A Research on a English・Japanese Conversational Text-Base and Its Retrieval System Keyed by the Semantic Information

(1)英日会话语篇收集及语义结构描述数据通过本项目，我们从NHK电台英语会话语篇中收集了英语和日语会话句语篇。句子的总数多达20万句，其中包括简短的句子，如“是的”、“嗨”、“哦，亲爱的”。至于语义结构描述数据(SD-Form数据)，总数约为8000句。会话文本和SD-Form数据数据库管理系统的开发我们为SD-Form语义模型环境(SDENV)开发了一个Prolog程序。在该核心系统的基础上，开发了会话式文本管理系统和SD表单管理系统。系统功能包括A.新概念标注添加、B.动态知识设置、C.主题管理等功能。(3)会话文本库的应用在本项目中，我们还开发了一个“英日会话文本语义检索”的原型系统。该系统配备了许多技术，以使系统运行更快。一种典型的技术是保留以前函数求值的结果，即将子函数的成功/不成功结果存储到PROLOG程序的数据库中。这项技术使系统运行速度提高了10倍。文本检索分为四种模式，每种模式都有不同的检索键。它们是，A.SD-Form，B.语句中的功能项SD-Form，C.概念标签，以及D.日语/英语文本的示例。在每种模式下，系统都会根据SD-Form语义模型中的概念细化关系计算语义差异度量。

(1)Collection of English-Japanese Conversation Texts and the Semantic-structure Description DataThroughout this project, we have collected English and Japanese conversational sentence texts from NHK RADIO ENGLISH CONVERSATION TEXT. The total number of sentence piece is as much as 200, 000 including short utterances like, "Yes", "Hi", and "Oh, dear. "As for the semantic structure description data(SD-form data), the total number is about 8, 000 sentences. We assembled all the data in a MS-DOS file.(2)Development of database management system for conversational texts and SD-form dataWe developed a program for SD-form semantics model environment(SDENV)as a Prolog program. Based on this core system, we also developed a Conversational text management system and a SD-form management system. The system function includes such function as A. New concept label appending, B. Dynamic knowledge installing, C. Topic managing, etc.(3)Application of the conversational textbaseIn the project, we also developed a prototype system of a "Semantic retrieval of the English/Japanese conversational texts". The system was equipped with many techniques to make the system operation faster. A typical technique is to retain the results of the previous values of functional evaluation, that is, we store the sucessful/unsuccessful results of the subfunctions into the database of the Prolog program. This technique enabled a 10 times faster system operation. Text retrievals are categorized into four modes, all of those have different retrieve keys. They are, A. An SD-form, B. Functional item in a statement SD-form, C. Concept label, and D. An example of Japanese/English text. In every mode, the system computes the semantic difference measure which is based on the concept elaboration relations in the SD-form semantics model.