JPS59132039A - Evaluating method of kana character string - Google Patents

Abstract

PURPOSE:To make the Kana (Japanese syllabary)-Kanji (CHinese character) conversion easy, by using the likelihood of plural Kana character candidates, which are attained on a basis of the pronunciation of a single syllable unit, and connection information among plural single syllables to attain m-syllable Kana character strings and selecting the Kana character string having a higher priority. CONSTITUTION:Characters are pronounced in single syllable units and are inputted to a voice input device 1. The voice input device 1 outputs corresponding Kana character candidates and their likelihood for every inputted single syllable and stores them in a storage device 2. A connection information storage device 3 stores connection information of Kana characters of the single syllable. An operating device 4 attains priorities of partial Kana character strings in accordance with Kana character strings in the device 2 and connection information by operation and selects, for example, 5 Kana character strings having higher priorities and stores selected them in a storage device 5. Five Kana character strings having higher priorities are selected though 5<2> two- syllable partial Kana character strings are generated, and selected them and the likelihood of the third single syllable are operated in 5<2> operations, and this operation is repeated to select the Kana charater string having a higher priority. Thus, the trouble of the Kana-Kanji conversion operation is reduced to make the operation easy.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention provides kana characters composed of one or more types of kana character candidates obtained for each monosyllable of a Japanese sentence divided into monosyllabic units and pronounced. The present invention relates to a kana character string evaluation method that evaluates a predetermined number of strings in descending order of priority.

In recent years, information processing devices such as computers have come to process Japanese sentences. but,
Japanese has kanji, hiragana and 2 hiragana.

Since it is a language that uses a wide variety of characters including alphanumeric characters and symbols, input problems have been considered the biggest technical obstacle when processing Japanese sentences. Currently, the biggest challenge in Japanese language processing is establishing an efficient and easy input method.

Currently, the Kana-Kanji conversion method is the mainstream Japanese input method. This is a method that inputs a desired sentence using a kana keyboard as it is read and converts it into a mixed kanji/kana sentence. It has the advantage of being able to be entered using a keyboard for kana characters. However, if the user has not received kana type m11 training, inputting from the kana keyboard will not be easy and will not only put a heavy burden on the user, but the input speed will also not be fast. Therefore, it cannot be said that the kana-kanji conversion method using input from a kana keyboard is sufficiently effective for general users who are not familiar with kana type.

Therefore, a Kana-Kanji conversion system that is equipped with a monosyllabic voice input device in place of the keyboard may be considered. According to this method,
When a user dictates the sentence they want to input, the monosyllabic voice input device converts it into a kana character string, which is then converted using the kana-kanji conversion method to obtain the desired kanji-kana-mixed sentence. This makes it possible to easily input Japanese sentences without having to receive a message.

However, since the recognition rate of monosyllable recognition by the monosyllabic voice input device is not 100 me when it comes to laughing, it is not always possible to convert the syllable string dictated by the user into the desired kana character string. This results in erroneous conversion or inability to convert into kanji/kana mixed sentences.

It is possible to consider a method in which the user checks the kana character string output from the monosyllabic voice input device and corrects errors in the kana character string using an auxiliary keyboard, etc., but this method is difficult to operate and reduces the benefits of voice input by half. It turns out. Therefore, it is desirable from the perspective of a man-machine interface to recognize the ambiguity of the voice recognition result and automatically perform processing to compensate for it, thereby minimizing manual correction operations via the keyboard.

On the other hand, when we look at Japanese as a series of monosyllables, we often find remarkable features regarding the conjunction of monosyllables. For example, after the single syllable ``a'', ``i'' and ``n'' are relatively common (in love, security, etc.), but ``zo'' and ``yo'' are relatively common.
Ha, b, I can't see it.

Additionally, although ``i'' and ``su'' are relatively often seen before ``a'' (e.g., in the case of scrapped 2-joint combinations), ``zo'' and ``nu'' are almost never seen.

By effectively utilizing such information regarding monosyllabic concatenation, it is possible to identify candidates for kana character strings that are not suitable for Japanese characters from among the candidate kana character strings that are a collection of kana character candidates output from the monosyllabic speech input device. (i.e., kana character strings consisting of kana character candidates with extremely low probability of concatenation), and conversely lower the priority of kana characters that are considered appropriate for Japanese (i.e., kana characters with extremely high probability of concatenation). By performing processing such as increasing the priority of kana character strings consisting of candidates, it is possible to evaluate more likely kana character strings as a result.

