08-20 Online:2025-09-05 Published:2026-05-10
Contact: Jie HU
About author:XU Tong， born in 2001， M. S. candidate. Her research interests include natural language processing.
ZHANG Yan， born in 1974， Ph. D.， professor. His research interests include software engineering， information security.
Supported by:
National Natural Science Foundation of China(61977021)

摘要/Abstract

摘要：

针对持续少样本事件检测（CFED）任务中面临的灾难性遗忘与小样本泛化难题，提出一种基于层次化自适应融合机制与类别边界蒸馏的CFED模型。首先，引入特征重构，结合全局平均池化与可学习映射，增强文本表征的结构建模能力并优化其特征分布；其次，设计层次化自适应融合机制，动态整合预训练模型浅层、中层与深层特征，引入高斯扰动增强特征鲁棒性，并通过自注意力机制实现跨层次特征的自适应加权融合；最后，提出类别边界蒸馏策略，利用KL（Kullback-Leibler）散度对齐新旧任务类别分布，结合余弦相似度优化决策边界特征，缓解知识遗忘。在数据集MAVEN和ACE2005上与9个基线模型以及大语言模型GPT-3.5-Turbo进行实验对比，结果表明所提模型在MAVEN上4?way 5-shot和4-way 10-shot的5个子任务平均F1值比次优模型HANet（Hierarchical Augmentation Networks）分别提升了2.92和1.80个百分点；在ACE2005上2-way 5-shot和2-way 10-shot的5个子任务平均F1值比次优模型HANet和Combined Retrain分别提升了1.83和2.00个百分点；相较于GPT-3.5-Turbo，所提模型在MAVEN上2-way 1-shot和2-way 2-shot的平均F1值分别提升了3.47和8.77个百分点，在ACE2005上分别提升了4.37和2.39个百分点，验证了该模型性能更优。

关键词: 持续少样本事件检测, 层次化自适应融合机制, 特征重构, 类别边界蒸馏

Abstract:

To address the challenges of catastrophic forgetting and limited generalization in Continual Few-shot Event Detection （CFED）， a new CFED model based on hierarchical adaptive fusion mechanism and category boundary distillation was proposed. Firstly， feature reconstruction was introduced by combining global average pooling with a learnable mapping to enhance the structural modeling of text representations and optimize feature distribution. Secondly， a hierarchical adaptive fusion mechanism was designed to dynamically integrate shallow， intermediate， and deep features from the pretrained model. Gaussian perturbation was introduced to improve feature robustness， and a self-attention mechanism was employed to achieve adaptive cross-layer feature weighted fusion. Finally， a category-boundary distillation strategy was proposed， which aligned the class distributions of old and new tasks using KL （Kullback-Leibler） divergence and refined the decision boundary features via cosine similarity， effectively mitigating knowledge forgetting. Experimental comparisons with 9 baseline models and the large language model GPT-3.5-Turbo were conducted on the MAVEN and ACE2005 datasets. On MAVEN， the proposed model achieved average F1 value improvements of 2.92 and 1.80 percentage points over the suboptimal model HANet （Hierarchical Augmentation Networks） across 5 subtasks under the 4-way 5-shot and 4-way 10-shot settings， respectively； on ACE2005， it outperformed the suboptimal models HANet and Combined Retrain by 1.83 and 2.00 percentage points across 5 subtasks under the 2-way 5-shot and 2-way 10-shot settings， respectively. Compared to GPT-3.5-Turbo， the proposed model achieved average F1 score improvements of 3.47 and 8.77 percentage points on MAVEN， and 4.47 and 2.39 percentage points on ACE2005 under 2-way 1-shot and 2-way 2-shot settings， respectively. The results demonstrate the superior performance of the proposed model.

Key words: Continual Few-shot Event Detection (CFED), hierarchical adaptive fusion mechanism, feature reconstruction, category boundary distillation

中图分类号:

TP391.1

胡婕, 徐彤, 张龑. 基于层次化自适应融合机制和类别边界蒸馏的持续少样本事件检测模型[J]. 计算机应用, 2026, 46(5): 1450-1459.

