JP2812495B2 - Syllabic input of language using kanji - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION 3.1 Field of Industrial Application The present invention relates to a word processor of a language that uses a sentence consisting only of kanji or a sentence in which kanji and phonetic characters are mixed.
More specifically, the present invention relates to Chinese, Japanese, and Korean word processors. More specifically, the present invention relates to phonetic characters (Chinese Sound, Kana or Romaji for Japanese, Hangul for Korean)
A syllable delimiter system that automatically reads the sentence reading in syllable units using, and automatically follows the input as well as a kanji conversion system that performs kanji conversion for each word separated by this syllable delimiter About.

3.2 Conventional technology 3.2.1 Weakness of kana character conversion by longest match method in Japanese word processor [Ai / Shoka] has difficulty in kana kanji conversion between "love / early summer" and "love book / house" (shaded lines) / Represents a word separator.
same as below). It is difficult to judge whether "Jokigen" is "upper / cheerful" or "steam / source". Although it is well converted at one time into "periodic onset" and "periodic oscillation", "shukyuekiseichi" becomes "shukyutekisemachi". When performing so-called collocation batch conversion, such delimitation of words frequently occurs, and sentence batch conversion has a function in a current Japanese word processor (hereinafter may be simply abbreviated as word processor). However, there is no performance enough to truly convince the user, and it is often not actually used.

There is also a Japanese word processor that performs a so-called word connection process in order to improve the normal conversion rate in the sentence batch conversion. In this word connection process, vocabulary is classified into main words such as nouns and verb stems and functional words such as particles and auxiliary verbs or inflected endings and prefixes and suffixes. Is introduced as ancillary data, and word separation is performed using the ancillary data. By adopting such a word connection process, the correct conversion rate is considerably improved, but even if the word connection process is performed, erroneous conversions are not small except for the inflected endings of verb adjectives. "Gohenkan" is either "return" or "return", and when you want to write "wrong conversion", it is likely to be wrong conversion in one shot unless you separate the word with "go / henkan". .

The main kana-kanji conversion method of the current Japanese word processor is
This is the longest matching word delimiter + word connecting process, and as long as the longest matching word delimiter is used, if syllables of various lengths are mixed to form a phrase, the conversion result is poor.

The present invention uses the statistical frequency of speech sounds (word readings) as a weight to determine which speech sound arrangement is optimal for the entire sentence.
The principle of minimum speech entropy is determined by the principle that the sequence of speech sounds that maximizes the product of the frequency of each speech sound that constitutes a sentence is the closest to true. Kanji conversion is performed individually for. The longest match method also optimizes the arrangement of phonemes, but it says that "song syllables of speech have longer weights."
It is based on the too simple principle. In the present invention, word division is performed by weighting with a frequency statistic, so it can be said that the method is theoretically firmly supported.

3.2.2 In Chinese word processors From the point of view of ergonomics, Chinese word processors are Chinese-style Roman characters Key in the sound The royal road is to convert sound to kanji and output kanji sentences. However, it has some cultural hurdles. The most important of them is the phonetic The fact that they are described in sound is that "Chinese are mistaken for word separation." Since ancient times, writing has been done using only the ideographic character of kanji, the Chinese sentence did not need to be separated. Originally, however, Chinese has many monosyllable vocabulary words, and sugar (eating candy) is established as a word either as sugar or as a sugar, and has a Hu theory (say something like Hu-stupid The word delimiter is not wrong in either the Hu theory or the Hu / theory. It is said that 90% of the vocabulary in modern Chinese is such a word with unclear word separation (co-authored by Takezoku and others, Kanji / Kanji Reform History, 1988, Hunan People's Press, p.257).

Developed to date and already available in China and Japan All sound-input Chinese word processors use the kanji conversion method on a word-by-word basis (Chinese does not have phrases like Japanese, so the longest match method that assumes line-line conversion cannot be used). Since 97% of the Chinese vocabulary is monosyllabic or disyllable, this method requires frequent human conversion keystrokes, and it is inevitable that human punctuation errors will occur all the time.
Therefore, in Chinese word processors, it is most desirable to introduce full-word kanji conversion using automatic word separation or sequential word separation kanji conversion following syllable input.

For the Chinese word processor that performs full-text batch kanji conversion by automatic word separation, see the invention "Chinese word separation method and speech kanji conversion method (Japanese Patent Application No.
−105030) ”and“ Chinese phonetic delimiter method (A.
63-172163) ". In these inventions, the statistical frequency of speech sounds (word readings) is used as a weight, and the optimal arrangement of speech sounds (that is, speech sound segmentation) for the entire sentence is determined by the principle of minimum speech entropy. Is adopted.

The simplicity of Chinese reading, as mentioned above, that monosyllabic and disyllable terms are the bulk of the vocabulary, is a great advantage when applying automatic word segmentation by phonetic frequency. It has been well proven by running tests (see FIG. 8). However, for sentences with more than three syllable words,
In the above two inventions, some maladaption may occur.
Such anxiety can be eliminated by the present invention.

3.2.3 Korean word processor and Hangul-Kanji conversion situation It is thought that all Korean is written in Hangul, but traditional Kanji-style Hangul sentences are still strong,
In addition, there is a tendency that the re-evaluation of kanji gradually progresses.
Currently, the mainstream of the Korean word processor is the Hangul element input Hangul conversion method or the method that adds Hangeul Kanji conversion for each kanji idiom. Word processors may spread. FIG. 13 shows an example of Korean corresponding to Japanese.

As shown in the example in Fig. 13, Korean is kanji (expressed in Korean kanji sounds in katakana) and pure Korean vocabulary (expressed in hiragana)
It is very similar to Japanese in that it consists of Grammar and word order are similar to Japanese. Therefore, in a word processor that handles Hangul sentences with kanji, the design policy may be basically the same as that of a Japanese word processor. Therefore, it will be described later that the present invention is sufficiently effective also for a word processor of a Hangul sentence composed of Chinese characters.

3.3 Problems to be Solved by the Invention As described above, the present inventor has already filed patent applications for two inventions for the Chinese word processor (Japanese Patent Application No. 63-105030: Chinese word separation system). And Japanese kanji conversion method, Japanese Patent Application No. 63-172163: Chinese word separation method). Each of these inventions Assuming phonetic input, for monosyllabic words and disyllable words, define the statistical frequency of speech as a frequency that is an information amount, import it into dictionary data, and minimize the sum of the frequency of words that compose the sentence. "Principle of minimum speech information entropy" that the arrangement of speech sounds that shows the most probable separation of speech sounds
Is applied to execute automatic speech separation. This technique of automatic speech separation is the "tone frequency method".

However, in the above two inventions, the processing target of the vocal frequent method is limited to monosyllabic words and disyllable words. The reason is that, when a speech sequence including speech up to M syllables (M is theoretically any positive integer) up to M syllables is generally processed by the tone frequency method, the processing becomes complicated and the processing time becomes longer, and the reading time becomes longer. This was because it was thought that there was a risk that the design conditions for a word processor required without short-time real-time processing of kanji conversion would not be satisfied.

In order to supplement the above two inventions, the present invention aims to provide as simple and simple a syllable speech processing algorithm as possible for syllable processing sections containing three or more syllables.

Compared to Chinese, Japanese and Korean have longer syllable lengths per word. Therefore, the present invention enables vocal phrasal speech segmentation processing in a relatively short processing time even for speech sounds having a length of at least three syllables, and also uses speech frequent method which can be applied to Japanese and Korean besides Chinese. It is intended to provide a system and a kanji conversion system.

