Robust Incremental Parsing based on Finite-State Approximation of Context Free Grammar

在本研究中，一直致力于开发具有同声传译功能的多语种之间的翻译技术，目标是以与语音输入相同的速度处理的句法分析技术的高速和渐进的发展。选择了利用语料库和句法树开发基于统计方法的近似转换技术，能够恰当地反映语法规则的使用频率等。具体推进了以下几个方面的研究。^*英语、日语和泰语语料库的维护。使用名古屋大学综合声音信息研究基地收集的同声传译对话语料库。在本研究中，它的翻译数据是使用24小时的日语数据，其中关于泰语的英语翻译数据是用原始数据制作的。语法树数据被提供给每个…更多的语料库以短语结构语法的形式出现。^*从大规模语料库中获取统计信息。从给出句法树的语料库中获取各种统计信息的技术。通过搜索语法树中出现语法规则的位置，并用上下文信息对其进行描述，对其进行统计分析。^*有限自动机近似转换技术的发展。研究了从上下文无关文法到有限自动机的转换技术。在转换中。自动机是通过将上下文无关文法表示为自反变迁网络的形式，并将其向下发展而形成的。根据概率计算，开发了优先开发使用频率较高的ARE的算法作为开发方法。^*语法学习、有限自动机近似和句法分析系统的设计与实现。为了实现可行的分析，使用了由大约5000万个区域组成的有限自动机。^*句法分析的比较评估。使用基准标记进行了英语、日语和泰语的句法分析实验。从准确度、时间、句法树的数量和形式等多个角度对该分析技术进行了比较评估，通过两年的研究，用上下文无关文法的自动机近似验证了健壮增量式句法分析的可实现性，对分析速度的影响得以证实。较少

In this research, it has been aiming at the development of high speed and gradual progress the parsing technology treatable by the same degree of the speed as the voice input to develop the translation technology between several languages that have the simultaneous interpreter function. The approach of developing the approximation conversion technique based on the statistical method that was able to reflect the use frequency etc. of the grammatical rule appropriately by using the language corpus with the syntax tree was selected. The research of the following items was concretely promoted.^*English, Japanese, and Thai corpus maintenance.The simultaneous interpreter conversation corpus that had been collected in the Nagoya University integration sound information research base was used. In this research, the translation data of it was made by using Japanese data of 24 hours and the English translation with the original data among these about Thai. The syntax tree data was given to each … More language corpus in the form of the phrase structure grammar.^*Acquisition of statistical information from large-scale corpus with syntax tree.The technique to acquire various statistical informations was examined from the language corpus from which the syntax tree was given. It was statistically analyzed by searching for the position in the syntax tree where the grammatical rule appeared, and describing it with the context information.^*Development of limited automata approximation conversion technique.The technique for converting it from the context-free grammar into limited automata was researched. In conversion. Automata were made by expressing the context-free grammar in the form of the reflexive transition network, and developing them descending. The algorithm that developed the are with high use frequency by priority was developed as a development method according to the probability calculation.^*Design and mounting of parsing gradual progress system.The system of grammatical acquisition, the limited automata approximation, and parsing was designed, and mounted. Limited automata that consisted of the are of about 50 million were made for the achievement of a practicable analysis.^*Evaluation for comparison of parsing.The parsing experiment in English, Japanese, and Thai was executed by using the bench mark. The evaluation for comparison of this analytical technique from a diversified viewpoint like accuracy, time, and the number and the form etc. of the syntax tree was executed as a result.The realizability of robust incremental parsing was verified by the automata approximation of the context-free grammar through the research on two years, and the effect on the speed-up of the analysis was able to be confirmed. Less