Studies on natural language processing systems based on brain-style architecture

本研究旨在构建基于脑式体系结构的柔性信息处理系统。我们取得了以下有意义的成果：1)发明了一种新的词的矢量表示方法；2)构建了以电子词典为知识库的神经网络。特别是，结果1)是非常重要的。到目前为止，我们还没有合适的词表示方法来反映词与词之间的距离：我们必须将一个向量中的每个元素关联到一个词。因此，为了制作一个实用的神经网络系统，我们必须处理一个数十万长度的向量。这意味着神经网络式的语言处理系统几乎不可能构建。由于本研究的结果，向量的长度减少到几百个，这使得创建各种复杂的神经网络风格的语言处理系统成为可能。对于结果2)，新提出的系统与1)的结果一起实现了类比推理。此外，我们还创建了另一种新的神经网络。它具有使用电子词典的层级结构。这个网络可以通过使用电子词典学习来进行归纳和演绎推理。该网络从概念词典中学习单词之间的关系，并结合知识，给出综合的推理结果。

This research aims at constructing flexible information processing systems based on brain-style architecture. We have obtained the following significant results :1) A new vector expression method for word is invented ;2) Neural network using an electronic dictionary as knowledge base is constructed. Especially, result 1) is very important. So far we did not have suitable expression method of word reflecting distance between words : we have to relate each element in a vector to one word. Therefore in order to make a practical neural network system, we have to treat a vector with hundreds of thousands length. This means neural network style language processing system is almost impossible to construct. Owing to the result of this research, the length of vector is reduced to several hundreds, Which makes possible to create various kinds of sophisticated neural network style language processing systems.As for the result 2), analogy like inference is realized by a newly proposed system together with the result of 1).In addition, we created another kind of new neural network. It has a hierarchical structure using an electronic dictionary. This network can do both inductive and deductive inference by learning using an electronic dictionary. The network learns the relation between words from a concept dictionary and knowledge is combined and gives integrated inference results.