3.4 Means for Solving the Problems The means provided by the present invention for solving the above-mentioned problems are syllables obtained by sequentially inputting syllables using phonetic characters in units of sentences in a language using kanji. Syllabic input speech to obtain the most accurate kanji word sequence by sequentially dividing the syllable sequence into speech sounds to obtain an optimal speech sequence, sequentially performing speech to kanji conversion for each speech in the speech sequence, In the sequential delimited kanji sequential conversion method, the pronunciation of each word having the same pronunciation and being statistically significant is defined as one speech sound, the statistical frequency of occurrence of each speech sound used in the sentence of the language is defined as f, the syllable length is set to s, when the total syllable total number of all speech in the speech system material and the F _t, Shikiritsu p of each word sound
, P = (f × s) / F _t, and the frequency I of each speech is an integer, where I = int (−log _a p), where a = 2, and a continuous speech sequence in the sentence of the language is input. Then, from the first syllable at the beginning of the word string, the nth most recently input
When the syllable string up to the syllable is a frequent phrase, the frequent phrasal is subjected to the sequential speech-separation-sequential kanji sequential conversion process at the latest time point, and when the sum of the frequent classes of the respective speech sounds in the syllable is frequently summed, In the above, the speech sounds of 1 to M syllable lengths (where M is an integer of 2 or more) which are statistically significant are used as headings, and the speech frequency frequent dictionary storing the frequency of each speech is used as the heading. A lexical kanji word dictionary in which kanji homonyms to be read are stored in the form of kanji character strings, and each time one syllable is input in the vocal phrase,
Searching for the frequency of each speech in the speech frequency dictionary, using each speech of M types having 1 to M syllable lengths ending with the syllable as a heading, and describing each speech and each frequency in the next section With respect to the speech frequency class search means to be sent to the optimal frequency class sum sequential calculation means, and the optimal frequency class sum sequential calculation means, the syllable input number in the vocal phrase is n (= 1, 2, 3,..., N)
, R _n1 , R _n2 , R _n3 ,..., R _nM , respectively, each of the M words having a length of 1 to M syllables ending with the n-th syllable input recently. The classes are _In1 , _In2 , _In3 , ..., _InM , respectively. The syllable segmentation type of the type that can be maximized in the syllable string of n syllable length is defined as the syllable length m of the last syllable. 1,2,
.., M and the readings are classified into M sets of R _n1 , R _n2 , R _n3 ,..., R _nM , respectively, and the minimum frequency sum for each of the M sets is P _n1 , P _n2 , P
_n3, ......, further the P _n1 and _{P nM, P n2, P n3} , ......, the optimum Shikikyu sum Pn to the smallest value in P _nM, by sequential syllable input, n is from 1 1 As the number increases,
With respect to obtaining the P _n, when n is in the range of 1 ≦ n ≦ M, when n is 1, 1 P _1, calculated by _{P nm = I 11 = P 1} , n is relative to M < When n ≦ M, n−1 P _nm for n−1 m of m ≦ n−1
Is calculated by P _nm = P _nm + I _nm , and for one m of m = n, one P _nm is calculated by P _nm = P _nn = I _nm , so that P _n1 , P _n2 , ..., P _nM M minimum frequency sums are obtained. When n is in the range of M <n for M, M minimum frequency classes for M m of 1 ≦ m ≦ M The sum is calculated by P _nm = P _nm + I _nm , and finally, the n minimum frequency sums of P _n1 , P _n2 ,..., P _nm are obtained. The minimum utterance class sum of the syllable delimiter type is added to each of the n or M minimum frequent class sums P _nm already obtained and stored by the processing immediately after the syllable input m m syllables before the current syllable input number n. , A minimum sum sum sequential calculation means obtained by adding each of the n or M frequency classes I _nm searched at present, and the minimum frequency sum successive calculation means each time one syllable is input. Is Receives the value of n or the M minimum frequent class sum, controller compares selection of n's or M's alternative to these minimum values among the n's or M's
Seeking P _nm as the optimal Shikikyu sum P _n, the store the value of P _n, frequent optimum seeking and only word sound column at the end of the speech R _nmo and number of syllables mo of word or sound with該頻class sum simultaneously Class sum _sectioning type selection means, receives the speech R _nmo obtained by the optimal frequency class sum _sectioning type selection means, and, with the R _nmo as a heading, a homophone kanji word of the syllable number mo ending with the currently input syllable. Among them, the most probable kanji word H _nmo at present is read from the phonetic kanji word dictionary and sent to the most probable kanji string calculating means in the next section, the most probable kanji word search conversion means, The determined mo and the most probable kanji word obtained by the most probable kanji word search conversion means of the preceding paragraph
In response to H _nmo , every time n is incremented by 1 each time the syllable is input with 1 as the initial value, the current most probable kanji character string K _nmo is obtained by adding the character string K _nmo = K _n-mo + H _nmo The most probable kanji character string calculating means as an output of the means of the invention. In summary, when the language has speech sounds from one syllable to M syllable, in a sentence of N syllables, the syllable number is n,
The initial value of n is set to 1 and the syllables are sequentially input to the syllable. Each time n increases by one, up to M syllables that can exist at the end of the sentence (n when n ≦ M) A speech sound and a frequent class are obtained, and up to M frequent sums of the already determined probable optimum speech sequences for the speech sequence from the beginning of the sentence to the syllable immediately before the end speech are obtained. By adding the frequent classes of the last vocal sounds, a sum of vocal sequences of up to M words having the length of n syllables is obtained. Is the current optimal phonetic punctuation, and
The ending speech of the speech sequence is determined as the current optimal ending speech, and the kanji sequence obtained by converting the vocabulary to kanji is designated as the ending speech kanji sequence, and the kanji conversion sequence for the known most probable vocabulary sequence prior to the speech is determined. Obtain a new kanji string connected to the last word kanji string,
A syllable-input-sequential-separated-kanji-sequential-sequential-conversion system for a language using kanji, characterized in that the kanji string is newly newly output each time n increases by one.

3.5 Action 3.5.1 All combinations of speech sounds of 1 to m syllables in n-syllable length syllable processing section and its system Fig. 1 shows 1 to 7 syllable word sounds in syllable processing section with 1 to 7 syllable length Here is an example of a list of possible speech-segmentation types when all of the above exist.

However, lowercase Roman letters a, b, c …… indicate syllables,
For example, abc indicates a speech of three syllables by three syllables a, b, and c. Also, / indicates a separation between speech sounds.

First, the main terms used in the description of the present invention are defined as follows.

The pronunciation of each word having the same pronunciation and statistically significant is regarded as one word sound, the frequency of statistical appearance of each speech used in the sentence of the speech is assumed as f, the syllable length of the speech is assumed as s, the speech statistics when the total syllable total number of all speech in the material and the F _t, Shikiritsu P of each word sound
Is defined as P = (f × s) / F _t, and the frequency I of each speech is an integer as I = int (−log _a p), where a = 2. When there is no two or more polysyllabic words ending with the one syllable word input to the syllable word, the point immediately before the one syllable word in the word sequence is set as a node, and the word sequence between two adjacent contact points is determined. A syllable sequence is referred to as a vowel phrase, and the sum of the continuations of each speech sound from the first syllable of the vowel phrase is referred to as a frequent sum.

As shown in the bottom column of FIG. 1, when the syllable length of the syllable processing section (the section of the syllable phrase to be processed) is n, the number U _n of theoretically possible speech segmentation types is 2 ⁿ⁻ There can be ^one species. Therefore, when n = 8, there are 128 types of partition types, and it can be seen that the calculation process of selecting one optimal partition type with the smallest frequent sum takes a considerable amount of time. However, the partitioning type can be systematically arranged with a certain order as shown in FIG. 1 (word-segmentation type tree structure), and as n increases by one, any type in the next n Exists can be calculated using the order.

The partition type shown in FIG. 1 is arranged for each n according to the following order.

(1) The type in the top row is a word string of n monosyllable words, and the type in the bottom row is a word string of one n-syllable word (particularly shown in white characters).

(2) The type of delimiter in the speech sequence is represented by the following binary number B
Expressed by

(A) If there is a delimiter / syllable between syllables, 1 is set;

(B) The first delimiter position (between a and b) of the syllable string is the first digit of the binary number, and the delimiter position between the second b and c is the second digit of the binary number. As a number, a binary number of n-1 digits is created for a word string of n syllables to represent a delimited type. For example
a / b / cd / efg is B = 001011.

(3) The above-mentioned binary number B is, for example, n = 7.
Starting from 000000 representing abcdefg on the bottom line and increasing in order by one upward and increasing to 1111 representing a / b / c / d / e / f / g on the top line
Leads to one. Therefore, B represents all 2 ^7-1 = 64 binary numbers, and the delimiter type represented by B covers all delimiter types of a word sequence of seven syllable lengths composed of speech sounds of one to seven syllables. You are doing.

Looking from left to right, the partitioning type in FIG. 1 shows that the partitioning type having the number of syllables n is classified into a group of types having the following structure.

(A) The type of a one-syllable word sound ends with the same one-syllable word sound added to the end of each type in the syllable number n-1.

(B) The type of the two-syllable word sound at the end is obtained by adding the same two-syllable word sound to the end of each type in the syllable number n-2.

(M) The type of m-syllable words at the end is the same m-syllable word at the end of each type in the syllable number nm.

3.5.2 Simultaneous progression of the process of finding the minimum frequency sum and the process of converting the optimal syllables into kanji. Looking at the above vertical and horizontal order from another aspect, the vertical order is the same as that of Fig. 1 It shows that the delimited type is covered, and the horizontal order is
One syllable word a at n = 1, two syllable word a at n = 2
b,..., m-syllable words abcd at n = m, abcd... m, etc., starting from a at n = 1, n is incremented by 1 in each of the above procedures (a) to (m). It shows that every type belonging to a new n can be systematically and sequentially created. FIG. 1 shows the resulting speech-segmented tree structure.

From the above perception of the order of the tree structure of the phonetic division type,
The following important conclusions regarding the sequential calculation of the least frequent class are drawn.

When the speech to be handled by the vocal frequency processing is limited to speech of 1 to M syllables, in the speech sequence of n syllable length, (1) the speech sum of the speech sequence of the last one syllable speech having the minimum frequency sum is: It is obtained by adding the frequency of the monosyllabic speech of the nth input syllable to the frequency sum of the speech sequence having the minimum frequency sum in the word sequence having n-1 syllable lengths. P _n1
And The number of syllables in the last word of the word string having the frequent sum is 1
It is.

