Basic Research on the Automatic Construction of Chemical Reaction Knowledge Base

为了从化学反应数据库中自动构建化学反应知识库，开发了一个原型软件系统dNet。该知识库软件接受特定反应描述的输入，并将其存储在识别网络中。当向系统提供查询响应时，它从网络中检索类似的反应。利用检索到的反应通过类比的方式估计查询反应的性质。下面给出dNet系统的概要。DENT系统可以处理以连接表表示的分子结构和化学反应。在连接表中定义锚部分，并且系统接收它以及连接表本身和条目的相关属性。锚定部分设置视点以对辨别网中的条目进行排序。例如，化学反应的锚定部分是发生键交换的子结构，输入条目唯一地位于识别网中，遵循连接表之间锚定的亚同构所定义的偏序。当我们构建知识库时，系统在定位的位置生成一个新节点，然后在该节点上注册条目。如果对系统给予查询反应，则放置在查询位置的位置的反应被认为是对该查询的类似反应。通过参照相似反应的各种类比推理方法来估计查询反应的性质，DNET系统已成功地自动构建了操纵数十个脱氢反应的知识库。我们将结合可以判断副反应发生的推理方法来开发dNet系统。

A prototype software system DNET is developed to construct the chemical reaction knowledge base automatically from the database of reactions. This knowledge base software accepts the input of specific reaction descriptions, and stores it in a discrimination net. When a query reaction is given to the system, it retrieves similar reactions from the net. The retrieved reactions are utilized to estimate the property of query reaction by analogy. The followings give the outline of the DNET system.DENT system can handle molecular structures and chemical reactions represented as the connection table. An anchor part is defined in the connection table, and the system receives it as well as the connection table itself and the associated properties of the entry. The anchor part sets the viewpoint to order the entries in the discrimination net. For example, the anchor part of a chemical reaction is a substructure where the bond interchange occurs.An input entry is uniquely located in the discrimination net following the partial order defined by the anchored subisomorphism between the connection tables. When we construct the knowledge base, the system generates a new node at the located position and then register the entry on this node. If a query reaction is given to the system, the reactions placed in the vlelnity of the query location are considered to be the similar reactions to the query. The property of a query reaction is estimated using the various analogical inference methods referring the similar reactions.DNET system has succeeded to construct the knowledge base automatically manipulating tens of dehydrogenation reactions. We are going to develop the DNET system by incorporation of the inference method which can judge the occurrence of side reactions.