Jie HU, Tong XU, Yan ZHANG. Continual few-shot event detection model based on hierarchical adaptive fusion mechanism and category boundary distillation[J]. Journal of Computer Applications, 2026, 46(5): 1450-1459.

图/表 12

图1 本文模型结构

Fig. 1 Structure of proposed model

表1 数据集属性及数据划分

Tab. 1 Dataset attributes and data partitioning

数据集划分	MAVEN		ACE2005
数据集划分	文档数	事件提及数	文档数	事件提及数
训练集	2 498	66 812	501	4 088
验证集	415	11 181	41	433
测试集	710	18 904	55	790

表2 事件子集划分及信息统计

Tab. 2 Event subset partitioning and information statistics

事件子集	MAVEN		ACE2005
事件子集	类别数	事件提及数	类别数	事件提及数
A	33	12 783	9	584
B	30	12 259	6	840
C	39	14 268	5	1 335
D	35	13 209	5	717
E	31	14 293	8	612

表3 各模型在MAVEN数据集上的评价指标对比 ( %)

Tab. 3 Evaluation index comparison among different models on MAVEN dataset

模型	4-way 5-shot						4-way 10-shot
	F₁					$F 1 ¯$	F₁					$F 1 ¯$
	A	B	C	D	E	$F 1 ¯$	A	B	C	D	E	$F 1 ¯$
Fine-tune	40.43	33.17	17.50	19.72	21.01	26.37	40.43	38.18	20.46	20.35	23.57	28.60
Combined Retrain	40.43	42.10	39.61	43.03	47.43	42.52	40.43	44.27	44.76	48.28	53.66	46.28
EWC	40.43	34.29	17.40	18.61	20.43	26.23	40.43	36.42	19.69	20.02	23.72	28.06
LwF	40.43	37.27	26.69	24.70	30.54	31.93	40.43	41.09	31.89	30.57	34.43	35.68
iCaRL	35.82	37.16	33.74	35.54	35.98	35.65	35.82	42.43	37.45	40.11	41.04	39.37
KCN	40.43	48.38	41.99	41.32	40.29	42.48	40.43	51.15	45.22	44.31	44.47	45.12
KT	41.04	40.19	35.21	32.69	33.77	36.58	41.04	44.39	40.00	39.42	37.87	40.54
EMP	40.17	30.95	31.21	22.90	22.25	29.50	40.17	32.33	32.95	26.68	28.16	32.06
HANet	41.91	51.39	43.21	43.53	43.89	44.79	41.91	53.17	46.71	46.36	48.12	47.25
本文模型	44.75	51.69	46.27	46.97	48.86	47.71	44.75	53.40	48.33	48.29	50.50	49.05

表3 各模型在MAVEN数据集上的评价指标对比 ( %)

Tab. 3 Evaluation index comparison among different models on MAVEN dataset

模型	4-way 5-shot						4-way 10-shot
	F₁					$F 1 ¯$	F₁					$F 1 ¯$
	A	B	C	D	E	$F 1 ¯$	A	B	C	D	E	$F 1 ¯$
Fine-tune	40.43	33.17	17.50	19.72	21.01	26.37	40.43	38.18	20.46	20.35	23.57	28.60
Combined Retrain	40.43	42.10	39.61	43.03	47.43	42.52	40.43	44.27	44.76	48.28	53.66	46.28
EWC	40.43	34.29	17.40	18.61	20.43	26.23	40.43	36.42	19.69	20.02	23.72	28.06
LwF	40.43	37.27	26.69	24.70	30.54	31.93	40.43	41.09	31.89	30.57	34.43	35.68
iCaRL	35.82	37.16	33.74	35.54	35.98	35.65	35.82	42.43	37.45	40.11	41.04	39.37
KCN	40.43	48.38	41.99	41.32	40.29	42.48	40.43	51.15	45.22	44.31	44.47	45.12
KT	41.04	40.19	35.21	32.69	33.77	36.58	41.04	44.39	40.00	39.42	37.87	40.54
EMP	40.17	30.95	31.21	22.90	22.25	29.50	40.17	32.33	32.95	26.68	28.16	32.06
HANet	41.91	51.39	43.21	43.53	43.89	44.79	41.91	53.17	46.71	46.36	48.12	47.25
本文模型	44.75	51.69	46.27	46.97	48.86	47.71	44.75	53.40	48.33	48.29	50.50	49.05