(2) The frequent sum of a speech sequence having a minimum frequency sum in a speech sequence having a two-syllable word sound at the end is the sum of n-1 in a speech sequence having a minimum frequency sum in a speech sequence of n-2 syllable lengths, and n-1. It is obtained by adding the frequency of a two-syllable word sound that is the sum of the syllable input at the nth position and the syllable input at the nth position. The value of the frequency sum is defined as P _n2 . The number of syllables of the last speech in the speech sequence having the frequency sum is 2.

…… …… (M) The frequent sum of the syllables ending with M syllables that has the least frequent sum is the frequent sum of the syllables that have the minimum number of sums in the mnemonic syllable length And a frequency of M syllable words obtained by sequentially connecting M syllables from the (n−M + 1) th input syllable to the nth input syllable. Let the value of the frequency sum be P _nM . The number of syllables of the last speech in the speech sequence having the frequency sum is M.

As a result, the partitioning type having the minimum frequency sum in the word sequence having n syllable lengths is the partitioning type P _n1 , P _n2 ,..., Of the above M minimum frequency sums (1) to (M). In P _nM , there is only one delimited type of the total minimum frequency sum with the smallest value. Let the value of the total minimum frequency sum be P _n . Based on the principles of sound frequent method compartment Setsugata is optimal speech separator type, the number of syllables end of speech and mo, if the frequent class of the speech and _{I mo, P n = P n} -mo + I mo It is.

n is defined as a syllable input number indicating the order of syllable input of syllable phrases. It is assumed that n has an initial value of 1 and increases by one as a syllable is input.

Assume that the syllable currently input (n = n) is at the end, and M words that read a syllable string having a length from one syllable to M syllables are R _n1 , R _n2 ,..., R _nM .

The frequencies of speech sounds R _n1 , R _n2 , ..., R _nM are represented by In ₁ , In ₂ , respectively.
…, _InM .

At the same time speech is R _n1, R _n2, ......, in the R _nM, H _n1, H _n2 seems now most certainly Kanji words each, ..., and H _nM.

The syllable length of the current syllable processing section is n and the maximum syllable syllable length is M
If _n is the maximum number of delimitable types, U _n = 2 ^n-1 when n ≦ M, and initial values are U _M = 2 ^M and U _M-1 = 2 when n> M. ^M-1 ……,
Assuming that U ₁ = 1, U _n = U _n-1 + U _n-2 +... + U _{nM. For} example, when M = 4, n = 10, U ₁₀ = 401. In actual writing language, in sound frequently treated section of the n syllable length, since not all words sound from R _n1 to R _nM actually exists, the value of U _n is usually less than the. However, u also M when also large it is clear that a large value U _n is considerably. It takes a lot of time to calculate the frequency sum for each of these delimited types, to select and determine the type with the minimum frequency sum, and to process the phonetic kanji conversion. The processing becomes difficult in a short time between one syllable input and the next syllable input. The successive tone frequency processing method described herein eliminates such difficulties.

The above U _n speech-segment delimiters are represented by R _n1 , R _n2 , ...
…, _Divided into M sets of R _nM . The minimum frequency sum P _n1 , P _n2 ,
..., P _{nM It} is not necessary to obtain a large number of sums for each set by comparing the frequent sums, where n is n−1, n−2,.
−M, the total minimum frequency sums P _n−1 , P _n−2 ,..., P _nM already obtained by the processing in the previous stage are In _n1 and In _n2 obtained at the current n. ,..., And _InM , each of which is easily obtained by the following equation.

When m is the number of syllables of the last speech in the speech segmentation type, When n is in the range of 1 ≦ n ≦ M, n−1 m for m ≦ n−1 of
The P _nm calculated by _{P nm = P nm + I nm} , for the one m of m = n, the one of the P _nm calculated by _{P nm = P nn = I nm} , after all P _n1, P _n2 ,..., P _nn are obtained as n minimum frequency sums. When n is in the range of n> M, m m P _nm for 1 ≦ m ≦ M m _Is calculated by P _nm = P _nm + I _nm , and finally, the M minimum sums of P _n1 , P _n2 ,..., P _nM are obtained.

A frequency sum having a minimum value is selected from the above P _{n1 to} P _nm by selecting one of the m values, and is selected as a total minimum frequency sum P _n . m
The alternative is executed by m-1 alternative selection processes.

When only one P _n is selected from P _n1 , P _n2 ,..., P _nm , the word segmentation type having the minimum frequency sum is also selected at the same time. Assuming that the number of syllables of the delimited end speech having P _n is mo, of the above m of P _n1 , P _n2 ,..., P _{nmo of} the subscript mo is P _n
It was selected as. That is, in the n-syllable syllable frequent processing section currently being processed by the above-mentioned m-choice selection, (1) the value P _n of the minimum frequent class sum of the optimal syllable segmentation type; The number of syllables mo is determined. By the way, _Pn calculated _Pn = _Pn-mo + _Imo
P _n-mo in the equation has already been obtained before n, where n is n-mo. That is, the total minimum frequency sum P
Similarly, the number of syllables of the last speech in the _n-mo speech segmentation type is known. If the processing stage is considered up to n = 1 by this logic, the operation of the true tone processing method can be reduced to the following two points.

(1) Each time a new syllable is input and the syllable syllable syllable length is extended by one syllable, the syllable length mo of the last syllable in the optimal syllable sequence indicating the minimum frequent class sum P _n is determined to update the optimal syllable delimiter ( Input tracking action).

(2) The optimal phonetic sequence is the optimal phoneme of the vocal phrasal whose syllable length is short by mo and the phonetic sequence obtained by adding the phoneme of the syllable length mo to the end, which was obtained before. The frequent syllable sum P _n is calculated as the minimum frequent syllable sum of syllable phrases with short syllable length by mo
This is the sum of the frequencies obtained by adding the frequencies of the words of mo (sequential action).

The processing required for the above-described input-tracking sequential speech segmentation function is completed by a frequency search of M times or less for obtaining P _n and mo, an integer addition, and a binary comparison of M-1 times or less. However, it is always the same regardless of the syllable input ordinal number or syllable syllable length n. In other words, the algorithm for the optimal word-sound segmentation processing by the true frequent processing method is not only very simple, but also the processing time following one syllable input is always the same regardless of the length of the syllable string to be processed. This is a great advantage in practical use.

When the processing related to the nth input syllable in a syllable is completed, the phonetic kanji conversion is performed by the following kanji character string addition formula using the mo obtained together with the result P _n of the above-described optimal syllable delimiter processing. Is done.

K _n = K _n-mo + H _mo ... (K) The above calculation is based on the kanji character string K _n-mo already obtained when n is n _-mo. The most probable kanji word H searched from the dictionary with the word sound R _n-mo as the heading
fulfill the function of connecting the _mo, most確漢string K _n is, n the new syllable input is determined sequentially in time proceeding 1.

FIG. 2 shows a sequential process for obtaining P _n from P ₁ to P ₂ when M = 4 and n is 1 to 7, an optimal word-segment separation sequential process for obtaining K _n progressing simultaneously, and a word-phone kanji. The algorithm A of the conversion processing is shown. This algorithm is used in Example A in FIG. 5 described later. K _n is for one syllable input, first to prepare the K _nm of the M (when n ≧ M), and K _n to choose one of the best in imitation from among them in the selection of P _n There is no necessity, as shown in FIG. 5, one K _nm may be calculated and obtained for each alternative selection of P _nm . This algorithm uses M =
This is useful for the second frequency processing.

The third figure, with M = 4, from P ₁ in the case of n = 1 to 7
Use mo determined for each processing of each P _n to P _7, 1
The algorithm B of speech-sound optimal delimiter sequential processing and speech-kanji conversion processing for finding K _n by executing H _mo search and expression (K) character string addition only once for syllable input is shown below. This algorithm is the simplest and clearest, and is useful for frequent processing in Japanese / Korean word processors requiring M of 3 or more. This algorithm B is used in the second embodiment shown in FIG. 7 described later.

When a word R is not in the dictionary, and thus neither the frequent class I nor the kanji word H exists, there is naturally no related frequent class sum. In that case, the corresponding alternative selection is unnecessary, and is omitted in the processing flow. In the actual tone frequency processing according to the present method, there are many non-existent words in Japanese, Chinese, and Korean, especially in words having a long syllable length. The number of syllables whose frequency is statistically high is overwhelmingly high for 1 and 2 syllables in Chinese, but different for Japanese and Korean. These circumstances, which have individuality depending on the language, are points to consider when designing processing software.

In any language, there is an upper limit on the syllable length of speech sounds (word readings). That is M. M may vary depending on the definition of "word" or the design policy of the dictionary, but is it about 6 in Japanese, 4 in Chinese, and 4 in Korean?

When n is smaller than M, there are only n P _nm .
Therefore, the procedure for obtaining P _n requires reduction. The maximum value N (maximum syllable length of one vowel phrase) and the maximum value of M differ depending on the language. The rate at which the above reduction procedure is used during the tone processing may also vary from language to language. In Chinese, this rate is quite high. This is a point to be considered when designing software.

