一种基于字词结合的汉字识别上下文处理新方法

A NOVEL METHOD BASED ON INTEGRATING CHARACTERS WITH WORDS FOR CONTEXTUAL PROCESSING OF CHINESE CHARACTER RECOGNITION

根据字、词信息之间的互补性 ,提出一种字、词结合的上下文处理方法 .在单字识别的基础上 ,首先利用前向 -后向搜索算法在较大的候选集上进行基于字 bigram模型的上下文处理 ,在提高文本识别率的同时可提高候选集的效率 ;然后在较小的候选集上进行基于词 bigram模型的上下文处理 .该方法在兼顾处理速度的同时 ,可有效地提高文本识别率 .脱机手写体汉字文本 (约 6 .6万字 )识别中的实验表明 :经字 bigram模型处理 ,文本识别率由处理前的 81.5 8%提高至 94 .5 0 % ,文本前 10选累计正确率由 94 .33%提高到 98.2 5 % ;再经词 bigram模型处理 ,文本识别率进一步提高至 95 .75 % .

According to the complementarity between Chinese characters and Chinese words, a novel contextual processing method is put forward, which integrates character based language model with word based language model. On the basis of isolated character recognition, character based bigram post processing using forward backward search is first executed on big candidate sets, which improves both the recognition rate of document (RRD) and the efficiency of candidate sets (the accumulated recognition rate of the top ten candidates is greatly boosted). Then, word based bigram post processing is executed on small candidate sets to further improve the RRD. This method effectively improves the RRD while giving attention to the processing speed in the meantime. Experimental results on off line handwritten Chinese documents (about 66000 characters) demonstrate the effectiveness of the novel method: character based bigram post processing improves the RRD to 94.50% from 81.58% RRD before post processing, and the accumulated recognition rate of the top ten candidates boosts from 94.33% to 98.25%. The 95 75% RRD is obtained after word based bigram post processing.