Semantic and Contextual Analysis using Common knowledge from Japanese Articles

研究的主要目的是开发一个实用的语义分析系统SAGE。为此，我们试图在原始原型的基础上提高精度和速度的性能。在精度方面，我们从四个方面进行了改进。(1)除了从EDR语料库中计算出的统计度量来确定词之间的深层格外，我们还补充了一个基于规则的程序来确定词之间的格，该程序基于粒子、词性和词义的信息。我们设法确定了一个不可退化词和它的修饰语之间的深层大小写。(3)用概念相近的已注册词代替未注册词，可以分析EDR词典中未注册词的深度格。(4)我们将包含括号的表达式分为补充、委托和补充3类，并在SAGE中加入括号分析。采用上述技术的系统对来自EDR语料库的句子的词义和深度格的准确率分别达到90.2%和90.0%，对来自互联网的新闻文章的准确率分别达到87.0%和86.8%。在加速方面，我们设计了一种线性顺序算法，包括确定词义和深度大小写两步。我们首先从下往上看修饰词如何影响修饰语的意思，然后用概率表示它。之后，我们使用自顶向下的方法确定每个子句的含义。我们使用该算法实现了大约1万倍的加速。此外，我们将访问EDR字典所需的时间减少到原来的1 / 5，并将整个系统从Prolog版本转换为C版本，从而实现了5倍的速度提升。

The main aim of the research is to develop a practical semantic analysis system SAGE. For this aim, we attempted to improve the performance both in precision and speed based on the original prototype.In regard to the precision, we made four efforts to improve it. (1) In addition to the statistical measures calculated from the EDR corpus to determine the deep case among words, we supplement a rule-based procedure to determine the case between two words, based on the information on particles, parts of speech, and word meanings. (2) We contrived to determine the deep case between an indeclinable word and its modifier. (3) It has been made possible to analyze the deep case of unregistered words in the EDR Dictionary by replacing them with the registered words holding similar concepts. (4) We classified the expressions containing brackets into 3 categories, supplementation, commission, and complement, and add the analysis of brackets into SAGE.The present system employing the above techniques has reached a precision of 90.2% on word meaning, 90.0% on deep case for the sentences from EDR Corpus, and 87.0% and 86.8% respectively for news articles from the Internet.In regard to the speed-up, we devised a linear-order algorithm including two steps for deciding word meaning and deep case. We first see how the modified word impacts the meaning of the modifier from bottom up, and then represent it with a probability. After that, we determine the meaning of each clause with a top-down approach. We implemented a speed-up of about 10 thousand times with the algorithm. Besides, we reduced the time required for accessing the EDR Dictionary to 1 fifth of the original time, and converted the whole system from a Prolog version to a C version and hence realized another speed-up of 5 times.