用大规模并行网络模型同时解决中文分词和语法分析

Intergrating Chinese Natural Language Segmentation and Parsing Using a Massively Parallel Network Model

提出用大规模并行网络的自然语言处理模型(即 MPNP 模型)处理中文的方法,它将中文的分词和语法分析合为一体加以统一的解决。以便实现多层次知识的综合。这种新的形式化方法吸收了当前连结主义模型的并行的优点,即用主动的节点—带局部记忆区和过程的节点和它们之间的连结关系显式地表达知识,用节点间的信息传送实现信息的并行处理。同时又保持了在较高层次模拟智能行为的特点。由于它的特殊的控制方法,它能够将自下而上的分析和信息传送与自上而下的期望扩散模型有机地结合起来。通过从左至右的扫描,并行地生成所有可能的分析,这一并行方法为新一代计算机的自然语言处理提供了一个有用模型。本文中从实现的角度较详细地阐述了这种处理方法的基本原理、结构和控制,以及中文分析系统的建造方法。

This paper presents a massivelly parallel model for Chinese natural lauguage processing that integrates the process of segmentation and parsing into a uniform framework in order to improve the knowledge integration. In the formallism, knowledge of grammar and word structure arc represented as a massively parcllcl notwork explicitly,in which nodes are active units with local memory and house-keeping procedures. The control scheme of the model combines the advantage of parallelism of conncctionist model and marker passing algorithms. It parses sentence bottom-up with the assistence of top-down expectation propagation to improve its effcciency. It scans a sentence from left to right, explores all alternatives in parallel and generates all plausible parses as output.