JP2527719B2 - Language processing method of language processing device - Google Patents

Description

TECHNICAL FIELD The present invention relates to a Japanese speech recognition apparatus and a language processing method for post-processing the output of a Japanese word processor. When multiple bunsetsu candidates are given at the time, each bunsetsu string is constructed so that the optimal bunsetsu string is constructed as a Japanese phrase or sentence, taking into account the reliability of those candidates and the consistency of the dependency between bunsetsu. The present invention relates to a language processing method for selecting a phrase one by one from the phrase candidates at the time, determining its syntax, and calculating the degree of eligibility of a phrase sequence obtained thereby as a Japanese phrase or sentence.

[Prior Art] In some Japanese processing devices that use bunsetsu units, a plurality of bunsetsu candidates are output at each time point. For example, there are many such Japanese speech recognition devices that use bunsetsu units, and even in Japanese word processors that use the kana-kanji conversion method, there are kanji with the same meaning, so one bunsetsu input It is considered that multiple candidates are output. Therefore, it is conceivable to select a bunsetsu from candidates at each time point so that the bunsetsu sequence with the highest qualification can be obtained from such a bunsetsu candidate in consideration of the connection relation between the bunsetsus.

In explaining this method, first, the structure of Japanese in terms of bunsetsu will be described.

A Japanese sentence, or a set of phrases, can be thought of as being formed by a broadly modified relation between units called bunsetsu. For example, in the Japanese sentence [S1] "I go to the office by train at 8 o'clock", "I am", "8 o'clock",
"By train", "to the office", "go",
"I am", "by train", "to the company" are all modified to "go", and "8 o'clock" is modified into a sentence by modifying "by train". I am configuring.

When clause x modifies clause y, "x is related to y, y
Receives x ". In addition, such a modification relationship is called dependency.

In order for a bunsetsu sequence to form a set of phrases or sentences in Japanese, it is considered that there is a dependency between those bunsetsu that satisfies the following conditions.

[C1] A clause other than the last clause relates to any one of the clauses after it.

[C2] Dependency between two bunsetsu does not intersect with dependency between two other bunsetsu.

[C3] To have a dependency between two clauses,
The types and meanings of these clauses must have a certain relationship with each other. For example, in [S1] above, "I" can relate to "go", but "8 o'clock"
Cannot be related to

The conventional parsing of Japanese sentences is nothing but searching for all dependency relations that satisfy such conditions for a given phrase sequence (eg, Masayoshi Yoshida, Syntactic analysis of Japanese sentences based on reception ", IEICE Theory '72 / 4, Vol.55-D, No.4, pp.238-243).

If the given phrase sequence is correct Japanese, parsing can be performed by such a method, but it is possible for the phrase sequence that contains an error that is often found in the spoken language or the output of the speech recognition device. , Parsing can get bogged down.

Therefore, the above condition [C3] is a more flexible condition [C3 '] A numerical value that represents the consistency of the set of two clauses x and y in that x is related to y depending on the type and meaning of those clauses. Is given.

There is a method of searching for the most highly qualified dependency relation as a whole. This will be described next.

Conditions [C1] and [C2] are defined as follows:
It can be represented by "syntax".

[D1] (1) When x is a clause, (x) is "syntax".

(2) When X ₁ , X ₂ , ..., X _m is “syntax” and x is a clause,
_{(X 1 X 2 ... X m} x) is the "syntax".

[D2] The phrase sequence x ₁ x ₂ ... x _n with proper parenthesis to make it a syntax is called the syntax on x ₁ x ₂ ... x _n . ₁ phrase sequence
The overall syntax on x ₂ ... x _n to be expressed as _{K (x 1 x 2 ... x} n).

Syntax _{(X 1 X 2 ... X m} y), X 1 = (... x 1), X 2 = (... x 2),
If we promise that x ₁ , x ₂ , ..., x _m in y, x _m = (... x _m ) are related to y, in the above sense, the condition [C1] becomes [C2] is satisfied, and conversely, the dependency relation in the clause sequence that satisfies the conditions [C1] and [C2] can always be expressed by the syntax in the above meaning.

Regarding the condition [C3 ′], it is assumed that the consistency of the fact that the clause x is related to the clause y is represented by a function PEN (x, y) taking a nonnegative value. The value of PEN (x, y) is 0, for example.
The range is from 1 to 100, and promises that the closer to 0, the higher the consistency. How to determine the function PEN is a very important issue, but this is already considered in the past and is not the main point of the present invention, so its explanation is omitted here.

Under the above preparation, the eligibility P (X) of the syntax X is recursively determined as follows.

[D3] (1) When X = (x), (x is a clause), P (X) = 0, (2) X = (Y ₁ Y ₂ ... Y _m x), Y ₁ = (... y ₁ ), Y ₂ = (...
When y ₂ ), ..., Y _m = (... y _m ), P (X) = P (Y ₁ ) + P (Y ₂ ) + ... + P (Y _m ) + PEN (y
_{1, x) + PEN (y} 2, x) + ... + PEN (y m, x 1) the value of the thus-defined P (X) is obtained by adding the value of PEN for any dependency in the X It will be clear that The following definition is made based on P (X).