表4 各模型在ACE2005数据集上的评价指标对比 ( %)

Tab. 4 Evaluation index comparison among different models on ACE2005 dataset

模型	2-way 5-shot						2-way 10-shot
	F₁					$F 1 ¯$	F₁					$F 1 ¯$
	A	B	C	D	E	$F 1 ¯$	A	B	C	D	E	$F 1 ¯$
Fine-tune	60.86	52.09	46.37	26.64	23.15	41.82	60.86	48.17	49.55	23.29	24.66	41.31
Combined Retrain	60.86	62.45	52.21	52.20	58.36	57.22	60.86	63.39	63.75	61.23	64.25	62.70
EWC	60.86	49.30	45.41	27.14	22.36	41.01	60.86	47.58	51.15	23.82	21.79	41.04
LwF	60.86	47.31	38.91	23.31	28.40	39.76	60.86	46.98	50.77	33.48	29.69	44.36
iCaRL	50.85	52.21	37.39	31.33	28.85	40.13	50.85	52.06	42.45	32.89	34.70	52.59
KCN	60.86	56.38	47.56	38.62	37.05	48.09	60.86	59.41	57.39	46.48	44.30	53.69
KT	53.16	42.55	33.93	38.48	31.27	39.88	53.16	59.12	50.02	49.02	28.54	47.97
EMP	54.78	40.49	24.32	27.15	22.53	33.85	54.78	37.28	19.60	34.69	24.19	34.11
HANet	61.16	63.07	57.50	53.21	54.31	57.85	61.16	66.84	64.68	58.02	54.37	61.01
本文模型	62.68	60.14	63.64	54.68	57.26	59.68	62.68	66.76	65.51	62.84	65.71	64.70

表4 各模型在ACE2005数据集上的评价指标对比 ( %)

Tab. 4 Evaluation index comparison among different models on ACE2005 dataset

模型	2-way 5-shot						2-way 10-shot
	F₁					$F 1 ¯$	F₁					$F 1 ¯$
	A	B	C	D	E	$F 1 ¯$	A	B	C	D	E	$F 1 ¯$
Fine-tune	60.86	52.09	46.37	26.64	23.15	41.82	60.86	48.17	49.55	23.29	24.66	41.31
Combined Retrain	60.86	62.45	52.21	52.20	58.36	57.22	60.86	63.39	63.75	61.23	64.25	62.70
EWC	60.86	49.30	45.41	27.14	22.36	41.01	60.86	47.58	51.15	23.82	21.79	41.04
LwF	60.86	47.31	38.91	23.31	28.40	39.76	60.86	46.98	50.77	33.48	29.69	44.36
iCaRL	50.85	52.21	37.39	31.33	28.85	40.13	50.85	52.06	42.45	32.89	34.70	52.59
KCN	60.86	56.38	47.56	38.62	37.05	48.09	60.86	59.41	57.39	46.48	44.30	53.69
KT	53.16	42.55	33.93	38.48	31.27	39.88	53.16	59.12	50.02	49.02	28.54	47.97
EMP	54.78	40.49	24.32	27.15	22.53	33.85	54.78	37.28	19.60	34.69	24.19	34.11
HANet	61.16	63.07	57.50	53.21	54.31	57.85	61.16	66.84	64.68	58.02	54.37	61.01
本文模型	62.68	60.14	63.64	54.68	57.26	59.68	62.68	66.76	65.51	62.84	65.71	64.70

