Kazuhiko Tsuda

Word classification and hierarchy using co-occurrence word information

By the development of the computer in recent years, calculating a complex advanced processing at high speed has become possible. Moreover, a lot of linguistic knowledge is used in the natural language processing (NLP) system for improving the system. Therefore, the necessity of co-occurrence word information in the natural language processing system increases further and various researches using co-occurrence word information are done. Moreover, in the natural language processing, dictionary is necessary and indispensable because the ability of the entire system is controlled by the amount and the quality of the dictionary. In this paper, the importance of co-occurrence word information in the natural language processing system was described. The classification technique of the co-occurrence word (receiving word) and the co-occurrence frequency was described and the classified group was expressed hierarchically. Moreover, this paper proposes a technique for an automatic construction system and a complete thesaurus. Experimental test operation of this system and effectiveness of the proposal technique is verified.

An incremental algorithm for string pattern matching machines

Aho and Corasick presented a string pattern matching machine (hereafter called machine AC) to locate multiple keywords. However, the machine AC must be reconstructed all over again when a keyword is appended. This paper proposes an efficient algorithm to append a keyword for the machine AC. This paper presents the time efficiency comparison with the original algorithm using the actual simulation results. The simulation results show the speed up factor, by the algorithm proposed, to be between 25 and 270 fold when compared with the original algorithm by Aho and Corasick which requires the reconstruction of the entire machine AC.

Fast and compact updating algorithms of a double-array structure

In many information retrieval applications, it is necessary to be able to adopt a trie search for looking at the input character by character. As a fast and compact data structure for a trie, a double-array is presented. However, the insertion time is not faster than other dynamic retrieval methods because the double-array is a semi-static retrieval method that cannot treat high frequent updating. Further, the space efficiency of the double-array degrades with the number of deletions because it keeps empty elements produced by deletion. This paper presents a fast insertion algorithm by linking empty elements to find inserting positions quickly and a compression algorithm by reallocating empty elements for each deletion. From the simulation results for 100 thousands keys, it turned out that the insertion time and the space efficiency are achieved.

A method for the expansion and contraction of text's size by using the string replacement algorithm

Abstract This paper describes a method for the expansion of the size of texts using an efficient string replacement algorithm. A method for the expansion of the size of texts by use of morpheme replacements is supported by the rules for expansion morphemes. This paper explains the text expansion method using the example of text reduction. Morpheme replacements are defined as a combination of a pattern of categories and restrictions on morphemes. A replaceable pattern is efficiently detected through an extension of Ahos string pattern-matching algorithm. This extension consists of the application of a matching algorithm that replaces the input morpheme. Results of simulations upon several texts show that reduced texts are between 5% and 10% shorter than their original counterparts. Also, by use of this pattern-matching algorithm for detection of the suitable reducing rules, the number of state translations can be reduced to 1 7 of that when using Ahos pattern-matching algorithm.

A new approach to phoneme recognition by phoneme filter neural networks

This paper describes a phoneme filter neural network (PFN) approach to phoneme recognition. Most conventional speech recognition neural networks have a serious drawback: the network output values do not correspond to candidate likelihoods. The PFN is a multilayer neural network with fewer hidden units than input units prepared for each of the phoneme categories. Each network is trained as an identity mapping by speech data belonging to one phoneme category. In the recognition process, the similarity between the input data and output data is computed for each network. The results of the experiment to apply the Japanese vowel recognition task showed that the PFN recognition rates for the top two or more choices are higher than those of a conventional three-layer neural network and the PFN outputs represented candidate likelihoods. It was also confirmed that the PFN had a mapping ability and recognition performance superior to those of the linear K-L transformation method because of the nonlinearity of the PFN.

Word category prediction based on neural network

This paper proposes a new method of the word category prediction for the speech recognition system. In order to improve the speech recognition results, not only the acoustical information but also certain linguistic information is needed. World category prediction is a very effective method to implement an accurate word recognition system. Traditional statistical approaches require considerable training data to estimate the probabilities of word sequences, and many parameters to memorize probabilities. And it is difficult to predict unseen data which does not include the training data. To solve this problem, NETgram, which is the neural network for word category prediction, is proposed. The performance of the NETgram is comparable to that of the statistical model although the NETgram requires fewer parameters than the statistical model. Also the NETgram performs effectively for unknown data, i.e., the NETgram interpolates sparse training data. Results of analyzing the NETgram show that the NETgram learns ...