人们在就业招聘时仍然经常使用人格测验，而人格测验的形式多为自陈量表，在找工作这类高利害测试环境中，社会称许性成为阻碍测验结果真实性的重要因素。了解影响社会称许性变化的因素，对于人格测验开发和修订时控制社会称许性，降低人们的社会称许性反应有重要意义。

本研究的目的是从人格测验题目文本角度出发，寻找对题目社会称许性产生影响的题目文本特征。过往研究显示人格测验题目的维度(如大五)、用词(如频率词)、主题(如成就)和语义对题目的社会称许性存在一定的影响，但是这些研究拘泥于单个量表，不完全涵盖大五人格框架下的人格测验题目。研究的词特征和主题特征也不够全面，针对语义的影响也没有直接进行验证。自然语言处理能提取更多的文本特征，处理文本的语义特征。因此，本研究不仅提取前人研究中的题目特征，也利用自然语言处理方法补充提取更多特征，在更多大五人格框架下的测验题目上探索哪些题目文本特征会影响题目的社会称许性。

对国际人格测试题库中部分大五框架下的测验题目进行编码和统计分析后，发现题目描述的内容是影响其社会称许性最主要的原因，与道德、情绪、交往、能力、工作和同情的题目具有较强的社会称许性。而题目中使用的频率词、否定词也对社会称许性产生一定影响，且更抽象的题目的称许性更强。统计大五维度下各类题目占比发现，大五维度与称许性的相关可以由维度下各类题目的分布解释，大五维度的称许性差异本质还是各维度组成题目的称许性差异。

“语言查询与字数统计”软件(Linguistic Inquiry and Word Count，LIWC)能够统计题目中使用的各类型词的百分比，对该软件输出的词类统计进行假设检验，发现情绪类、成就类、现在时态类、休闲类4类表示实际意义的词类能较好的区分称许性的强弱。长单词占比较高和题目单词数较多这些表示句子复杂性的指标也影响题目的社会称许性。

将以上特征放入决策树，通过决策树的变量选择来验证以上推论的正确性，决策树最终保留的变量基本符合上述推论的预期。最后利用双向编码表征转换器(Bidirectional Encoder Representations from Transformers，BERT)只依据题目文本对其称许性进行分类，分类正确率达到75.6%。得到了较好的分类效果，证明题目的语义特征对题目社会称许性的确存在影响。

   因此，题目中使用频率词，题目内容涉及道德、情绪、交往、能力、工作和同情，题目描述更为抽象，题目包含更多与成就、描述当前的词时，题目的社会称许性较强；题目内容为兴趣、艺术、对事物的看法时社会称许性较弱。在题目中加入程度副词对题目的社会称许性没有明显影响，而使用双重否定，能够降低对正面事物描述的题目的社会称许性。

    Personality scale is often used in employment recruitment. In job hunting or other high-stake situation, social desirability could destroy the validity of scale. Understanding the causes of social desirability is of great significance for reducing people's social desirability responses during the development and revision of personality scale.
    Therefore, the purpose of this study is to find out the textual features of personality items that influence their social desirability. Previous studies have shown that big5, frequency words, achievement and semantics have a certain influence on the social desirability of the items, but these studies are limited to a single scale. The word features and topic features of the study are not enough, and the influence of semantics is not directly verified. Natural Language Processing (NLP) can extract more text features and deal with the semantic of text. So this study not only extracted the features from previous studies, but also used NLP to supplement, and explored which text features would affect the social desirability of more personality items under the framework of the Big Five.
    After analyzing some of big5 personality items from the International Personality Item Pool (IPIP), it is found that the content of the item is the most important feature. Items of morality, emotion, social communication, ability, work and compassion have strong social desirability. The frequency words and negative words also have influence on social desirability, and if the items are abstract, their social desirability would be stronger. It is also found that the correlation between the big5 and social desirability can be explained by the distribution of different types of items.
    "Linguistic Inquiry and Word Count" (LIWC) can count the percentage of some type of words. It is found that the four types of words representing the practical meaning of emotion, achievement, words means present and leisure can better distinguish social desirability. The proportion of long words and the number of words in each item also have influence on its social desirability.
    The decision tree is used to classify social desirability. All the upper features are put into the tree. The final reserved variables of the tree conform to the expectations. Finally, “Bidirectional Encoder Representations from Transformers” (BERT) is used to classify the social desirability only according to the item, and the correct ratio is over 75%, which proves that the semantic characteristics of the item do influence its social desirability.
    Therefore, items that contain frequency words, more words describing achievements or current, items related to morality, emotion, communication, ability, work or compassion, and items that are more abstract have stronger social desirability. When items’ content is interest, art, views on things, its social desirability is weak. The inclusion of degree adverbs has no significant effect on the social desirability, while double negation can reduce the social desirability of the items describing positive things.