表5 MAVEN数据集上的消融实验结果 ( %)

Tab. 5 Ablation experimental results on MAVEN dataset

模型	4-way 5-shot						4-way 10-shot
	F₁					$F 1 ¯$	F₁					$F 1 ¯$
	A	B	C	D	E	$F 1 ¯$	A	B	C	D	E	$F 1 ¯$
本文模型	44.75	51.69	46.27	46.97	48.86	47.71	44.75	53.40	48.33	48.29	50.50	49.05
-边界感知蒸馏	44.75	51.65	46.06	46.02	47.35	47.17	44.75	53.36	47.95	47.35	50.24	48.73
-类别分布蒸馏	44.75	51.65	46.34	46.37	48.44	47.51	44.75	53.40	48.13	47.72	50.36	48.87
-类别边界蒸馏	44.75	51.62	45.81	45.89	47.39	47.09	44.75	53.33	47.88	47.32	50.11	48.68
-特征重构	42.08	51.63	45.69	46.87	47.76	46.81	42.08	53.12	48.21	47.91	50.24	48.31
-上下文增强策略	44.72	51.69	46.13	46.24	47.45	47.25	44.72	53.32	47.87	47.35	48.21	48.29
-浅层和中层特征	44.52	51.60	45.76	45.28	45.85	46.60	44.48	53.17	47.28	45.24	46.92	47.42
-层次化自适应融合机制	44.46	51.66	44.57	44.91	45.11	46.14	44.47	53.17	46.85	44.85	46.39	47.15

表5 MAVEN数据集上的消融实验结果 ( %)

Tab. 5 Ablation experimental results on MAVEN dataset

模型	4-way 5-shot						4-way 10-shot
	F₁					$F 1 ¯$	F₁					$F 1 ¯$
	A	B	C	D	E	$F 1 ¯$	A	B	C	D	E	$F 1 ¯$
本文模型	44.75	51.69	46.27	46.97	48.86	47.71	44.75	53.40	48.33	48.29	50.50	49.05
-边界感知蒸馏	44.75	51.65	46.06	46.02	47.35	47.17	44.75	53.36	47.95	47.35	50.24	48.73
-类别分布蒸馏	44.75	51.65	46.34	46.37	48.44	47.51	44.75	53.40	48.13	47.72	50.36	48.87
-类别边界蒸馏	44.75	51.62	45.81	45.89	47.39	47.09	44.75	53.33	47.88	47.32	50.11	48.68
-特征重构	42.08	51.63	45.69	46.87	47.76	46.81	42.08	53.12	48.21	47.91	50.24	48.31
-上下文增强策略	44.72	51.69	46.13	46.24	47.45	47.25	44.72	53.32	47.87	47.35	48.21	48.29
-浅层和中层特征	44.52	51.60	45.76	45.28	45.85	46.60	44.48	53.17	47.28	45.24	46.92	47.42
-层次化自适应融合机制	44.46	51.66	44.57	44.91	45.11	46.14	44.47	53.17	46.85	44.85	46.39	47.15

表6 ACE2005数据集上的消融实验结果 ( %)

