【作者】 韩旭；

【导师】 王博； 赵瑞明；

【作者基本信息】 天津大学 ， 工程（专业学位）， 2021， 硕士

【摘要】 在心理学领域,偏见指用户针对特定事物的倾向性看法。准确识别和分析个人或群体的偏见不仅具有心理学领域的科学意义,同时也是人工智能领域包括个性化推荐、智能对话、社会计算等下游研究的重要基础。具体而言,偏见可以分为外显偏见和内隐偏见。其中,外显偏见指人们在公开场合所直接表达的偏见,内隐偏见指人们内心中真实的偏见。心理学研究显示,个人或群体的外显偏见与内隐偏见常常可以存在差异,这为偏见的识别与分析,尤其是内隐偏见的识别与分析带来了巨大挑战。在心理学领域,内隐联想测验(IAT测试)是测量内隐偏见的权威方法。但是内隐联想测验只能够在实验室中,通过被试者的主动配合来完成测试,这极大的限制了该方法在大规模人群中的使用。近年来,基于人工智能方法,尤其是自然语言处理技术,从个人或群体的言论信息中来发现人们的偏见成为包括人工智能伦理在内的若干研究领域共同关注的热点问题。当前已有的方法主要依靠词汇级的表示学习和词汇的语义表示间的语义距离来实尝试模拟心理学领域的内隐联想测验,并取得了一定的进展。但是,当前基于自然语言处理技术的偏见识别与分析方法仍然存在一些不足。首先,现有的方法实质上并没有区分外显和内隐偏见。本文揭示,现有基于自然语言处理的偏见识别方法所识别出的偏见是外显和内隐偏见的混合结果。其次,偏见表达主要存在于句子级,而当前的方法主要应用了词汇级别的表示学习技术,而没有能够挖掘句子级语言表示学习在外显及内隐偏见识别语分析中的潜力。虽然部分方法探索了包括BERT在内的主流句子级偏见识别的方法,但其本质上是围绕句子中的个别词汇来建构句子语义,而没有探索同时包含了偏见对象和偏见属性的句子,即包含完整的偏见表达的句子的表示学习对于偏见分析的作用。为了解决上述问题,本文的工作主要体现在以下四个方面:(1)心理学内隐测量方法与自然语言处理技术的关联:本文首先尝试从理论层面,建立心理学内隐联想测验要素与语言要素之间的关联,通过提出概念词和属性词及其在表达中的组合与内隐联想测验中的刺激物、属性,以及语义关联与内隐联想距离的关联,为基于语言的外显及内隐偏见测试提供理论基础。(2)基于词汇级表示学习的外显及内隐偏见区分测量:本文在当前基于词汇级表示学习的偏见分析方法(WEAT)的基础上,基于前述关联分析,提出一种通过合理划分语料来分别自动测量外显及内隐偏见的词汇级测量方法。(3)基于句子级表示学习的外显及内隐偏见区分测量:本文构建由概念词和属性词通过不同关系组成的表达语句,并利用前沿的句子级表示学习方法建模表达的语义,进而实现句子级的外显及内隐偏见测量方法。(4)基于深度学习的外显及内隐偏见测量方法的验证:在实验中,本文通过将自动测量结果与心理学测量结果进行定性的对比,来验证本文方法的有效性。同时,本文通过对所测量的外显及内隐偏见的时序演化规律进行观察,得出新的具有一定科学意义的分析结论,即内隐偏见较之外显偏见具有更高的时序稳定性。更多还原

【Abstract】 In the field of psychology,prejudice refers to the user’s tendency to view specific things.Accurately identifying and analyzing the biases of individuals or groups is not only of scientific significance in the field of psychology,but also an important basis for downstream research in the field of artificial intelligence,including personalized recommendation,intelligent dialogue,and social computing.Specifically,prejudice can be divided into explicit bias and implicit bias.Among them,explicit prejudice refers to the prejudice that people express directly in public,and implicit prejudice refers to the real prejudice in people’s hearts.Psychological research shows that there can often be differences between the explicit and implicit biases of individuals or groups,which brings great challenges to the identification and analysis of bias,especially the identification and analysis of implicit bias.In the field of psychology,the implicit association test(IAT)is the authoritative method for measuring implicit bias.However,the implicit association test can only be completed in the laboratory through the active cooperation of the subjects,which greatly limits the use of this method in large-scale populations.In recent years,based on artificial intelligence methods,especially natural language processing technology,discovering people’s prejudices from the speech information of individuals or groups has become a hot issue in several research fields,including artificial intelligence ethics.The existing methods mainly rely on the semantic distance between vocabulary-level representation learning and the semantic representation of the vocabulary to try to simulate the implicit association test in the field of psychology,and have made certain progress.However,the current bias recognition and analysis methods based on natural language processing technology still have some shortcomings.First of all,the existing methods essentially do not distinguish between explicit and implicit biases.This article reveals that the biases identified by the existing bias recognition methods based on natural language processing are a mixed result of explicit and implicit biases.Secondly,prejudice expression mainly exists at the sentence level,while current methods mainly apply vocabulary-level representation learning techniques,but fail to tap the potential of sentence-level language representation learning in the analysis of explicit and implicit bias recognition.Although some methods have explored sentence-level bias recognition methods including BERT,they essentially construct sentence semantics around individual words in the sentence,instead of exploring sentences that contain both the object of bias and the attributes of bias,that is,it contains The effect of the expression learning of the sentence of complete prejudice expression on the analysis of prejudice.In order to solve the above problems,the work of this article is mainly reflected in the following four aspects:(1)The relationship between psychological measurement and natural language technology: This article first attempts to establish the relationship between psychological implicit association test elements and language elements from a theoretical level,by proposing concept words and attribute words and their combination and expression in expressions.The stimuli,attributes,and the relationship between semantic association and implicit association distance in the implicit association test provide a theoretical basis for language-based explicit and implicit bias testing.(2)Distinguishing between explicit and implicit biases based on vocabulary-level representation learning: Based on the current bias analysis method based on vocabulary-level representation learning,based on the aforementioned association analysis,this paper proposes an automatic measurement by reasonably dividing the corpus Vocabulary-level measurement methods for explicit and implicit biases.(3)Distinguishing between explicit and implicit biases based on sentence-level representation learning: This paper constructs expression sentences composed of concept words and attribute words through different relationships,and uses cuttingedge sentence-level representation learning methods to model the semantics of expressions,and then achieve sentence-level The method of measuring explicit and implicit biases.(4)Verification of the Measurement Method of Explicit and Implicit Bias Based on Deep Learning: In the experiment,this paper verifies the effectiveness of this method by qualitatively comparing the automatic measurement results with the psychological measurement results.At the same time,this paper observes the time-series evolution law of the measured explicit and implicit biases,and draws a new analytical conclusion with certain scientific significance,that is,implicit biases have higher time-series stability than explicit biases.更多还原