Development of a Coding Supporting System for Responses to Open-ended Questions with Applying Natural Language Processing Technique

主要的研究成果是，我们已经基本完成了NANACO系统，该系统支持社会调查中的职业编码。在职业编码中，对受访者职业的开放式问题的回答必须归入近200个职业代码中的一个进行统计分析，这很耗时，有时还会错误分类。首先，NANACO系统通过机器学习(支持向量机)和手工生成的规则自动分类的5个职业代码作为候选代码显示在编码器的监视器上。其次，NANACO系统还提供关于受访者的信息，如年龄、性别和教育背景，这有助于编码员了解答复。第三，NANACO系统具有多种功能，例如一个“字典”的查看窗口，它定义了每个职业代码，以便编码者轻松工作。NANACO系统将包括编码员设置的职业代码在内的数据保存为逗号分隔值(CSV)文件。此外，在JGSS(日本一般社会调查)或SSM(社会分层和社会流动性)调查等社会调查中使用NANACO系统的编码员要求我们提供信心衡量标准，即排名第一的候选人对其信心决策的信心衡量标准。因此，我们提出了一种新的估计置信度的方法--类成员概率，该方法不仅使用分类器的第一类分数，而且还使用第二类分数。实验结果表明，该方法比以往的方法更有效。在以后的工作中，我们将在NANACO系统中设置类成员概率。显然，NANACO系统可以很容易地扩展到对具有预定义类别的不限成员名额问题的任何答复。

The main research result is that we have almost completed for coders the NANACO system, which supports an occupation coding in social surveys. In the occupation coding, responses to open-ended questions about respondents' occupation have to been classified into one of nearly 200 occupation codes for statistical analysis, which is time consuming and sometimes misclassified. First the NANACO system presents 5 occupation codes automatically classified with machine learning (support vector machines) and hand-crafted rules as candidates on coder's monitor. Second the NANACO system also presents information about respondents such as age, sex and educational background which help for coders to understand responses. Third the NANACO system has various functions such as a view window for a "dictionary" which defines each occupation code for coder to work easily. Finely the NANACO system saves data including occupation codes set by coders as a Comma Separated Values (CSV) file. Furthermore coders who used the NANACO system in social surveys such as JGSS (Japanese General Social Surveys) or SSM (Social Stratification and social Mobility) survey, asked us to supply a measure of confidence a measure of confidence for the first-ranked candidate of their confidence decisions. Therefore, we proposed a new method to estimate a measure of confidence, class membership probability, by using not only the first-class score but also the second-class score from a classifier. We showed that the proposed method was more effective than previous methods. In future work, we will set class membership probabilities into the NANACO system. It is clear that the NANACO system can be easily extended for any responses to open-ended questions which have predefined categories.