As the analysis performance of the Natural Language Processing (NLP) systems has improved, grammar rules and lexicons to be used with NLP systems become complicated and large than before. On the other hand, it must be expected that new application filed using a NLP system is enlarging. The underlying task is to develop and modify the NLP systems with reconstruction of the grammar rules and lexicons that has the desired performance. Since the work needs many staffs and much time for the development, reconstruction of grammar rules and lexicons is hard work and also expensive. Methods to develop more efficiently the desired system which can be applied to new application filed are required. In this paper, a new method of a systematic description of grammar rules is proposed for the purpose of creating grammar rules and modify them. The way of abstracting grammar rules from linguistic phenomena has been studied with special reference to the formation of morph-syntactic structure of Japanese. Based on the generalization of the way, the new method was made for further improvement for grammar writing. The paper also describes the grammar rules created by using the proposed method to illustrate the effectiveness of the method. Thegrammar rules was applied to analysis of a set of newspaper editorial sentences. First, the linguistic phenomenon that the analysis was not completed were specified and extracted. We considered that the phenomenon were a demand specification of meeting a new analysis performance. According to the specification, the grammar rules were extended using the same procedure of the method. Next, we checked that the analysis performance of the extended grammar rules was increasing. The result shows that it becomes possible to raise the analysis performance, maintain consistency of grammar rule description by following the method through the experiment. Generality raises for development work procedure of a grammar rule on engineering. Development time can also be shortened. The grammar rules described as an experiment runs on computer and are published from ICOT (Institute for New Generation Computer Technology). The proposed method in the paper is considered to be a trial which systematizes know-how of description of grammar rules as a knowledge base.

View full abstract

This paper describes the re-interpretation stage of understanding metaphor/simile. We define the meaning of a noun by using probability. The meaning of a metaphor/simile is the meaning of the tenor which is affected by the vehicle. We define “clarity” and “novelty”, which are based on the quantity of information, as measures of the meaning difference between metaphor/simile and tenor. The clarity of an attribute of a metaphor/simile measures the decrease in entropy. The novelty of an attribute of a metaphor/simile measures the increase in rareness. We measured the meanings of similes by using Semantic Differential and calculated their clarity and novelty. We showed that the clarity of an attribute of a simile agreed well with the salience pattern of the vehicle, tenor, and simile. We also showed that the clarity and novelty of a simile agreed with the comprehensibility of the simile.

View full abstract

This paper proposes a learning method that automatically acquires English verb selection rules for machine translation using a machine learning technique. If the rules are learned from only real translation examples, many examples are necessary for good translation quality. It is, however, difficult to gather a sufficiently large number of real translation examples. The main causes are verbs of low frequency and the frequent usage of the same sentences. To resolve this problem, the proposed method learns English verb selection rules from hand-made translation rules and a small number of real translation examples. The proposed method has two steps: generating artificial examples from the hand-made rules, and then putting the resultant artificial examples and real examples into an internal learner as the training set. The internal learner outputs the final rules with improved verb selection accuracy. The most notable feature of the proposed learner is that any attribute-type learning algorithm can be adopted as the internal learner. To evaluate the validity of the proposed learner, English verb selection rules of NTT's Japanese-English Machine Translation System ALT-J/E are experimentally learned from hand-made rules and real examples. The resultant rules have better accuracy than either those constructed from the real examples or those that are hand-made.

View full abstract

In this study we propose a method of syntactic analysis which approaches parsing as a search problem, controls the agenda of the chart parser following the strategy of A algorithm and generates only the required number of syntactic structures, from the most adequate in descending order of goodness. Given a set of costed rules of syntactic analysis in CFG form, at the pre-processing stage, for A to become admissible, every sub-structure which can be generated by these rules receives an estimate of the cost from the present state, corresponding to the sub-structure itself, to the goal state, corresponding to a complete structure of which this sub-structure is a constituent part. Then, at the stage of syntactic analysis proper, all competing structures are processed in the order from the lowest-costing, where cost is the sum of the estimate and the actual cost of generating the structure, and thus the lowestcosting structure is guaranteed to be obtainable. This method allocates appropriate cost values to each parsing rule. This makes it possible to prioritize syntactic structures in great detail and also to modify priorities, if necessary, simply by altering the cost values. Costed CFG rules can be seen as extending the descriptive power of Probabilistic CFG rules.

View full abstract

This paper describes a method of evaluating quality for developers of machine translation systems to easily check imperfections in their own systems. Our method employs test-sets in which example sentences, their model translations, yes/no questions for evaluating the system output, similar examples (if any), and grammatical explanations have been systematically aligned. The example sentences have been collected focusing on wide coverage of both basic linguistic phenomena and linguistic phenomena problematic to MT systems. The questions in the test-sets are designed to clarify the evaluation viewpoints. Given the system outputs for each example sentence in question, the system developer needs only to answer the question assigned to the example sentence. This judgment does not vary among evaluators, thus enabling an objective evaluation.

View full abstract