Research on Computational Aids to Intellectual Discovery

英日联合研究计划的主题是“用于科学和技术发现的机器智能工具”。我们特别研究了三种方法的使用：*归纳逻辑编程*神经网络和相关的统计学习技术*数据库挖掘。更详细地，我们研究了如何使用由麻省理工学院开发的归纳逻辑编程系统PROGOL实现计算机发现和创造。结果发现，在PROGOL的输入中通过在背景知识中引入新的谓词来实现高阶概念学习和类比推理是可能的。此外，我们还研究了如何从数据库中发现知识，发现了从数据库中提取相关信息作为PROGOL输入的必要性，并发现了创建适当的“反例”和“背景知识”的重要性。然后，我们探索了对PROGOL进行并行化的可能性，以达到处理大型数据库的高效率。作为ILP的一个应用，我们选择了自然语言处理。通过认知心理学实验考察了日语对动词Break的亲和度，得出了日语对新名词亲和度的预测规则。在科学发现中，能够通过只调查有限数量的假设来反驳整个假说空间是至关重要的。我们发现，对于相当大的一类假设空间，可以实现这样的假设空间的反驳。

The subject of the joint Anglo-Japanese research initiative is"Machine intelligence tools for discovery in science and technology."In particular, we studied the use of three approaches :*Inductive Logic Programming*Neural network and related statistical learning techniques*Database miningIn more detail, we investigated how to realize discovery and creation by computer using an Inductive Logic Programming system PROGOLdeveloped by Muggleton. As a result, we found that it is possible toachieve higher order concept learning and analogical reasoning by introducing new predicates in background knowledge in the input of PROGOL.Furthermore, we investigated how to discover knowledge in database and found the necessity to extract relevant information from database to adjust as an input to PROGOL.Also we found the importance of creating appropriate"negative examples"and"background knowledge". Then, we pursued the possibility to parallelize PROGOL in order to achieve high efficiency for dealing with large database. As an application of ILP,we picked up natural language processing. We investigated the Japanese correspondence to the verb"break"by investigating the degree of affinity to various nouns by cognitive psychological experiments and obtained a rule for predicting the affinity to a new noun. In scientific discovery, it is essential to be able to refute a wholehypothsis space by just investigating finite number of hypothesis. We found it possible to achieve such refutability of hypothsis spaces for a fairly large class of hypothsis spaces.