The frequent class I (positive integer variable) and the kanji word string H (character variable) are updated each time a syllable is input. Therefore, the variables needed for these are actually I _{1 to} I _M for I and H.
It suffices to provide _M for each of H _{1 to} H _M , and it is not necessary to place M for each n as shown in FIGS. 2 and 3. Minimum sum P _n1
With respect to .about.P _nM, for the same reason, it is sufficient to secure only M pieces.

For the n-th syllable input, P required for calculation of P _{n1 to} P _nM is M of P _{n−1 to} P _nM . Therefore, P
And in the process of sequentially obtaining the K, nM-1 pieces of old data from P ₁ to P _nM-1 is unnecessary. Similarly, also in the top確漢string processing, data necessary for the sequential processing requires only the M K _n-1 ~K _nM.

Japanese Patent Application No. 63-1050 already filed by the present inventors
In Japanese Patent Application No. 30 and Japanese Patent Application No. 63-172163, regarding "nodes", "when there is no speech that crosses a virtual point between two consecutive syllables, the node is connected. The frequent processing section is cut off at the node. . " When a “node” is found in the middle of the syllable processing, n is reset to an initial value (1 in the present description), and the process proceeds assuming that a new syllable processing section starts from the next syllable.

The definition and utility of the "breakpoint" in the above-mentioned invention of the application are invalid when M> 2.

As described above, the conversion of vowel-word-separated kanji by sequential processing can be practiced by a very simple algorithm even when N and M are considerably large.

3.5.3 Frequent sum sum tying process In the process of tone frequent processing, when the frequent sum and the total frequent sum are executed by alternative magnitude comparison, the frequent sum is a positive integer. A case arises where the two are equal. The following measures are available.

(1) An algorithm is set so as to select the one with the larger number of syllables of the last speech from the two speech strings.

(2) An algorithm is set so as to select one of the two speech strings that has a smaller number of syllables of the last speech.

Which of these should be chosen depends on the characteristics of the target language. In the embodiment described later, (1)
The method is adopted.

3.6 Example 3.6.1 Configuration example of frequent-word-speech-separated sequential kanji conversion processing Fig. 4 shows the "phonetic phrase" based on the principle of "action" in 3.5.
FIG. 1 is a block diagram showing a configuration of an embodiment of the present invention in which an optimal speech sound separation is performed on a utterance and a speech sound selected by the speech sound separation is converted into a kanji word. Referring to FIG. 4, the speech separation and speech / kanji conversion method according to the present invention will be described in detail.

In FIG. 4, the initial syllable length of a word sound is M, and the syllable input number n in the syllable processing section is 1 as an initial value, and the processing in the case of n <M and in the case where it does not exist in the dictionary to be handled is omitted. In addition, no consideration is given to efficient memory use and processing order.

The function of each means block in FIG. 4 is as follows.

1: Syllabic input means Input "reading" from a keyboard or the like. In particular, the "syllable" input means is that Chinese, Japanese, and Korean, which are the target languages of the present invention, have a pronounced syllable structure in all phonemes (Chinese kanji reading is one syllable; Japan The word is one mora
This is because even if the syllables and Hanguls are perfect syllabic characters themselves, even if they are input in Roman characters, it is highly efficient to store the headings R of the phonetic sounds in the dictionary in syllable units.

3: Dictionary The following R, I, H subscripts indicate the number of syllables. The vocabulary of the syllables 1 to M is stored in the dictionary.

(1) Speech headings R represented by syllable strings (subscripts indicate the number of syllables. The vocabulary of speech sounds of 1 to M syllables is stored in the dictionary) (2) Frequency I of each speech sound (3) Each speech sound This is storage means for storing data such as the same kanji vocabulary H (kanji code expression) for each sound.

2: Each time the syllables R _n is input from the speech generation means syllable input means 1, 2,
The syllables R _n-1 , R _n-2 , ... already entered and stored in ₂
…, Concatenate R _{n-M + 1} and R _n , R _n1 , R _n2 ,..., R _nM are generated and stored.

4: Speech frequency kanji data search means 3 The speech sounds R _n1 , R _n2 ,.
.., _RnM is searched for the next data, and stored.

(1) Each frequency class I _n1 , In ₂ , ..., I of speech sounds R _n1 , R _n2 , ..., R _nM
nM (2) Kanji words H _n1 , H _n2 ,… with speech sounds R _n1 , R _n2 , ..., R _nM
…, H _nM (Select the most probable of the same phonetic words by self-learning such as the recent access priority formula.) 5: Minority class sum generation means 2 and frequency classes In ₁ , In ₂ , ……, I Given _nM , for m from 1 to M, the M minimum frequency sums P by the formula P _nm = P _{n (nm)} + I _nm
Calculate and store _n1 , _Pn2 ,..., _PnM .

6: Kanji string generating means 2 receives the kanji words H _n1 , H _n2 ,..., H _nM from M, and, for m of 1 to M, M kanji strings by the formula K _nm = K _{n (nm)} + H _nm
K _n1 , K _n2 ,..., K _nM are obtained and stored.

7: Total minimum frequency class selection means 4 receives the minimum frequency class sums P _n1 , P _n2 ,..., P _nM from among them, selects the one with the minimum value among them, and stores it as P _n .

8: The most probable kanji string generation means 5 receives the kanji strings K _n1 , K _n2 , ..., K _nM from among them, and among them, P _n
Those corresponding to the storage as the most確漢string K _n.

9: Kanji string display means _Kn is displayed on the display.

12: Kanji word sound judging means When a symbol or the like other than Kanji is input, it is displayed on the display as a non-Kanji character, and the syllable input number n is reset to 1 to close the frequent syllable processing section. Prepare for the start of

3.6.2 First Embodiment FIG. 5 shows one syllable speech, two syllable speech, or four syllable speech input in units of syllables, which is the core of the present invention. 5 is a flowchart showing a processing procedure in a speech-sound separating and speech-kanji conversion apparatus (first embodiment of the present invention) using a sequential kanji conversion method. The algorithm in this example uses FIG. 2 described above.

In this figure, the initial values of the syllable syllable length are 1 to 4 and the syllable input number n is 1.

[Explanation of variables]

 n Positive integer variable.

The initial value is 1. Syllable input number, reset to 2 by node detection. The reason why 1 is reset to 2 will be described later.

m Positive integer variable. In FIG. 5, 1-4.

Subscripts of various variables are used in combination with n. This is the number of syllables of the last speech in the speech separation type.

R Character variable.

 One syllable entered each time.

R _m character variable. In FIG. 5, R _{1 to} R ₄ .

It is the same as R _{n1 to} R _nm described in Section 3.6.1, and it is not a two-character character variable like R _n1 because these variables are updated for each syllable input. is there.

The initial values of R _{1 to} R ₄ are all non-strings (nstr). Immediately after inputting R, the initial value of the subscript m is set to 4, and the calculation of R _m = R _m-1 + R is repeated while reducing m by 1 until m reaches 2, (R _m-1 is n Is one stage before), R ₄ , R ₃ , and R ₂ are obtained. Finally, for m = 1, R ₁ = R
Then, R _{1 to} R ₄ are obtained. At this time, when m> n, the above calculation is skipped, and only the processing of m = m-1 is calculated. When n is smaller than 4, all logically unnecessary R _m remain nstr. In the flow diagram of Figure 5, this processing is executed by R ₁ to R ₄ generation block (1). Note Details of R ₁ to R ₄ generating block is shown in Figure 6.

I _m positive integer variable. In FIG. 5, I _{1 to} I ₄ .

Frequencies of each of the speech sounds R _{1 to} R ₄ . For each syllable input obtained from the dictionary and the R _m heading. In FIG. 5, the dictionary search process is not shown. Note that even if there is no speech for one syllable reading, it is assumed that speech, I, and H exist formally. This is called an imaginary word sound. In this case, however, I = 32 is set so that it is irrelevant to the tone frequency processing. It is not necessary to define the I of two or more syllable speeches without real speech.

H _m Kanji variable. In the FIG. 5 H ₁ to H _4.

Speech R ₁ to R ₄ each in the converted kanji word homophones in the those likely currently most reliable. For each syllable input, obtained by subtracting the dictionary and the R _m heading (omitted in the figure). Kanji language H ₁ for 1 syllable words sound without actual words sound has been left "reading". That is, H ₁ = R ₁ . It should be noted that in Japanese and Korean, the actual words sound when the output representation Yu is "pseudonym" and "Hangul", and H _m = R _m. For example, the Japanese particle "no" is I ₁
= 5, H ₁ = “of”. The imaginary word “n” is I ₁ = 32, H ₁
= "N". Migooto "from" is _{"to" I 2 = 8, H 2} =, homophones H ₂ is a "shell", "sky", "Tang". Regardless of whether H is a kanji or a kana or a mixture of both, in the present invention, all Hs are called "kanji" in a logical form.

