Video events recognition is a challenging task for high-level understanding of video se- quence. At present, there are two major limitations in existing methods for events recognition. One is that no algorithms are av...Video events recognition is a challenging task for high-level understanding of video se- quence. At present, there are two major limitations in existing methods for events recognition. One is that no algorithms are available to recognize events which happen alternately. The other is that the temporal relationship between atomic actions is not fully utilized. Aiming at these problems, an algo- rithm based on an extended stochastic context-free grammar （SCFG） representation is proposed for events recognition. Events are modeled by a series of atomic actions and represented by an extended SCFG. The extended SCFG can express the hierarchical structure of the events and the temporal re- lationship between the atomic actions. In comparison with previous work, the main contributions of this paper are as follows： ① Events （include alternating events） can be recognized by an improved stochastic parsing and shortest path finding algorithm. ② The algorithm can disambiguate the detec- tion results of atomic actions by event context. Experimental results show that the proposed algo- rithm can recognize events accurately and most atomic action detection errors can be corrected sim- ultaneously.展开更多
针对现有的事件因果关系识别方法未考虑导入外部知识后产生的噪声干扰,导致事件表示歧义增加,从而影响识别效果的问题,提出了基于外部词库和超图降噪的事件因果关系识别模型(event causality identification model based on external vo...针对现有的事件因果关系识别方法未考虑导入外部知识后产生的噪声干扰,导致事件表示歧义增加,从而影响识别效果的问题,提出了基于外部词库和超图降噪的事件因果关系识别模型(event causality identification model based on external vocabulary and hypergraph denoising,EHDM)。首先,从外部词库中检索事件的背景知识来丰富事件的语义信息,并对带有背景知识的事件描述进行编码。然后,根据事件背景知识中多个关系对应的知识特征构建超图,通过超图卷积神经网络和多头注意力机制进一步处理特征,得到降噪后的事件特征表示。接着,对事件及其上下文进行编码得到基于上下文的特征表示,并与降噪后的事件特征表示一起通过门单元进行特征融合。最后,将融合的特征表示输入多层感知器得到预测值,实现因果关系识别。结果表明,EHDM在因果时间库(causal-timebank,CTB)数据集句内方面的F1分数相比关系图卷积网络(relation graph convolutional networks,RGCN)模型提高了1.5个百分点,在事件情节链(event story line,ESL)数据集句内方面的F1分数相比RGCN模型提高了2.4个百分点,跨句、总体方面的F1分数相比事件关系图变换器模型分别提高了2.1、3.0个百分点。该研究证实了EHDM能有效应用于事件因果关系识别领域。展开更多
基金Supported by the National Natural Science Foundation of China(60805028,60903146)Natural Science Foundation of Shandong Province of China (ZR2010FM027)+1 种基金SDUST Research Fund(2010KYTD101)China Postdoctoral Science Foundation(2012M521336)
文摘Video events recognition is a challenging task for high-level understanding of video se- quence. At present, there are two major limitations in existing methods for events recognition. One is that no algorithms are available to recognize events which happen alternately. The other is that the temporal relationship between atomic actions is not fully utilized. Aiming at these problems, an algo- rithm based on an extended stochastic context-free grammar （SCFG） representation is proposed for events recognition. Events are modeled by a series of atomic actions and represented by an extended SCFG. The extended SCFG can express the hierarchical structure of the events and the temporal re- lationship between the atomic actions. In comparison with previous work, the main contributions of this paper are as follows： ① Events （include alternating events） can be recognized by an improved stochastic parsing and shortest path finding algorithm. ② The algorithm can disambiguate the detec- tion results of atomic actions by event context. Experimental results show that the proposed algo- rithm can recognize events accurately and most atomic action detection errors can be corrected sim- ultaneously.
文摘针对现有的事件因果关系识别方法未考虑导入外部知识后产生的噪声干扰,导致事件表示歧义增加,从而影响识别效果的问题,提出了基于外部词库和超图降噪的事件因果关系识别模型(event causality identification model based on external vocabulary and hypergraph denoising,EHDM)。首先,从外部词库中检索事件的背景知识来丰富事件的语义信息,并对带有背景知识的事件描述进行编码。然后,根据事件背景知识中多个关系对应的知识特征构建超图,通过超图卷积神经网络和多头注意力机制进一步处理特征,得到降噪后的事件特征表示。接着,对事件及其上下文进行编码得到基于上下文的特征表示,并与降噪后的事件特征表示一起通过门单元进行特征融合。最后,将融合的特征表示输入多层感知器得到预测值,实现因果关系识别。结果表明,EHDM在因果时间库(causal-timebank,CTB)数据集句内方面的F1分数相比关系图卷积网络(relation graph convolutional networks,RGCN)模型提高了1.5个百分点,在事件情节链(event story line,ESL)数据集句内方面的F1分数相比RGCN模型提高了2.4个百分点,跨句、总体方面的F1分数相比事件关系图变换器模型分别提高了2.1、3.0个百分点。该研究证实了EHDM能有效应用于事件因果关系识别领域。