An object of the present invention is to provide kana character candidates obtained for each monosyllable by using conjunctive information between a plurality of monosyllables when the recognition result of a monosyllabic speech input device is not uniquely determined. The present invention provides a Japanese language information processing device that uses speech input, including a phonetic kana-kanji conversion method, and provides a method for evaluating kana character strings consisting of a pre-specified number of kana character strings in descending order of priority. The objective is to improve the performance and operability of the system.

An invention related to the present invention is a kana character string determination method (patent application filed in 1973).
No. 7-092757) and a kana character string priority determination method (Japanese Patent Application No. 57-092755), but in this kana character string determination method, only one kana character string candidate was obtained. In addition, in this kana character string priority determination method, from all combinations of kana character candidates (i.e., for an n-character word with 5 kana character candidates for each monosyllable, 5n combinations) Because we were considering kana character strings consisting of the following characters, the processing time was long and the amount of memory was enormous.

In the present invention, the following method is used to improve these two points and to evaluate a pre-specified number of kana character strings starting from those with higher priority.

In other words, when connecting the m+1 syllable kana character candidate to the m-syllable partial kana character string to generate the m+1 syllable partial kana character string, the m+1 syllable partial kana character string is selected as a partial kana character string with a lower priority (appropriate for Japanese). We have decided to exclude cases (the possibility of which is extremely low). This reduces the number of partial kana character strings to be evaluated, which shortens processing time and reduces storage capacity.

The present invention will be described below with reference to the drawings and specific examples; however, the device configuration used here may take a form other than this example, and the scope of the present invention is limited. It's not something you do.

FIG. 1 is a block diagram showing one embodiment of the present invention. l
2 is a monosyllabic speech input device, and 2 is a kana character/likelihood storage device for temporarily storing kana character candidates output from the monosyllabic speech input device 1 and the likelihood of the kana character candidates. 3 is a conjunctive information storage device for storing monosyllable conjunctive information; 4 is a kana character candidate, the likelihood, and conjunctive information storage in the kana character/likelihood storage device 2; 5. A kana character string/priority calculation device that outputs partial kana character strings made up of the kana character candidates in descending order of the priority of the partial kana character strings based on concatenation information in the device 3;
is a kana character string/priority storage device that stores the partial kana character string outputted from the kana character string/priority calculating device 4 and the column precedence of the partial kana character string.

The user inputs a monosyllable string into the monosyllabic speech input device 1 by uttering a Japanese sentence divided into monosyllabic units or by outputting recorded speech on a tape recorder. The monosyllabic speech input device 1 outputs, for each input monosyllable, a kana character candidate corresponding to that monosyllable and a likelihood representing the probability of the kana character candidate to the kana character/likelihood storage device 2.

FIG. 2 is a conceptual diagram showing an example in which kana character candidates and their likelihoods are stored in the kana character/likelihood storage device w, 2P'3.

In FIG. 2, A(1,j) (1,j are both natural numbers) is a kana character which is the j-th candidate of the first syllable input by the monosyllabic speech input device 1, and B(1 ,j)(1
, j are both natural numbers) is the likelihood expressed numerically as the probability of A(1, j).

On the other hand, monosyllable conjunctive information is stored in the conjunctive information storage device 3.

FIGS. 3(at), (bl, and (e)) are conceptual diagrams showing an example of the storage state.

In Figure 3 fal, M(i) (i is a natural number) and N(i) (i is a natural number) are written in kana characters, C[M(t
), N(i)) id force + letter M(1) t force to letter N(i
) is a numerical value expressing the concatenation information, specifically, for example, the concatenation frequency of character strings.

Figure 3 (b) is an example in which the conjunctive frequency is given in the format shown in Figure 3 (a). It shows that you have it.

In addition, tel in Figure 3 is an example in which the data in fbl in Figure 3 is organized into an index using kana characters as keys, but depending on the storage format of the linked information, it may be possible to search for it efficiently. It is possible.

