An Estimation Method of the Words Tendency Based on Time-Series Variation

摘要

Recently, there are many electronic text and computers are more and more processing for them. Frequencies of the words in the texts change according to the time-series variation. Frequently, these words are considered as keywords because they have strong relationships with the subject of the texts. However, traditional document processing systems do not consider the time-series information of the words, during calculating their importance. This paper presents an estimation method of the word trend considering the time-series variation. First, we made an example, which show us the re-arrangement of the similar texts retrieval system by using traditional methods and after using the method of the word trend based on time-series variation. By using the decision tree, the proposed method of this paper classifies words into three classes: increasing, constant, and decreasing, which effect in the stability class of words. This classification is acquired by learning five attribute values of the words, such as: slope and slice of regression line, correlation coefficient, the angle between two regression straight lines, and some special nouns attributes, and then we estimate the class of the new words. These attribute values are defined in order to measure the frequency change of each word quantitatively, and we find that these attributes have efficiency on the behavior of recall and precision. Among the evaluation, we obtained the attribute values of 1,069 proper nouns extracted from 8,216 articles of CNN newspapers (1997-1999) "This data called Learning-Data", where these articles discuss about the professional baseball. By learning the attribute values to the decision tree, 696 proper nouns that extracted from 1,272 articles of CNN newspaper (2000) are classified "This data called Test-Data". According to comparing the decision tree results evaluation with human evaluation results, it is estimated that, F-measures of increasing class, constant-class, and decrease-class are 0.847,0.851, and 0.768 respectively.