作者:
Gao, DuoLi, QiPeking Univ
Inst Remote Sensing & GIS Ctr Spatial Intelligent Comp Beijing 100871 Peoples R China
In this paper we present preliminary results and lessons learned in an effort to design an automatic Chinese address segmentation algorithm using self-organizing neural network (SONN). The SONN design is composed of a...
详细信息
ISBN:
(纸本)9780819469137
In this paper we present preliminary results and lessons learned in an effort to design an automatic Chinese address segmentation algorithm using self-organizing neural network (SONN). The SONN design is composed of an input layer and an output layer, fully connected with random initial weights. Unparsed addresses are transformed into activation patterns on the input layer by activating corresponding neurons simultaneously. The training process is controlled by the Hebbian learning rule and the "kwinnerstakeall" (kWTA) competing function. After the self-organization process, each unparsed address will be represented by k (multiple) winners with each winner representing a sub-pattern of the address. The goal of the algorithm is to make these sub-patterns correspond to the word segments of the addresses. Two tests were carried out using a prototype system implementing the algorithm. The results show that the current design can achieve the purpose conditionally and we will discuss the limitations of the SONN for actual applications and the possible improvements of the model will be also discussed.
暂无评论