The kana character string/priority arithmetic device 4 is constructed from the kana character candidates based on the kana character candidates in the kana character/likelihood recording device 2, the likelihood of the kana character candidates, and the linkage information in the linkage information storage device 3. This is a device that evaluates a pre-specified number of partial kana character strings starting from those with higher partial kana character string priorities. A concrete example of the arithmetic unit 4 for kana character strings and priority is as follows.

When a kana character candidate and the likelihood of the kana character candidate are given in the format shown in FIG. 2, the m-syllable partial kana character string always takes the form of equation (1).

A(1*x1)A(2,xl)+++・+A(m,xm
) (1) However, x, (1=1.2......, m)
is a kana character candidate number. The processing of equations (2) and (3) below is repeatedly applied starting from the first syllable to generate a partial kana character string in the form of equation (1), until the final n-th syllable is processed. Excluding those that are particularly inappropriate for Japanese (i.e., kana character strings consisting of kana character candidates with extremely low possibility of concatenation), and conversely, those that are considered to be appropriate depictions for Japanese. (In other words, non-kana character strings composed of kana character candidates with extremely high possibility of concatenation) can be evaluated by a predetermined number based on priority.

TI (PI) = 9' [B (1, i), B (2, J), f
(C(A(1,t).

A(2,j) month] (2) However, 1=1,2. ......, m1j=1, 2,...
......, m1 P1°1, 2, -=, ml+m鵞mk: Number of kana character candidates for the 1st syllable %; Set of +1 syllable partial kana character string priorities TJ (P = 9 (B (j+1.j), f(C(A(J,
i), A()+1゜* j) month, Tl-1(1)) (3) where i=1+2+=・=・*ml("J+1""N)
j=1, 2,..., m4+□ PJ"1e2t-""+"j""J-1-1""2+3
+...”-・, n1 In equations (2) and (3), the function f is a function that uses the concatenation frequency of kana characters as a variable, and specifically, the ordering is based on the concatenation frequency with a prespecified threshold value. By using a suppressing function, it is possible to avoid the situation where the accuracy rate is lower than the original data due to the connection frequency factor being too large, and it is also possible to increase the accuracy.

Function 9 also calculates the likelihood x of kana character candidates, (i=1.2°・
..., j), the value y obtained by substituting the concatenation frequency into the function f, and the priority flJz of the partial kana character string. Specifically, For example, f(
xl, xl, -=, x,, y, z)=x1+x@+=+
It can be expressed by a mathematical formula such as x, +y+t. Function 9′2g
" are 7.0 respectively), f" (y, z) mi' (OrOr
"""eO*7+").

<2), Equations (3) are explained as follows.

First, in equation (2), the value obtained by substituting the concatenation frequencies of the first and second characters of m, ・m1 through, and the first and second characters.
A priority set TI is obtained using the likelihood of the characters. This T1
are rearranged in descending order to obtain a descending priority set TF. at the same time,
The two-syllable partial kana character string is parallelized with LX(1*,j*)(1
*=1.2j*=1, 2, -=, ms) t' is obtained. However, at this stage, the number m of kana character candidates for the second syllable is set to a prespecified number N.

From the 2nd character onward, the value obtained by substituting the conjunctive frequency of the 91st character and 1+1 character of m□・m□□ to the function f in equation (3), the likelihood of the 2+1 character, and the 1st syllable part. A priority set TJ is obtained using the kana character string priorities. This is the descending priority set TJ and A(ij) (1=1.2."'=', 1+
IJ=1.2. """, mj+1) is obtained.

At this time, move to m□ or □. The processing of this equation (3) is 1=
2.3. By repeating this process for n-1 characters, it is possible to evaluate a predetermined number of kana character strings starting from the highest priority.

In this way, by storing a predetermined number of kana character strings in the kana character string/priority storage device 5 starting from those with high priority, the kana character strings corresponding to the input voice can be prioritized. You can obtain a pre-specified number of 4 from the highest.

Furthermore, although the explanation up to now has dealt with only connections between two kana characters, it is also possible to deal with connections between three or more kana characters.

Figure 4 is an example of memorizing the conjunctive frequency between three characters, and the weights have values such as ``Aia'' is 0, ``Aii'' is 5, ``Aie'' is 2, and ``Ai-ryoku'' is 4. It is shown that. In this case, the priority set T mentioned above can be expressed as follows.

