基于上下文词频词汇量指标的新词发现方法

A NOVEL APPROACH FOR CHINESE NEW WORD IDENTIFICATION BASED ON CONTEXTUAL WORD FREQUENCY-CONTEXTUAL WORD COUNT

提出一种基于上下文词频词汇量的统计指标。该指标通过修改信息熵公式中参数的定义,即将邻接字符串在语料集中出现的次数改成邻接字符串集合的大小,克服了左右信息熵在识别新词时特征不够明显的缺点。同时提出一种递归的基于邻接关系的字符串连接方法,克服了N-gram方法采用固定滑动窗口大小的缺点。实证分析表明该新词发现方法有较高的准确率,通过选取不同的词频词汇量指标值作为阈值,能够在发现更多新词和提高发现新词的准确率方面进行灵活调整,为新词发现提供一种实用的方法。

This article presents a statistic index which is based on contextual word frequency-contextual word count ( W F -C W C ). W F -C W C , by modifying the definition of the parameters in information entropy formula, i. e ., changing the occurrence frequency of adjacent strings in corpus to the size of the adjacent strings collection, overcomes the defect of left and right information entropies being not significant in characteristics when identifying new words. Meanwhile, this paper presents a recursive and adjacent relation-based string concatenation method, which overcomes the disadvantage of the fixed sliding window size in N-gram model. Empirical analysis indicates that this new word identification method has higher accuracy. Through selecting different W F - C W C as the thresholds, it can make flexible adjustment in finding more new words or improve the accuracy of new words identification, and this provides a practical approach for new words identification.