Tab. 6 Ablation experimental results on ACE2005 dataset

模型	2-way 5-shot						2-way 10-shot
	F₁					$F 1 ¯$	F₁					$F 1 ¯$
	A	B	C	D	E	$F 1 ¯$	A	B	C	D	E	$F 1 ¯$
本文模型	62.68	60.14	63.64	54.68	57.26	59.68	62.68	66.76	65.51	62.84	65.71	64.70
-边界感知蒸馏	62.68	60.07	61.55	53.89	56.34	58.91	62.68	66.48	64.45	61.37	65.21	64.04
-类别分布蒸馏	62.68	59.14	62.73	54.21	56.75	59.10	62.68	66.65	65.14	62.74	65.83	64.61
-类别边界蒸馏	62.68	60.02	61.23	53.44	56.12	58.70	62.68	66.45	64.33	58.79	64.03	63.26
-特征重构	61.13	59.98	62.34	54.33	56.67	58.89	60.75	66.57	64.69	60.38	63.61	63.20
-上下文增强策略	61.24	60.04	61.65	54.52	56.17	58.72	62.14	66.75	65.23	62.74	64.19	64.21
-浅层和中层特征	60.45	58.77	57.24	54.87	55.54	57.37	62.02	66.58	65.43	61.77	63.25	63.41
-层次化自适应融合机制	59.21	58.63	56.53	54.61	55.29	56.85	61.55	66.62	64.35	60.13	60.65	62.66

表6 ACE2005数据集上的消融实验结果 ( %)

Tab. 6 Ablation experimental results on ACE2005 dataset

模型	2-way 5-shot						2-way 10-shot
	F₁					$F 1 ¯$	F₁					$F 1 ¯$
	A	B	C	D	E	$F 1 ¯$	A	B	C	D	E	$F 1 ¯$
本文模型	62.68	60.14	63.64	54.68	57.26	59.68	62.68	66.76	65.51	62.84	65.71	64.70
-边界感知蒸馏	62.68	60.07	61.55	53.89	56.34	58.91	62.68	66.48	64.45	61.37	65.21	64.04
-类别分布蒸馏	62.68	59.14	62.73	54.21	56.75	59.10	62.68	66.65	65.14	62.74	65.83	64.61
-类别边界蒸馏	62.68	60.02	61.23	53.44	56.12	58.70	62.68	66.45	64.33	58.79	64.03	63.26
-特征重构	61.13	59.98	62.34	54.33	56.67	58.89	60.75	66.57	64.69	60.38	63.61	63.20
-上下文增强策略	61.24	60.04	61.65	54.52	56.17	58.72	62.14	66.75	65.23	62.74	64.19	64.21
-浅层和中层特征	60.45	58.77	57.24	54.87	55.54	57.37	62.02	66.58	65.43	61.77	63.25	63.41
-层次化自适应融合机制	59.21	58.63	56.53	54.61	55.29	56.85	61.55	66.62	64.35	60.13	60.65	62.66

图2 不同蒸馏损失权重参数的性能对比

Fig. 2 Performance comparison with different distillation loss weight coefficients

图3 不同神经元数的性能对比

Fig. 3 Performance comparison with different numbers of neurons

表7 不同参数的F1ˉ对比

Tab.7 Comparison of F1ˉ with different parameters

参数		$F 1 ¯$ /%
		MAVEN		ACE2005
		4-way 5-shot	4-way 10-shot	2-way 5-shot	2-way 10-shot
正则化权重 $λ$	0.000 00	46.67	47.47	58.50	61.65
	0.000 01	47.71	49.05	59.68	64.70
	0.001 00	46.24	47.08	58.22	62.43
温度参数 $T$	1	47.23	48.83	58.36	63.44
	3	47.71	49.05	59.68	64.70
	4	47.57	48.19	58.77	64.31
边界感知蒸馏超参数 $α$	0.1	47.23	48.74	59.56	63.82
	0.5	47.71	49.05	59.68	64.70
	0.9	47.45	49.31	59.54	64.49

表7 不同参数的F1ˉ对比

Tab.7 Comparison of F1ˉ with different parameters

参数		$F 1 ¯$ /%
		MAVEN		ACE2005
		4-way 5-shot	4-way 10-shot	2-way 5-shot	2-way 10-shot
正则化权重 $λ$	0.000 00	46.67	47.47	58.50	61.65
	0.000 01	47.71	49.05	59.68	64.70
	0.001 00	46.24	47.08	58.22	62.43
温度参数 $T$	1	47.23	48.83	58.36	63.44
	3	47.71	49.05	59.68	64.70
	4	47.57	48.19	58.77	64.31
边界感知蒸馏超参数 $α$	0.1	47.23	48.74	59.56	63.82
	0.5	47.71	49.05	59.68	64.70
	0.9	47.45	49.31	59.54	64.49