[D4] For a given subsequence of x ₁ x ₂ ... x _n , OPTP (x _i x _{i + 1} ... x _j ) is a numerical value indicating the degree of eligibility of the clause sequence x _i ... x _j as a sentence or a group of phrases.
Also, OPTK (x _i x _{i + 1} ... x _j ) is the most highly qualified syntax on the clause sequence x _i x _{i + 1} ... x _j .

Now consider the following situation: [J1] At each time point from 1 to n, the set of clauses B ₁ ,
Suppose B ₂ , ..., B _n are given. Also, each clause set
For B _k , a non-negative real-valued function S _k is defined as follows: S _k : B _k → [0, ∞) S _k (x) is a clause x in the clause set B _k . Is assumed to be a value from 0 to 100, and the closer to 0, the higher the reliability. Taking voice recognition as an example,
S _k (x) is a numerical value indicating how likely the recognition device has recognized the recognition result x, and most speech recognition devices have come to output such a numerical value together with the recognition result. There is.

Taking a Japanese word processor that uses the Kana-Kanji conversion method as an example, multiple bunsetsu candidates that have the same reading are output because of the presence of homophones, but each candidate follows the frequency of use of Kanji or idioms. A numerical value indicating the reliability can be attached.

Under such circumstances, the problem of language processing handled by the present invention can be stated as follows.

[problem] And the clause sequence x ₁ x ₂ … x _n and OPTK (x ₁
Find x ₂ ... x _n ).

Conventionally, the solution to this problem is to use a full-width search, that is, for every element X of every possible clause sequence x ₁ x ₂ … x _n and K (x ₁ x ₂ … x _n ), Suzuki Was calculated, and the minimum value and the syntax giving the minimum value were calculated (for example, Shoichi Matsunaga, Masanori Yoshida "Post-processing of phrase recognition based on dependency consistency", The Acoustical Society of Japan). Proceedings of Spring Research Presentation 1986 1-1
-23).

[Problems to be Solved by the Invention] Here, except for the calculation of the function PEN that indicates the suitability of dependency, the basic operations for solving the above problems are addition of real numbers and comparison operation. Letting L [n] be the number of syntaxes on a clause sequence having a length of n, that is, the original number of K (x ₁ x ₂ ... x _n ), L [n] = L [n−1] · L [1] + L [n-2] L
[2] + ... + L [1] · L [n-1] holds. When this L [n] is used, the amount of calculation in the conventional full-width search is as follows. However, the number of clauses in each clause set was all M.

Adding: (L [n] · ( n-1) + (n-1)) M n ^{comparison: L [n] · M n} -1 these equations M = 5,10, n = 5,10,15 Table 1 shows the results calculated for 20,20,25,30.

As can be seen from Table 1, the number of possible phrase sequences and K (x ₁
Since the original number L [n] of (x ₂ ... x _n ) increases exponentially with the length n of the bunsetsu sequence, the conventional method requires a very large amount of calculation and It was extremely difficult to apply to the problem.

Therefore, an object of the present invention is to improve the above-mentioned drawbacks of the conventional technique and to significantly reduce the calculation amount as compared with the conventional method, that is, the cube of the bunsetsu string length and the square of the original number of each bunsetsu set. It is to provide a language processing method for performing desired language processing with a calculation amount of

[Means for Solving Problems] In order to achieve such an object, the present invention, when a sequence of bunsetsu sets and a numerical value indicating the reliability of each bunsetsu in the bunsetsu set sequence are given, 1 is obtained from each bunsetsu set by minimizing or maximizing the sum of the degree of dependency matching between two bunsetsu and the sum of numerical values indicating the reliability of each bunsetsu.
In the language processing method of the language processing apparatus, which determines the optimum phrase sequence under the condition that each phrase is selected, the optimum syntax above it, and the degree of its eligibility, the length of the phrase set sequence is Two-dimensional upper triangular matrix first and second tables having the same number of rows and columns are prepared on the memory, and each grid of the first table and the second table is assigned to its column number. The first table and the second table are three-dimensionalized in advance by dividing them into the same number of terms as the original number of the clause sets of equal numbers, and i rows, i columns, and p terms have i A numerical value representing the reliability of the p-th phrase in the n-th phrase set is set in the holding circuit, and k = i,
, J−1, each item in the i-th row and the k-th column on the memory,
The calculated value is stored in k + 1 row, j column, q term, and if the storage is done, the calculated value of i row, k column, p term of the first table in the memory is stored. , K + 1 row, j column, the value of the q term, and the degree of consistency that the p-th clause of the k-th clause set relates to the q-th clause of the j-th clause set are added in an operator. , The minimum value or the maximum value of k and p of the addition result and the value of k and p that gives the minimum value or the maximum value are detected by the detector, and the detected minimum value or the maximum value is stored in the i row on the memory. , column j, q, and store the set of values of k, which is the optimum segment point that gives the minimum or maximum value, and p, which is the optimum clause number, in row i of Table 2 on the memory. , j column, q term, and the first table and the second table in the memory are sequentially filled with the calculated values, and the memo is stored. Table 1 above and the second
When the table is completely filled with calculated values, the first
The final degree of eligibility and the optimum clause number in the last clause set are obtained by finding the minimum or maximum value in each term in the upper right corner of the table, and the optimum required for constructing the optimum syntax. It is characterized in that all the partition points and the optimum phrase numbers are obtained in the second table on the memory.