PN _m positive integer variable. In the FIG. 5 PN ₁ to PN _4.

The minimum sum of the respective frequencies in the phonetic sequence of the last syllable having one to four syllables. By calculating the _{PN m = P nm + I m} , determined previously by adding I to the precalculated P. Updated for each syllable input.

P _A positive integer variable.

PN ₁ and PN ₂ and compares P _A = PN ₂ if _{PN 2 ≦ PN 1, PN 2} > is P _A = PN ₁ if PN _1. If there speech of R ₁ and R ₂ are both P
_A is not determined, but no countermeasures are required in Japanese because no two or more non-sounding readings such as "n" or "tsu" do not continue. On the other hand, there is no syllable without speech in Chinese. P _A is updated every syllable input.

K _A character variable.

If PN ₂ ≦ PN ₁ in selection of the above _{P A K A = K n-} 2 +
If H ₂ , PN ₂ > PN _1, then K _A = K _n-1 + H ₁ . The selection of K _A
As already described in the "action" of 3.5, choose the kanji column with Kanji word H to equalize P _A tail speech and number of syllables of the total minimum frequent class sum resulting separated type in singles trailing K _A It is. K _A is updated each time a syllable is input.

P _B Positive integer variable.

PN ₃ and PN ₄ are compared. If PN ₄ ≤ PN _3, P _B = PN ₄ and if PN ₄ > PN _3, P _B = PN ₃ . When there is no speech both of R ₃ and R ₄ and P _B = 255. This is because the meaningful maximum value of I is 18, and the frequent syllable processing section having the length of 14 syllables is inexperienced.

K _B character variable.

In selection of the above P _B, if _{PN 4 ≦ PN 3 K B =} K n-4 +
And H _4, PN _4> PN ₃ if _{K B = K n-3 +} H 3. When P _B = 255 is the K _B = nwrd.

P _n positive integer variable. In Figure 5 P ₀ to P _n.

The total minimum frequency sum at that n. It is determined by M-choice selection of PN _m . In the example of FIG. 5, since m = 1 to 4, in order to obtain one P _n , the alternative selection is 4-1 =
You need three times. The process of obtaining P _n is the third selection. P _B ≦ P _A if P _n = P _B, a P _B> P _A if P _n = P _A.
P ₀ = 0 all the time.

K _n character variable. In FIG. 5, K _{0 to} K _n .

The kanji word string by the optimal word separation at that n. In parallel with each stage of one alternative selection process in which P _n is required
The process of obtaining K _n proceeds. If P _n = P _B , K _n = K _B , P _n = P
If _A, then K _n = K _A. K When the nwrd both of R ₂ and K _{B
Let n} = nwrd. This indicates that the node immediately before the currently input syllable R is a node. K ₀ is a non-string (nst
r).

[Explanation of process]

Fig. 5 shows a Japanese word processor that reads in kana or romaji and automatically separates words and then converts them into kana sentences with kanji characters. Korean word processor, which converts to
And romaji A main part of an embodiment of a sequential word-segmented-word-to-kanji conversion apparatus suitable for any Chinese word processor that automatically converts words into kanji sentences after inputting with sounds will be described. However, the maximum number of syllables of speech sounds is four. Hereinafter, the flow of the processing will be described.

Immediately after the start, to 1 syllable input number n, the initial value of R ₄ from four speech R ₁ in the non-string (nstr), and P ₀
Set the initial values of K ₀ to 0 and nstr, respectively.

 Enter one syllable or symbol R.

0: The non-character processing block processes and displays an input other than a syllable having a reading, and performs “free separation” (the present invention originally intended to eliminate the need for a conversion operation using a human conversion key. This does not preclude the manual partitioning. This is referred to as "free partitioning." When a key is pressed, the display returns to the display without display.

₍₁₎ R 1 ~R ₄ generating block, shown in detail in Figure 6.

Hereinafter, speech of Shikikyu I _m, such as Chinese language H _m variables (m =. 1 to
4) comes out Some of the description, these are retrieved from the dictionary as a heading reading R _m. If R _m is present, find these values. In the case of nothing, R _m = nwrd (non-word). However, FIG. 5 omits the search process.

1. First-stage alternative selection block: Investigate the presence of one-syllable speech and two-syllable speech at the end of the speech sequence, and determine which speech sequence has the optimal speech segmentation. Then, kanji conversion is performed on the word sequence.

Process when the R ₁ and R ₂ are both free, although not necessary, as described above, in the fifth view is written together with for convenience of explanation.

Find the frequent sum PN ₁ of the phonetic sequence ending with one syllable word,
Prepare for PN selection. When n = 1, the subscript n-1 of P =
Although it is 0, it was necessary to set P _n-1 = P ₀ to 0 in preparation for it.

When speech of R ₂ is a free utilizes the node determined as R ₂ = nwrd.

The end is previously obtained the Shikikyu sum PN ₂ of the word sound column of the second syllable words sound.

Compare PN ₁ and PN ₂ and select the smaller one. If the values are the same, PN ₂ is selected (there is no statistical basis).

Determining the P _A and K _A in accordance with the results. The speech sequence and the kanji sequence of the speech delimiter of the speech sequence having a small frequent sum are determined. A kanji string having the same word separation as that of the minimum frequency sum is determined here for the first time.

2. Second-stage alternative selection block: Checks for the presence of 3-syllabic and 4-syllabic speech at the end of the speech sequence, and determines which speech sequence has the optimal speech segmentation Then, kanji conversion is performed on the tested word sound sequence.

Send to '3 syllable speech R ₃ is a process when chromatic'.
When there is nothing, send to '.

'Seek Shikikyu sum PN ₃ of the word sound column at the end are three syllable words sound.

'4 when syllable speech R ₄ is closed Handles' send.
When there is nothing, send to '.

'Ending seek 4 Shikikyu sum PN ₄ of the word sound column of sound words sound.

′ Compare PN ₃ and PN ₄ and select the smaller one. If the values are the same, PN ₄ wins (no statistical basis).

Determine the P _B and K _B in accordance with the result of '''. The speech sequence and the kanji sequence of the speech delimiter of the speech sequence having the small sum are determined.

'4 when syllable speech R ₄ is closed Handles' send.
When there is nothing, send to '.

'The processing is the same as', but since there is no speech R ₃ , no alternative selection between PN ₃ and PN ₄ is required and the processing is sent directly to'.

'When both R ₃ and R ₄ are absent, set P _B = 255 to ensure that P _n = P _{A in} the next selection of the third stage (a). 255 is used for both Japanese and Chinese, statistically significant frequency I
Is the maximum value of 18, and the fact that the length of the frequent syllable processing section from one node to the next node is 14 syllables is the result of experience (18 × 14 = 252). Also, set K _B = nwrd.
This is for use in the node determination routine.

3. The third stage alternatively Selection Block: and P _A of the optimum word sound string of words sound column at the end decided in alternative of the first stage is 1 or 2 syllable words sound, two parties of the second stage It compares the P _B of the optimum word sound string of words sound column at the end decided in alternative 3 or 4 syllables sound,
The one with the smaller value is finally determined as a speech sequence having an optimal speech segmentation. The frequency sum P _n and the kanji word string K _n are determined and stored for the next R input and processing.

A node refers to a point between two adjacent one-syllable speech frequency classes where all paths must pass through in a frequency network. There cannot be any more than two syllable speech that spans a node. The nodes separate the tone processing zones.
Is a node determination routine for a point between the current R and the R input immediately before. R ₂ = nwrd indicates that two syllables speech across the that point no, K _B = nwrd has no 3 and 4 syllables speech.

If the point is not a node, the process returns to the next R input routine after incrementing n.

When that point is a node, it is necessary to process the current R as the first syllable of the next syllable. In the preparation routine, R ₁ = R, R ₂ to R ₄ are set to nstr, I _{1 is set} to P ₁ , and H _{1 is set} to K ₁ . Then, to display the K ₁ of the first character first, after setting the n = 2,
Return to the R input routine.

As described above, it is possible to successively perform the phonetic-segmentation-word-to-kanji conversion for all syllables whose syllable lengths are 1 to 4.

In FIG. 5, the maximum syllable length (M in FIG. 6) of the speech to be subjected to frequent processing is set to 4. But in theory M
You can make it as large as you want. For example when the = 8, obtains a P _A from PN ₁ and PN ₂ with a first stage of the alternative,
Seeking P _B from the PN ₃ and PN ₄ in the second stage, the P _C from PN ₅ and PN ₆ in the third stage, obtains a P _D from PN ₇ and PN ₈ in the fourth stage, then P _A and P _B
From the P _X, determine the P _C and P _D and P _Y, thereby obtaining the P _n from P _X and P _Y in the final seventh stage. In parallel with the processing of these PN and P to go in search of P and K for each stage, reaching the last K _n. It should be noted that the algorithm in each stage
This is exactly the same type as the second and third stages in the figure.
The number of alternative selection is M-1, the processing time required to determine the K _n for one R input may likely roughly proportional to M when M is large. Even if the complexity of the frequent phone network increases as described above, the processing time only increases in relation to the first order of M, which is an advantage of the real phonetic word segmentation processing.