图4 不同特征融合策略的性能对比

Fig. 4 Performance comparison with different feature fusion strategies

表8 在MAVEN和ACE2005数据集上与GPT-3.5-Turbo的F1ˉ对比 ( %)

Tab. 8 Comparison of F1ˉ with GPT-3.5-Turbo on MAVEN and ACE2005 datasets

数据集	模型	2-way 1-shot						2-way 2-shot
		F₁					$F 1 ¯$	F₁					$F 1 ¯$
		A	B	C	D	E	$F 1 ¯$	A	B	C	D	E	$F 1 ¯$
MAVEN	本文模型	67.23	46.36	38.17	44.21	43.65	47.92	67.69	55.92	51.27	53.11	51.98	55.99
MAVEN	GPT-3.5-Turbo	54.22	55.25	41.60	37.88	33.31	44.45	57.00	58.51	43.64	40.39	36.56	47.22
ACE2005	本文模型	61.34	52.07	41.86	42.58	35.66	46.70	61.31	57.84	42.72	43.49	45.22	50.12
ACE2005	GPT-3.5-Turbo	42.20	50.29	40.51	43.46	35.21	42.33	56.36	49.72	45.16	44.44	42.96	47.73

表8 在MAVEN和ACE2005数据集上与GPT-3.5-Turbo的F1ˉ对比 ( %)

Tab. 8 Comparison of F1ˉ with GPT-3.5-Turbo on MAVEN and ACE2005 datasets

数据集	模型	2-way 1-shot						2-way 2-shot
		F₁					$F 1 ¯$	F₁					$F 1 ¯$
		A	B	C	D	E	$F 1 ¯$	A	B	C	D	E	$F 1 ¯$
MAVEN	本文模型	67.23	46.36	38.17	44.21	43.65	47.92	67.69	55.92	51.27	53.11	51.98	55.99
MAVEN	GPT-3.5-Turbo	54.22	55.25	41.60	37.88	33.31	44.45	57.00	58.51	43.64	40.39	36.56	47.22
ACE2005	本文模型	61.34	52.07	41.86	42.58	35.66	46.70	61.31	57.84	42.72	43.49	45.22	50.12
ACE2005	GPT-3.5-Turbo	42.20	50.29	40.51	43.46	35.21	42.33	56.36	49.72	45.16	44.44	42.96	47.73