Tt, (Pt) = /CB (1, s), B (2, j), B
(3, k), f(C(A(1,1), A(2,j), A
(3, k) month] (4) However, l = 1, 2, -...
, m1j=1, 2,...-, m1 k=1, 2,...2m.

P1=1, 2, -=, mllmCm11** TL(PJ,)=f(B(L+2.j), f(C(A(
L, s), A* (L+z, j) month, T□−□(1))'(5) However”
""'l"J, -1 (mj+2=N)j=1.2.-1...emL+t Pj=1, 2*-=*mL+1°nll+2j=2.3
．． −・・・−, n−2 In addition, Fig. 5 +81
, (b), it is also possible to include a symbol denoting the beginning of the character string and a symbol denoting the end.

FIGS. 5(a) and 5(b) show an example thereof, where ■ is a symbol that means the beginning of a character string, and ■ is a symbol that represents the end of a character string.

Figure 5 (al means "■""A" (i.e., meaning that it starts with "A") has a value of 100, and the value is "N""■" (i.e. "
The weighting of ``meanings ending with ``'' indicates that the value is 50. And Fig. 5 fbl is "■", "power", "n" (
``Charcoal'' (starting with ``N'') is 45, ``I'', ``Ri'', ``■'' (
Ending with ``i'' and ``ri'') indicates that it has a value of 30. In these cases, the priority τ can be expressed as follows.

Two-character concatenation; Ts(Pt)=f'(B(1,l), f(C(■, A(
1, i) Month] (6) However, i = 1, 2, ・・・・・−,
m1P1,=1.2. ......, m1 * Tj (Pi) = 9 (B (J, j), f (C (A (j-1)
,i),A(j,j) sword.

*T□-1(1))(7) However, I=1+2, -""1m2-+(=2""N)j
= 1, 2, ......, m4 p, z = 1 + 2 + = "" + ml - s m no j = 2, 3,
......, n ** Tn+1(Pn+1"""(r(C(A(n, t), ■
month, Tn(i):) (8) where i=1.2.・・・・・・
...rmnP=1.2.・・・・・・+mn n+1 3-character concatenation T+(Pt)-2'(B(1,+), B(2,j), r
(C(■, A(1,i).

A (2, j) month] (9) where t=i, 2,..., m1j=1, 2,...
...1m.

P,=1.2. ......, ml・m.

Tz (Pz) = 2CB (J + 1, j), f (C (A (z
-i, 1), A** * (ttt), A(z-+'1.j) month, TL-□(1)
+IQ However, I=1.2. ”'=', mL (m44.=
N) j=1, 2,...'-, mJ+□ pJ:=1t2+-=t=2"mz4-tj=2.3.
・-・=, n-1 Tn(Pn)=su[f(C(A*(n-1,i), A*
(n, i), October.

αη T,, (1)] where i=1, 2,...1mnP=1.2.・
-...1mn Figure 6 (a) is an example showing the data structure in the kana character string/priority calculation device 4. The white mark indicates the priority order, 7 indicates the partial kana character string, and 8 indicates the partial kana character string. It has a priority of 7.

If the pre-specified number (N in the above explanation) is 5 as shown in Figure 6(a), for example, in a word of n letters with 5 kana character candidates for each monosyllable, (2) .

If we apply the processing in equation (3), we can generate a two-syllable partial kana sentence string in 5×5 ways, from the top candidates of the two-syllable partial kana character string and the third syllable kana character candidate. This is repeated in 5×5 ways to generate a syllable partial kana character string, so only partial kana character strings consisting of a total of 5×5X(n-1) combinations are evaluated. Compared to the column priority determination method (Japanese Patent Application No. 57-092755), which evaluates partial kana character strings consisting of 5n combinations, the processing time is shortened and the storage capacity is also reduced.

