组合原则表明句子的语义由其构成成分的语义按照一定规则组合而成,由此基于句法结构的语义组合计算一直是一个重要的探索方向,其中釆用树结构的组合计算方法最具有代表性(Tai et al.,2015).但是该方法难以应用于大规模数据处理,主要问题是其语义组合的顺序依赖于具体树的结构,无法实现并行处理.本文提出一种基于图的依存句法分析和语义组合计算的联合框架,并借助复述识别任务训练语义组合模型和句法分析模型.一方面图模型可以在训练和预测阶段采用并行处理,极大缩短计算时间;另一方面联合句法分析的语义组合框架不必依赖外部句法分析器,同时两个任务的联合学习可使语义表示同时学习句法结构和语义的上下文信息.我们在公开汉语复述识别数据集LCQMC(Liu et al.,2018)上进行评测,实验结果显示准确率接近树结构组合方法,达到79.54%,而预测速度提升高达30倍.
基于图神经网络的汉语依存分析和语义组合计算联合模型
The semantics of a sentence is composed of the meaning of its constituent components and the combination method.Therefore,syntax-based semantic composition has always been an important research direction in NLP.The semantic composition method using tree structure has became the most representative method(Tai et al.,2015).However,such methods are difficult to be applied to large-scale data.The main problem is that the order of its semantic composition depends on the structure of the specific tree,and parallel computation cannot be supported.In this paper,we present a joint framework for graph-based dependency parsing and semantic composition.The model does not need to rely on an external syntax parser for providing structural information,and the semantic composition method based on graph neural network can support parallel computation,which greatly reduces the computation time.Moreover,the joint learning of two tasks enables the model to learn the syntactic structure and semantic contextual information.Experimental results on LCQMC(Liu et al.,2018)dataset show that the accuracy is close to the tree-based semantics composition method,reaching 79.54%,and the prediction speed is increased by up to 30 times.
句法分析语义组合图神经网络复述识别
汪凯、刘明童、陈圆梦、张玉洁、徐金安、陈钰枫
展开 >
北京交通大学计算机与信息技术学院,北京100044
句法分析 语义组合 图神经网络 复述识别
Chinese National Conference on Computational Linguistic
Haikou(CN)
19th Chinese National Conference on Computational Linguistic
NETL