参考文献 20

[1]	朱艺娜，曹阳，钟靖越，等.事件抽取技术研究综述［J］.计算机科学，2022，49（12）：264-273.
	ZHU Y N， CAO Y， ZHONG J Y， et al. Survey on event extraction technology［J］. Computer Science， 2022， 49（12）： 264-273.
[2]	LEE S W， KIM J H， JUN J， et al. Overcoming catastrophic forgetting by incremental moment matching［C］// Proceedings of the 31st International Conference on Neural Information Processing Systems. Red Hook： Curran Associates Inc.， 2017： 4655-4665.
[3]	SHI G， CHEN J， ZHANG W， et al. Overcoming catastrophic forgetting in incremental few-shot learning by finding flat minima［C］// Proceedings of the 35th International Conference on Neural Information Processing Systems. Red Hook： Curran Associates Inc.， 2021： 6747-6761.
[4]	ZHANG C， CAO P， CHEN Y， et al. Continual few-shot event detection via hierarchical augmentation networks［C］// Proceedings of the 2024 Joint International Conference on Computational Linguistics， Language Resources and Evaluation. ［S.l.］： ELRA and ICCL， 2024： 3868-3880.
[5]	TAO X， HONG X， CHANG X， et al. Few-shot class-incremental learning［C］// Proceedings of the 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway： IEEE， 2020： 12180-12189.
[6]	KEMKER R， McCLURE M， ABITINO A， et al. Measuring catastrophic forgetting in neural networks［C］// Proceedings of the 32nd AAAI Conference on Artificial Intelligence. Palo Alto： AAAI Press， 2018： 3390-3398.
[7]	CHEN J， LIN H， HAN X， et al. Honey or poison？ Solving the trigger curse in few-shot event detection via causal intervention［C］// Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. Stroudsburg： ACL， 2021： 8078-8088.
[8]	LI X， ZHOU Y， WU T， et al. Learn to grow： a continual structure learning framework for overcoming catastrophic forgetting［C］// Proceedings of the 36th International Conference on Machine Learning. New York： JMLR.org， 2019： 3925-3934.
[9]	KIRKPATRICK J， PASCANU R， RABINOWITZ N， et al. Overcoming catastrophic forgetting in neural networks［J］. Proceedings of the National Academy of Sciences of the United States of America， 2017， 114（13）： 3521-3526.
[10]	LI Z， HOIEM D. Learning without forgetting［J］. IEEE Transactions on Pattern Analysis and Machine Intelligence， 2018， 40（12）： 2935-2947.
[11]	REBUFFI S A， KOLESNIKOV A， SPERL G， et al. iCaRL： incremental classifier and representation learning［C］// Proceedings of the 2017 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway： IEEE， 2017： 5533-5542.
[12]	CAO P， CHEN Y， ZHAO J， et al. Incremental event detection via knowledge consolidation networks［C］// Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing. Stroudsburg： ACL， 2020： 707-717.
[13]	YU P， JI H， NATARAJAN P. Lifelong event detection with knowledge transfer［C］// Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. Stroudsburg： ACL， 2021： 5278-5290.
[14]	LIU M， CHANG S， HUANG L. Incremental prompting： episodic memory prompt for lifelong event detection［C］// Proceedings of the 29th International Conference on Computational Linguistics. ［S.l.］： International Committee on Computational Linguistics， 2022： 2157-2165.
[15]	王翔，胡学钢.高维小样本分类问题中特征选择研究综述［J］.计算机应用，2017，37（9）：2433-2438.
	WANG X， HU X G. Overview on feature selection in high-dimensional and small-sample-size classification［J］. Journal of Computer Applications， 2017， 37（9）： 2433-2438.
[16]	LAI V D， DERNONCOURT F， NGUYEN T H. Extensively matching for few-shot learning event detection［C］// Proceedings of the 1st Joint Workshop on Narrative Understanding， Storylines， and Events. Stroudsburg： ACL， 2020： 38-45.
[17]	DENG S， ZHANG N， KANG J， et al. Meta-learning with dynamic-memory-based prototypical network for few-shot event detection［C］// Proceedings of the 13th International Conference on Web Search and Data Mining. New York： ACM， 2020： 151-159.
[18]	ZHANG R， WEI W， MAO X L， et al. HCL-TAT： a hybrid contrastive learning method for few-shot event detection with task-adaptive threshold［C］// Findings of the Association for Computational Linguistics： EMNLP 2022. Stroudsburg： ACL， 2022： 1808-1819.
[19]	ZHAO K， JIN X， BAI L， et al. Knowledge-enhanced self-supervised prototypical network for few-shot event detection［C］// Findings of the Association for Computational Linguistics： EMNLP 2022. Stroudsburg： ACL， 2022： 6266-6275.
[20]	KIM T， OH J， KIM N Y， et al. Comparing Kullback-Leibler divergence and mean squared error loss in knowledge distillation［C］// Proceedings of the 30th International Joint Conference on Artificial Intelligence. California： ijcai.org， 2021： 2628-2635.

基于层次化自适应融合机制和类别边界蒸馏的持续少样本事件检测模型

Continual few-shot event detection model based on hierarchical adaptive fusion mechanism and category boundary distillation

RichHTML

PDF

可视化

摘要/Abstract

引用本文

使用本文

图/表 12

参考文献 20

相关文章 1

编辑推荐

Metrics