Sound frequent treatment of the present invention, regardless of the length of the sound Shikiku, time processing necessary for obtaining the K _n with respect to one of the R inputs,
It is always the same for one increase of n. This is another advantage of the present invention, which responds to the demand for reading-to-kanji conversion immediately following each syllable input.

3.6.3 Second Embodiment FIG. 7 shows a second embodiment. Compared to the first embodiment in FIG. 5, variables m _A , m _B , and mo (all positive integer variables) are added. These are the number of syllables of the last speech of the minimum speech separation type. The subscript A or B is the value of the first selection or the second selection, and mo is the value of the final selection.

The processing algorithm in FIG. 7 differs from the algorithm in FIG. 3 in the following points.

(1) K _A =
K _A is not determined by an expression such as K _n-1 + H ₁ . ,, ′,
As in 'like, obtains the m _A and m _B. As shown in FIG. 5, kanji word search and kanji character string addition are not performed.

(2) without asking K _n from P _A and P _B in the third stage alternatively selected,
Ask for mo only.

(3) 5 and 6 syllable speech R ₅ and R ₆ are added, it is performed speech kanji conversion by the longest match method for these. Next, only the longest-matching partition block at the fourth stage in FIG. 4 will be described.

(A) If both R ₅ and R ₆ are present, the processing flow is →→
→→ ₃
The syllable words survive the other words and are converted to kanji at ₃ . Kanji column K _n is, K _n-6 + H _6, and the Chinese language H ₆ of 6 syllables sound is determined as the last word of the kanji column.

(B) When R ₅ is present and R ₆ is absent, the processing flow is →
→→ ₃ next, mo = formula 5 with _{3 K n = K n-5 +} H 5
Next, kanji word H ₅ 5 syllables sound is determined as the last word of Kanji example K _n.

(C) R ₅ is without, when R ₆ is Yes, the process flow in →→→ _3, Chinese character word H ₆ of 6 syllables sound is determined as the last word of the kanji column.

When (d) R ₅ and R ₆ are free both, unless the node is detected, the process is direct from to, K _n for alternative selection of the third stage is the answer.

(4) Through the above processing, speech kanji conversion _is 1,
_It is performed for the first time in the final stage of the processing such as ₂ , _3, or the like. In the processing before that, only integer variable calculation processing such as addition of frequent class sum, comparison of frequent class sum magnitude, determination of the number of syllable syllables at the end of a word string, and the like proceed. Finally, only one phonetic-kanji conversion is performed. With this algorithm, the number of time-consuming dictionary searches is minimized and processing time is reduced.

As in the second embodiment, it is possible to perform the delimiter processing by connecting the vocal utterance delimiter and the longest match delimiter. The significance is as follows.

(1) In general, when it is worthwhile to put a long-reading word sound in a dictionary, it is limited to a case where the word sound is a frequent synthetic word or collocation. There are not many compound words of this kind. When those words are detected in the sentence, there is only one corresponding word, and the probability that the word is correct is almost 1.

(2) In the above case, the longest match delimiter is truly effective. In Japanese, for example, "sangiin" means "no matter what particle syllable before it,
There can be no other.

(3) In the Japanese language, the vocal utterance delimiter is effective for judging interference and duplication of one to four syllables on a speech network. The longest match method is particularly useful for separating long speech sounds.

In FIG. 7, the processing prior to the first-stage alternative selection block is largely omitted. Also, 1
Syllable words are written on the assumption that they exist for any syllable. This is to simplify the expression.

3.7 Effects of the Invention 3.7.1 Effects on Chinese Sentences One of the characteristics of Chinese sentences is that syllable lengths are short. In office sentences, one or two syllables make up 97% of all syllables. Second, a particular syllable word is often used. Third, three or four syllable words have few homonyms. Such a feature indicates that the frequent phonetic word-separation sequential kanji conversion method of the present invention is suitable for a Chinese word processor. Hereinafter, the effects of the present invention will be described with reference to example sentences. The frequent pronunciation class statistics are described in “Modern Chinese Product Indices” (ed.), Edited by Beijing Language Institute (1986) (total syllables F _t = 1,807,40)
5) was used.

[Example sentence] (Institutional application sentence, (From Draft Publishers, 1885, p.83) Oil is an important strategic commodity. Self-serving Since the oil crisis that occurred in the early 70's, there has been a rational use of petroleum-saving petroleum. Immediately, the quantity of heavy oil and crude oil burned off every year in Japan, the total production of occupied crude oil is around forty percent, and the use of some of them is unreasonable. Energy saving resources, essential high-strength compression baking oil, light spinning raw materials for further use of petroleum used, average production favorable refined oil, satisfying four-dimensional construction demand. For this reason, the directive was issued as follows: [Latin alphabet Sound] Shiyou shi zhongyao de zhanle wuzhi. Zicong seventy niandai chu fasheng shiyou weiji yilai, heli shi
yong he jieyue shiyou, yijing chengwei quan shijie
pubian guanzhu de wenti. Muqian, woguo meinian shao
dao de zhongyou he yuanyou shuliang hen da, zhan yu
anyou zongchanliang de hundred fenzhi forty zouyou, zhizhong xi
angdang yi bufen shiyong bu heli. Weile jieyue nen
gyuan, bixu dali yasuo shao you, shi shiyou geng duo
de yong zuo qingfang huagong yuanliao, bing shengc
han geng hao de chengpin you, yi manzu sihua jiansh
e de xuyao. weici, fabu ruxia zhiling: Fig. 8 shows the romanization of the above example An example of the effect of inputting a syllabary and then outputting it to a kanji sentence by the automatic vocabulary separation and sequential kanji conversion method of the present invention will be described.

In the figure, a pattern such as a two-stage brickwork composed of a few vertical lines connecting three horizontal lines and two horizontal lines in some places is a tone frequency network. FIG. 9 shows two parts of the row in FIG. 8 (all of which contain syllables).
Is excerpted. In FIG. 9, the left is a frequent phrase A, and the right is a frequent phrase B. 1A and 1B are phony networks in which speech and frequency are entered, and 2A and 2B are in which only the frequency is entered. 2A includes 5.3.1 of each of the 1,2,3 syllable words, and there are eight types of word segmentation from (1) to (8). The paths shown by bold lines indicate the delimited paths, and the white numbers on the right side of the network are the sums of the frequencies. Among them, the partition of (8) gives the minimum frequency class sum 30.

FIG. 10 shows the progress of each input of one syllable in the alternative syllable word segmentation and kanji conversion for the five syllable phrasal phrases A described above. Figure (1) the sound frequently networks fill Shikikyu, (2) shows the course of to seek the total minimum Shikikyu sum P _N at each stage of the N sequentially in the alternative expression. The optimal speech delimiter is determined at the same time as P _N. (3) is the progress of speech-kanji word conversion that proceeds simultaneously with the progress of the optimal speech-segment separation. The right-hand column shows an example of a kanji display method on a display when the syllable input is a double-hit type of a vowel (consonant) and a vowel (vowel). When the initial key is pressed, an uppercase initial is displayed, and when the final key is pressed, the temporary kanji is displayed at that position. This is a very easy-to-see display method.

Figure 8 uses the alternative word separation and sequential kanji conversion methods described in Figures 9 and 10 to convert the above example sentence (Chinese State Council directive) into romaji. This is the result of executing word separation and kanji conversion from the sound to the original kanji sentence. The conversion works only at each point immediately before the punctuation mark / kanji (Kanji is the direct key input without conversion). In the figure, the bold dashed path indicates “re-separation” for “different separation”.
Indicates a later pass. There is one for each. There are 141 kanji (syllables) excluding kanji. The difference between them is that ab / c is incorrectly a / bc and a / b is incorrectly ab. The fact that so-called sentence batch conversion is performed on 141 syllables and only such a division error occurs indicates that the method of the present invention is considerably effective. Note that the kanji words shaded in the figure indicate homophone differences. Although the punctuation is correct, there are a total of nine kanji that make the same phonetic mistake. This is not much.

As described above, the automatic word-separation and sequential word-kanji conversion system of the present invention provides an effective automatic word-segmentation result for a Chinese input-kanji conversion word processor.

3.7.2 Effect on Japanese sentence In Fig. 11 (a) to (c), a frequent network is created for Japanese sentence sentences in all kana books, The path of the speech punctuation is obtained, and the result from the punctuation mark to the next punctuation mark is collectively indicated. From FIG. 11, it can be seen how high the word separation accuracy of the phonetic method of the present invention is in Japanese sentences.

In order to perform phonetic word separation, speech frequency statistics are indispensable. Here, the National Institute for Japanese Language, ed., “Vocabulary survey of junior high school textbooks” (Hideei Publishing, 1986) (total syllables Ft = 45)
7,845). Based on the vocabulary and frequency in this document, each speech and its frequency were calculated, and the speech network shown in the figure was created. In the figure, hiragana on the network is a syllabic stick and numbers in the network are frequent. For understanding, an example of the processing according to the present invention is shown below in FIG. 14 (s is the number of syllables of speech).