The data structure can also be in a cell format as shown in FIG. As control information, a table pointer 11 indicating which partial kana character string in the candidate table 9 is pointed to, a syllable number 12 indicating whether a partial kana character string for a side syllable is included in the candidate table 9, and the corresponding partial kana character string. A character string priority 13 and a cell pointer 14 pointing to the next highest priority cell after that cell are stored.Furthermore, 15 is the maximum cell pointer to the cell with the highest priority 13, and 16 is the maximum cell pointer. This is a control table consisting of a pointer 15 and a control cell 10. By using a cell-format data structure and controlling partial kana character strings with the table and pointer, rearranging becomes simple, reducing processing time. In addition, by making the size of the candidate table 9 constant, the number of candidates for partial kana character strings (i.e., narrowing down by the value of N in the above explanation of 9 (i.e., for example, the size of the candidate table 9 is 2).
Assuming that there are 5 characters, in the case of a 5-character word, 12 candidates can be stored up to the 2nd character, but as the 3rd, 4th, and 5th characters are processed, 8 candidates with higher priority can be stored. Toshi,
The number of candidates for partial kana character strings changes from 6 letters to 5 letters), so we leave many possibilities open at the beginning, and as we get closer to the end, lower kana character strings (i.e., concatenated It is also possible to suppress the risk that a kana character string (consisting of kana character candidates with extremely low probability) will appear among the top candidates, and to shorten the processing time.

FIG. 7 is an example of kana character candidates and likelihoods outputted from a monosyllable string uttered by a user to the kana character/likelihood storage device 2 through the monosyllabic voice input device 1. In the figure, the numbers in parentheses are It shows the likelihood. FIG. 8 also shows the kana character candidates in the kana character/likelihood storage device 2 and the likelihoods of the kana character candidates (FIG. 7) and the connection information storage device 3 using the kana character string/priority calculation device 4. The kana character strings composed of the kana character candidates are output to the kana character string/priority storage device 5 in descending order of priority based on the concatenation information.
(al processes equations (6), (7), and (8); (bl processes equations (9) and (10).

This is the result of implementing 0υ type processing. Note that the numbers in parentheses in FIG. 8 indicate the priority.

As shown in FIGS. 7 and 8, even if the first candidate as a monosyllable recognition result is incorrect, a correct kana character string can appear as the first candidate by using the concatenation information. Even if you do not appear as the first candidate, there is a high possibility that you will appear among the top candidates.

As described above, according to the present invention, when the g recognition result of a monosyllabic speech input device is not uniquely determined, kana character strings consisting of kana character candidates are sorted in descending order of priority. Therefore, the user's work such as inefficient correction of kana characters can be reduced, and an efficient phonetic kana-kanji conversion method can be realized.

Kana characters and likelihood used to explain the present invention.

Information storage formats such as linkage information may take a format other than this example, and the scope of the present invention is not limited thereto.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing an embodiment for realizing the present invention. In the figure, 1 is a monosyllabic voice input device;
2 is a kana character/likelihood storage device, 3 is a conjunction information storage device,
4 is a kana character string/priority calculation device, and 5 is a kana character string/priority storage device. Fig. 2 is a conceptual diagram showing an example of the storage format of kana character candidates and likelihoods; Fig. 3 fa), (b), (c) and Fig. 4 and Fig. 6 la, (b) are all connected. A conceptual diagram showing an example of an information storage format, FIG. 6 1al, (b) shows a kana character string/
Conceptual diagram showing an example of the data structure in the priority calculation device 4, agent Bentoshi INJi + ξ Shinmi Suga η Togi 7 mouth

Claims (2)

[Claims] (1) Likelihood indicating the probability of one or more types of kana character candidates for each monosyllable of a Japanese sentence pronounced in monosyllabic units, and the likelihood between multiple monosyllabic characters stored in advance When evaluating the kana character string composed of the kana character candidates and the priority of the kana character string using the linkage information, one or more types of m-syllable portions composed of m syllables of the kana character candidates. the connection information relating to the kana character candidate in the kana character string and the kana character candidates after the m+1 syllable; the m-syllable partial kana character string priority representing the certainty of the m-syllable partial kana character string; and m+1. The tn+1 syllable partial kana character string priority is determined using the above-mentioned likelihood of the syllable, and a pre-specified number of m+1 syllable partial kana character strings are selected from the m+1 syllable partial kana character strings with high priority. A kana character string evaluation method, characterized in that one type or plurality of kana character strings and the priority of the kana character strings are evaluated by repeating a generation process. (2) When generating one or more types of partial kana character strings, the number of partial kana character strings to be generated is made variable depending on the length of the partial kana character strings. The kana character string evaluation method described in (1).