Current Location:home > Detailed Browse

Article Detail


Submit Time: 2022-01-05
Author: 张雨馨 1 ; 刘斌 1 ; 黄思奇 1 ;
Institute: 1.西南财经大学;


[目的] 本文针对中文出版物中不同字体、不同书写系统的阅读绩效进行客观对比研究。 [方法] 具体地,将汉字渲染成其对应字形的图像,并进一步按照语序把句子中的汉字图像折叠成为三维的句子张量。对于同一段中文文本,用不同的字体或者简体、繁体会得到句子的视觉差异化的张量表达。通过进一步将得到句子张量输入到我们设计的深度语言模型,进行文本分类等任务的测试,可以客观地比较字体和书写系统对阅读绩效的影响。 [结果] 通过在两个中文文本分类数据集上的测试发现,一些特殊不常用字体相较于常用字体的机器识别准确度较低,并且常用字体中不同字体的阅读绩效也有差异。 [结论] 通过假设检验得出使用楷体和黑体的数据集在文本分类任务上的准确度存在显著性差异,楷体相比于黑体来说阅读绩效更高。简体中文和繁体中文的阅读绩效存在显著性差异。
[英文摘要] " [Objective] We study the reading performance of different fonts and writing systems that are using in Chinese publications. [Methods] Specifically, the Chinese characters in a sentence are rendered into their corresponding glyph images, then fold those images into a three-dimensional sentence tensor according to the word order. For different fonts or simplified/traditional Chinese text, we can get the corresponding representations with visual differences. By inputting the obtained sentence tensor into the proposed deep language model, we test them on text classification, which can objectively study the influence of font and writing system on reading performance. [Results] According to the experiments on two real-world Chinese text classification datasets, Toutiao and Thucnews, we found that the accuracy of text classification on some uncommon fonts is lower than that of common used fonts, and the text representation efficiency of different fonts in the common fonts is also different. [Conclusions] Through a hypothesis test, we found that there is a significant difference in the accuracy of using the data sets of regular script and bold script for text classification task, and the efficiency of regular script is higher than that of bold script. There are significant differences in reading performance between simplified and traditional writing systems. " "
Download Comment Hits:7272 Downloads:492
From: 张雨馨
Recommended references: 张雨馨,刘斌,黄思奇.(2022).利用深度学习研究中文书写系统、字体对阅读绩效的影响.[ChinaXiv:202112.00118] (Click&Copy)
Version History
[V2] 2022-01-05 11:04:46 chinaXiv:202112.00118V2 Download
[V1] 2021-10-24 14:42:57 chinaXiv:202112.00118v1(View This Version) Download
Related Paper


Current Browse

Change Subject Browse

Cross Subject Browse

  • - NO