As shown in FIG. 14, the optimal word separation and kanji string for this sentence example are obtained as follows. Optimum speech separation:
10 / id 15 (frequent sum: 10 + 15 = 25 is the minimum. ∴ solution is “ko / ido”. Kanji conversion selection: high latitude → light latitude → school latitude → high latitude. Thereafter “high” is “high” Conversion to.

The syllable string “Kouido” of (1) is processed by the frequent word utterance separation method.

The speech sounds that can be included in (1) are the seven types in (2). (3) is quoted from the vocabulary statistics having each word sound. The number to the right of the word is the word frequency in the statistical data.
The syllable frequency in (4) is the word frequency multiplied by the number of syllables that make up the word. Frequency (5) indicates the syllable frequency as the total number of syllables in the statistical data Ft =
Divided by 457,845. The frequency of (6) is obtained by converting the frequency to the amount of information.

For the syllable string (1), a speech network of (9) and a corresponding frequent network of (8) are formed using the speech (2). The path on which the sum of the frequent classes is the minimum is the optimal path, and the type and order of the speech on the path give the optimal speech separation of the syllable string (1). The paths indicated by thick lines in (8) and (9) are solutions.

FIG. 11 shows a frequent class network for the processing of the frequent sound method for example sentences. The thick line in the figure indicates the optimal path. No.
For all 495 syllables in Fig. 11, there are only two distinctions. The thick broken line shows the result of correcting the difference. The phonetic word sound segmentation process has a property that the segmentation process can be performed no matter how long a syllable string to be processed is. Therefore, it is essentially a processing method suitable for "all sentence collective kana-kanji conversion". However, the structure of the upper class network in FIG. 11 seems to be considerably complicated. Especially in Japanese, there are many syllables in speech sounds, so the frequency network is much more complicated than in Chinese.

The sequential speech segmentation processing method of the present invention is sufficiently effective even for such a complicated frequent network. No.
Fig. 12 shows an example of the progress of sequential speech segmentation, using sentence examples.

In FIG. 12, (1) shows a syllable string R of a sentence example.
(2) shows a frequent class network of a sentence example. In the network, a point where there is no speech sound over that point is a "node". In this example, there are six nodes from to.
Thereby, the example "sentence" is divided into five "syllable processing sections", and the syllable input number n starting from 1 is reset to 1 every time a node is detected. The vocal frequency processing, that is, the optimal speech separation, is performed for each phonological processing section, and the length of the syllable string of the phonological processing section does not become infinite.
In the syllable processing section, like the first section, one syllable ""
There are also short examples.

(4) shows a process of sequentially calculating the minimum frequency sum P _nm according to the algorithm of the present invention according to the syllable input number n, and then calculating the total minimum frequency sum P _n . Here, m is the number of syllables of the last speech in the speech sequence. In this sentence example, there is a case where the minimum frequency class sum does not fully exist at a certain n. For example, P ₄₁ , P
_42, there is a P _44, there is no P _43. The reason is that there is no word "dai". In the case of n = 2, P ₂₂
There is no P ₂₁ if there is. This is also because Japanese does not have the word "n". In Chinese, in principle, every single syllable has speech. Japanese is not so.

In Japanese, there are no monosyllable words such as "n", "tsu", "ni", etc. There are no polysyllabic sounds that begin with "n" or "tsu". Therefore, in the Japanese frequent class or speech network, the position occupied by these non-existing speech sounds is a “hole”. In the two-syllable word sound "Imi" ending with the fifth syllable in the fifth ward, the next word sound is "N", so "Imi /
N "is not satisfied, cause the P ₆₁ in the n = 6 is absent. In FIG. 12, shaded classes are existing speech sounds that become invalid as a result of the lack of these speech sounds. The tone frequency processing algorithm of the present invention, for the above-mentioned speech lack phenomenon,
It should work effectively.

The thick-line paths in each frequency network in FIGS. 11 (a) to 11 (c) are paths for optimal speech separation processed by the frequent processing algorithm of the present invention. In the following, the effect of the present invention on the sound segmentation of Japanese sentences will be evaluated based on the results obtained in FIG.

(1) “no”, “ha”, “ga”, “wo”, “ni”
Particles such as "to" are almost 100% surely separated. Immediately before and after these are nodes at a high rate.

(2) In the conventional "longest match method", a delimitation frequently occurs in one kanji word + two kanji word collocations, for example, "ki / kinzoku → precious metal", "so / kinzoku / sei → total metal", " Now /
Ryusan → concentrated sulfuric acid, etc., succeeds with ease. These examples are illustrated in [C] to [K] in FIG. 11 (c) and in sentence example 6 in FIG. 11 (b).

(3) As described above, the accuracy of the present invention in speech separation processing for Japanese sentences is higher than that in Chinese. The difference is only two places and six syllables in all 495 syllable sentences. Moreover, the reason why the word “ha / chu” in (1) of [B] in FIG. 11 (c) was mistaken for “hachu” is that the source material of the speech frequency was based on the vocabulary frequency statistics of junior high school society and science textbooks. This is probably because the frequency of the infrequently-speaking word sound "Hachichu-Reptile" is too high.
With more universal statistics, it is presumed that there is no difference here. If this is the case, the difference in the whole text example is that only the three syllables at the beginning of the fourth tier of sentence example 3 in (b), where “Kishi / be” is wrong with “ki / shibe”, have an erroneous delimitation rate of Only 3/495 = 0.6% in syllables. The speech separation ability of the present invention can be said to be epoch-makingly effective.

(4) The basis of the idea of the present invention lies in that the function of "reading kanji conversion" is divided into "sequential word separation" in the first stage and "word kanji conversion" in the second stage. The present invention relates to the first stage. The function of "full-text batch conversion" of "there is no need for a conversion key as an ideal type of word processor" cannot be achieved at all unless there is a condition that the function of automatic word separation is excellent. The present invention has an effect of sufficiently satisfying the conditions and advances one step toward the goal of an easy-to-use Japanese word processor.

(5) The automatic sequential speech segmentation function of the present invention
After performing speech separation with an accuracy close to%, word separation processing is performed on each separated speech using the grammar rules between collocations, and finally the correct conversion rate of Yomi-Kanji conversion Should be raised.

(6) The syllable separation method of the present invention follows the syllable input one after another, and
It is the one that is most probable from the standpoint of information theory, and is executed each time a syllable is input. Thus, the word separation system of the present invention is essentially a sequential process that is performed in immediate response to syllable input.

3.7.3 Effect of Korean on Hangul sentences with Chinese characters As described in section 3.2.3 above, Korean and Japanese are both linguistic linguistics,
They are very similar to each other, such as grammar and the number of words derived from Chinese.
Although the present inventor has not examined the effect of the present invention on Korean sentences at the same level of detail as in Japanese and Chinese, based on natural common sense in linguistics, the present invention is based on the Korean kanji hangul. Of course, it is also effective for word processors.

3.7.4 Effects of the present invention on input processing of languages using kanji (1) There are three languages that use kanji in the currently written language: Japanese, Chinese, and Korean. Kanji is used partially or completely in sentences in these languages. The use of ideographic kanji has a gap between "reading" and "notation". This is because the appearance of homonyms is inevitable in kanji words.

(2) The three words of Japan, China, and Korea happen to be written in syllable units. One Chinese kanji reading is one kanji
It is a syllable. One Korean Hangul reading is one syllable, and the Korean Kanji sound is one syllable. You can write Japanese kanji with one-syllable one-syllable kana. Therefore, the three words CJK are
It has a common feature that it is better to input in syllable units.

(3) When three Japanese, Chinese, and Korean words are input in syllable units, they must be converted to kanji characters before they can be converted into sentences. At that time, the sentence is delimited in units of something, and the kanji conversion is performed for each unit. In Japanese and Korean, the unit is “bunsetsu”. Both are "bunsetsu"
Starts with kanji and ends with kana or Hangul. The “space” in Korean Hangul sentences is placed for each phrase. Therefore, it is required to automatically delimit kanji words within a phrase. Chinese has no concept of a clause. There is no habit of “separating” each word, so it is difficult to divide a Chinese person into words if they are divided into words. Therefore, in Chinese reading input processing,
There is a need for a technique for automatically performing word separation and reading kanji conversion in a so-called whole sentence package. If the technology of full-text batch delimited kanji conversion effective in Chinese is invented,
This technique should be effective for the purpose of reading and kanji conversion in Japanese and Korean.

(4) If the "rule of minimum linguistic information entropy", which is the basis of the present invention, is specifically expressed, "a sentence is constructed such that the sum of the information amounts of the readings of each word in the sentence takes the minimum value. Yes ". The reason for the establishment of this rule is that when writing a sentence of a fixed length, the condition required for the information transmitted in each section of the sentence is not a large amount of information, but a small amount, that is, clarity. There is ". This requirement should be the same for written and spoken sentences.

