Computational models of inductive reasoning and their psychologicale examination:Towards an induction-based search-engine-猫眼课题宝

归纳推理的计算模型及其心理检验：走向基于归纳的搜索引擎

DOI：

发表时间：

2005

期刊：

影响因子：

通讯作者：

中科院分区：

文献类型：

作者：

Kayo Sakamoto;Asuka Terai;M. Nakagawa

本研究的目的是提出人类归纳推理的计算模型，使用日语语言数据的统计分析，并开发一个基于归纳推理的搜索引擎。Osherson等人（1990）提供了一个基于前提和结论之间的相似性以及对包括前提和结论在内的范畴的知识的归纳推理的心理模型。这种模型被称为基于类别的模型。Sloman（1993）提出了一个模型，其中归纳推理仅基于论点的特征（基于特征的模型）。这些模型是根据预先选择的关于论点及其属性之间关系的心理评估结果构建的。然而，很难客观地识别涵盖人类一般知识的所有属性，而这对于构建模拟人类归纳推理过程的通用模型是必要的。此外，对大量特征进行心理评估所涉及的成本意味着它们是不切实际的。为了避免这样的问题，本研究提出了三种类型的模型（神经网络模型，主观概率模型，贝叶斯模型），利用统计分析的结果，在计算两个词的共现概率语料库，而不是使用心理评估。通过一个关于归纳推理的心理学实验对模型进行了评价。将神经网络模型和主观概率模型的实验结果与仿真结果进行比较，发现两者具有良好的相关性。我们还成功地实现了一个试用版的基于归纳推理的搜索引擎使用的主观概率模型。Osherson等人（1990）提供了一种基于前提和结论之间的相似性以及对包括前提和结论在内的范畴的知识的归纳推理的心理模型。这种模型被称为基于类别的模型。相比之下，Sloman（1993）发展了另一种归纳推理的心理模型，其中只有前提和结论之间的相似性，基于它们的特征，具有重要作用。这种模型被称为基于特征的模型。Sloman认为，基于特征的模型能够更好地解释涉及条件句归纳推理的实验结果。Sloman的模型是一种神经网络。模型的输入节点表示结论的特征模式，输出节点表示结论与前提的相似程度。输入节点和输出节点之间的权重的强度根据一种delta规则从前提的特征模式计算。然而，虽然该模型仅能够处理包括肯定表达式的肯定前提，并且因此不能接受由否定表达式组成的否定前提，但是基于类别的模型对于否定前提是有效的（Osherson等人，1991年）。此外，Sloman的模型需要实验数据来确定前提和结论的特征强度。然而，很难确定所有的功能，涵盖一般的人类知识，是必要的，以构建一个通用的模型，模拟人类的归纳推理过程。除了这个问题，对大量特征进行心理评估所涉及的成本意味着它们是不切实际的。Osherson的模型也面临类似的问题，因为它也需要实验数据来确定前提和结论之间的相似性强度，以及前提和包括前提和结论在内的类别之间的相似性。为了解决这些问题的归纳推理的一般模型的建设，本研究提出了三种类型的模型，使用的结果，从语言语料库的统计分析，而不是心理评估的对象和它们的属性之间的关系。分析结果用于计算两个词之间的同现概率（在基于特征的模型的情况下）以及计算词和类别之间的同现概率（在基于类别的模型的情况下）。本文提出的基于特征的模型有一个输入层和一个输出节点，其中输入层和输出层之间的权重

The purpose of the present study is to propose computational models of human inductive reasoning, using a statistical analysis of Japanese linguistic data, and to develop a searchengine based on inductive reasoning. Osherson, et al. (1990) provided a psychological model of inductive reasoning based on the similarity between the premise and the conclusion and on knowledge of the category including the premise and the conclusion. Models of this kind are known as categorybased models. In contrast, Sloman (1993) proposed a model where the inductive reasoning is based only on the features of arguments (the feature-based model). These models were constructed based on the result from psychological evaluations concerning the relationship between arguments and their attributes which were selected in advance. However, it is difficult to objectively identify all the attributes that cover general human knowledge, which is necessary in order to construct a general model that simulates the human process of inductive reasoning. Moreover, the costs involved in conducting psychological evaluations for the sheer numbers of features means that they are prohibitively impractical. In order to avoid such problems, the present study proposes three types of models (a neural network model, a subjective probabilistic model, and a Bayesian model) that utilize the results of statistical analysis for a language corpus in computing co-occurrence probabilities for two words, rather than using psychological evaluation. A psychological experiment concerning inductive reasoning was conducted to evaluate the models. In comparisons of the experimental results and the simulation results for the neural network model and the subjective probabilistic model, good correlations were observed. We also successfully implemented a trial version of a searchengine based on inductive reasoning using the subjective probabilistic model. Introduction Osherson, et al. (1990) provided a psychological model of inductive reasoning based on the similarity between the premise and the conclusion and on knowledge of the category including the premise and the conclusion. Models of this kind are known as category-based models. In contrast, Sloman (1993) developed another type of psychological model for inductive reasoning, where only the similarity between the premise and the conclusion, based on their features, has an important role. This model is called the feature-based model. Sloman argues that the feature-based model is better able to account for the results from an experiment involving inductive reasoning for conditionals. Sloman’s model is a kind of neural network. In the model, input nodes represent the feature pattern of the conclusion, and the output node indicates the similarity of the conclusion to the premises. The strengths of weights between the input nodes and the output node are computed from the feature patterns of the premises according to a kind of delta rule. However, while this model is only capable of handling positive premises that include affirmative expressions, and so cannot accept negative premises consisting of negative expressions, the category-based model is valid for negative premises (Osherson, et al., 1991). Furthermore, Sloman’s model requires experimental data to determine the feature strengths of the premises and conclusion. However, it is difficult to identify all the features that cover general human knowledge that is necessary in order to construct a general model that simulates the human inductive reasoning process. Beyond this problem, costs involved in conducting psychological evaluations for the sheer numbers of features means that they are prohibitively impractical. Osherson’s model also faces similar problems, because it also needs experimental data to determine the strength of similarity between the premises and the conclusion, as well as the similarity between the premises and the category including the premises and the conclusion. In order to solve these problems for the construction of a general model of inductive reasoning, the present study proposes three types of models that use the results from the statistical analysis of a language corpus, instead of psychological evaluations of the relationships between objects and their attributes. The analysis results are used to compute co-occurrence probabilities between two words (in the case of the feature-based model) and to compute cooccurrence probabilities between a word and a category (in the case of the category-based models). The feature-based model proposed here has an input layer and one output node, where the weights between input

Computational models of inductive reasoning and their psychologicale examination:Towards an induction-based search-engine

归纳推理的计算模型及其心理检验：走向基于归纳的搜索引擎

文献摘要

被引文献

相似文献