(5) In the "phonetic word sound separation method" at the center of the present invention, the amount of information of a word is obtained from the statistical frequency of "reading" of the word. The definition of the word reading frequency I = int (−log ₂ p) indicates that the frequency I is the information amount itself. The fact that a sequence of phonemes that minimizes ΣI in a sentence provides the best phonetic separation for that sentence has been empirically established for many example sentences in Chinese and Japanese. The practical effect of the present invention comes from the correctness of the above theory.

(6) When the frequent phonetic processing section is long and the speech network is complicated, there is a concern that the amount of processing for finding the optimal phonetic speech segmentation is complicated and large, and the practicality is lost, no matter how valid the theory of the phonetic frequency method is. The method of sequentially finding the minimum frequent class sum and the optimal word-sound delimiter, which is a feature of the present invention, following the syllable input is not affected by the length of the frequent phonetic processing section. Even if the upper limit of the syllable syllable length is large, the time required for the automatic syllable segmentation processing is not significantly increased.

(7) In Japanese word processors, full-text batch kanji conversion is a technology that has always been desired for the purpose of popularizing operations, but a truly practical and highly accurate technology has not yet been developed. The automatic vocabulary separation according to the frequent method according to the present invention has the effect of enabling whole sentence kanji conversion in a practical dimension, and at the same time, enabling kanji conversion immediately following syllable input.

(8) The functional requirements required of a Chinese word processor are severe requirements for automatic word separation. The present invention answers that.

(9) For Korean word processors, the present invention is effective for the automatic division of kanji and Hangul in Korean phrases that have already been segmented in paragraph units, and for the conversion of long kanji strings to kanji. Will be considered.

[Brief description of the drawings]

FIG. 1 is a diagram showing a speech segmentation type tree structure including speech sounds of 1 to N syllables, and FIG. 2 is a diagram showing a speech-kanji conversion process executed in synchronization with a sequential process for obtaining a minimum frequency sum and an optimal segmentation type. Diagram,
FIG. 3 is a diagram showing the syllable number kanji conversion processing by mo after calculating the syllable number mo of the last utterance of the minimum frequency sum and the optimal delimiter.
The figure shows a block diagram of the tyrannical word-separated kanji conversion device.
Figure sequential diagram showing a first embodiment of the speech separator and speech kanji conversion apparatus, Figure 6 shows the details of the R _m generated and the node processing reading speech figure 7 Figure sequential speech separator and speech kanji conversion FIG. 8 is a diagram showing a second embodiment of the apparatus, FIG. 8 is a diagram showing an example of automatic syllable-separated sequential kanji conversion of a Chinese sentence by the phonetic method, FIG.
FIG. 10 is a diagram showing an optimal speech separation of a speech network, FIG. 10 is a diagram showing an example of alternative sequential speech separation Kanji conversion, FIG. 11 is a diagram showing an example of speech separation of a Japanese sentence by the phonetic method, FIG. 12
The figure illustrates the sequence of sequential speech separation of Japanese sentences by the frequent method, FIG. 13 illustrates the correspondence between Korean and Japanese, and FIG. 14 illustrates the application of the present invention to Japanese. It is a figure which shows the example of the obtained speech separation.

Claims (1)

(57) [Claims] In a sentence of a language using kanji, a syllable string obtained by sequentially inputting syllables using phonetic characters as a unit is divided into syllables, and an optimal syllable string is sequentially determined. In the syllable-input-sequential-sequential-separated-Kanji-sequential conversion method that obtains the most probable Kanji-word sequence, the phonetic conversion is performed sequentially for each individual sound in the word-sound sequence. reading a single speech, the statistical frequency of occurrence of individual speech used sentences of the language is f, and s the syllable length of word or sound, the total syllables total of all speech in speech statistics F _t when the, the Shikiritsu p of each word sounds and p = (f × s) / F t, the Shikikyu I of each word sounds _{I = int (-log a p)} , but to an integer as a = 2, the languages When entering a continuous phonetic sequence in a sentence,
A syllable string from the first syllable at the beginning of the syllable string to the n-th syllable that was recently input is regarded as a frequent phrasal phrase, and the frequent phrasal phrase is subjected to a sequential speech / separation / sequential kanji sequential conversion process at the most recent time. When the sum of the frequency of each speech is defined as a frequency sum, the speech of 1 to M syllable lengths (where M is an integer of 2 or more) which is statistically significant in the language is used as a heading, and the frequency of each speech is stored A vocabulary lexical class dictionary, and a vocabulary kanji vocabulary dictionary in which kanji homonyms reading the vocabulary words are stored in the form of kanji character strings, with each vocabulary as a heading. Each time it is entered, each type of speech is searched for in the dictation dictionary, and each vocabulary and each frequency are searched in the dictation dictionary. Sentence to the optimal frequency sum successive calculation means described in the next section For the search means and the optimal frequency class sum sequential calculation means, the syllable input number in the syllable is n (= 1, 2, 3,..., N)
, R _n1 , R _n2 , R _n3 ,..., R _nM , respectively, the M words having the length of 1 to M syllables ending with the n-th syllable input recently. The classes are _In1 , _In2 , _In3 , ..., _InM , respectively. The syllable segmentation type of the type that can be maximized in the syllable string of n syllables length is the syllable length m of the last syllable. 1,2,…
…, M where the readings are R _n1 , R _n2 , R _n3 , ……, R _nM respectively
And the minimum frequency sum of each of the M sets is P _n1 , P _n2 , P _n3 ,
...... further the P _n1 and _{P nM, P n2, P n3} , ......, the optimum Shikikyu sum Pn to the smallest value in P _nM, by sequential syllable input, n is 1 Increment by 1 The P _n
With respect to seek, when n is in the range of 1 ≦ n ≦ M, when n is 1, the P _1, calculated by _{P nm = I 11 = P 1} , n 1 <n ≦ against the M when in range of M, relative to m ≦ n-1 of the (n-1) m, (n-1) and P _nm calculated by _{P nm = P nm + I nm} , 1 single m of m = n , One P _nm is calculated by P _nm = P _nn = I _nm , and finally, the n minimum frequency sums of P _n1 , P _n2 ,..., P _nn are obtained. On the other hand, when it is in the range of M <n, M minimum frequency sums are calculated by P _nm = P _nm + I _nm for M m of 1 ≦ m ≦ M. As a result, P _n1 , P _n2 , ……, P _nM is calculated as the minimum sum of the most frequent syllables. After all, the minimum sum of the most uttered speech segmentation type for each of the n sets or the M sets is m syllables more than the current syllable input number n. Already obtained and stored by the processing immediately after the previous syllable input To each of the n or the M minimum Shikikyu sum P _nm that, the minimum Shikikyu sum sequential calculation means for calculating by adding the each of the n or the M Shikikyu I _nm by the currently searched, one syllable Is received, the values of the n or M minimum frequency sums obtained by the minimum frequency sum sequential calculation means are received, and a magnitude comparison selection of n or M alternatives is received. The minimum value P among the n persons or the M persons
seeking _nm as the optimal Shikikyu sum P _n, optimum Shikikyu which stores the value of the P _n, obtains a unique word sound column at the end of the speech R _nmo and number of syllables mo of word or sound with該頻class sum simultaneously Japanese _language separation type selection means, receives the speech R _nmo determined by the optimal frequency Japanese style division type selection means, and, using the R _nmo as a heading, of the same-sound kanji word of the syllable number mo ending with the currently input syllable , _{Read the} most probable kanji word H _nmo from the phonetic kanji word dictionary,
The most probable kanji word search and conversion means to be sent to the most probable kanji string calculation means of the next section, the mo obtained by the optimal frequency class sum delimitation type selection means, and the most probable kanji word search and conversion means obtained by the preceding section Kanji word H
_The present most probable kanji character string K _nmo is obtained by adding a character string of K _nmo = K _n-mo + H _nmo every time n is incremented by 1 each time a syllable is input, in _response to _nmo. Means for calculating the most probable kanji character string as an output of the means. In summary, when the language has speech sounds from one syllable to M syllable, in a sentence of N syllables, let the syllable number be n,
Is set to 1 and is sequentially input to the syllable syllable. Each time n increases by 1, M which may be present at the end of the sentence
(M when n ≦ M) and m syllables are obtained, and up to M syllables from the beginning of the sentence to the syllable immediately before the last syllable are obtained. To each of the frequency sums of the optimal speech sequence, the frequency of each of the up to M ending speech words is added, and eventually a frequency sum of up to M speech sequences having a length of n syllables is obtained. The only speech sequence having the smallest frequency sum in the sum is the speech sequence of the current optimal speech segmentation, the last speech of the speech sequence is determined as the current optimal speech, and the speech is converted to kanji. A kanji string is defined as a last kanji string, and a new kanji string obtained by connecting the last kanji string to a converted kanji string for a known most probable vocabulary string prior to the utterance is obtained.
A kanji sequence sequentially output a new kanji string each time the kanji is incremented by one.