(A) Basic recursive equation Before describing the configuration of the present invention, a recursive equation that plays a basic role will be described first.

X, in Y _{syntax, Y = (Z 1 Z 2} ... Z m y), ( where, Z _1,
…, Z _m is a syntax, and y is a clause), the syntax [XZ ₁ Z ₂ … Z _m y) created by inserting X at the beginning of Y is written as XY, and the following [E1] is established. It's easy to see.

[E1] In the clause sequence x ₁ x ₂ ... x _n , for (1) K (X _i ) = (X _i ), (1 in) (2) 1 _i <jn Further, for [E2] X = (... x), Y = (... y), P (XY) = P (X) + P (Y) + PEN (x, y). In _{fact, Y = (Z 1 Z 2} ... Z m y), Z 1 = (...
_{z 1), z 2 = (} ... z 2), ..., z m = (... if _{z m), P (XY)} = P ((XZ 1 Z 2 ... Z m y)) = P (X) + P
(Z ₁ ) + P (Z ₂ ) + ... + P (Z _m ) + PEN (x, y) + PEN (z
₁ , y) + PEN (z ₂ , y) + ... + PEN (z _m , y) = P (X) + P
Using (Y) + PEN (x, y) [E1] and [E2], the following important relation [E3] is obtained.

[E3] (1) OPTP (x _i ) = 0 (2) For i <j In fact Now, for each x _j εB _j , far. Then, by [E3] Therefore, by writing B _k = {x _{k, 1} x _{k, 2} , ..., x _{k, n (k)} }, the following recurrence equation for OPTPS is obtained.

[E4] (1) For 1pN (i) OPTPS (i, i; xi, p) = S _i (x _{i, p} ) (2) For i <j, 1qN (j) The k and p that give the minimum value in the above equation are called the optimal segment point and the optimal clause number for i, j, and q, respectively. If we write OPTKS (i, j; x _{j, q} ) the optimal syntax above the optimal clause sequence that gives OPTPS (i, j, x _{j, q} ), the following holds from [E4]: .

[E5] (1) OPTKS (x _{i, p} ) = (x _{i, p} ) (2) When the optimal segment points and optimal clause numbers are always uniquely determined Optimal segment points and optimal clause numbers for i, j, q Are written as k and p respectively, OPTKS (i, j; x _{j, q} ) = OPTKS (i, k; x _{k, p} ) OPTKS (k +
(1, j; x _{j, q} ) (3) When there are multiple optimal segment points or optimal clause numbers, OPTKS becomes a set, and sets of optimal segment points and optimal clause sequences are (k1, p1), (k2, p2), ..., (kL, pL) (B) OPTPS, optimal segmentation point, optimal clause number calculation method (2) of [E4] is OPTPS (i, k;
x _{k, p} ) and OPTPS (k + 1, j; x _{j, q} ), (ikj−1,1)
It has been shown that if pN (k)) has already been calculated, OPTPS (i, j; x _{j, q} ) can be calculated by solving the minimization problem of a one-variable function twice. So (1) of [E4]
And (2), OPTPS (i, j; x _{j, q} ) j-1 is 0
It is possible to start from the part of (1) to sequentially increase to a larger part, and at the same time, determine the optimum segmentation point k and the optimum clause number p.

(C) Optimal Syntax and Calculation Method of Its Eligibility For the sake of simplicity, a case will be described in which the optimal segment point and the optimal clause number are always uniquely determined. First, because there is, Calculates the degree of eligibility for the optimal syntax on the optimal clause sequence. Also, Then, the optimal clause sequence and the optimal syntax above it are given by OPTKS (i, n; x _{n, qo} ). This can be calculated more concretely as follows.

If the optimal segment points and optimal clause numbers for 1, n, q0 are s1 and t1, respectively, then OPTKS (i, n; x _{n, q0} ) = OPTKS (1, s1 ; x _{s1, t1} ) OPTKS (s
1 + 1, n; X _{n, q0} ) holds. If s1 ≠ 1, then OPTKS (1, s1;
x _{s1, t1} ) is OPTKS (1, s1; x _{s1, t1} ) ＝ OPTKS (1, s2; x _{s2, t2} ) OPTKS, using the optimal segmentation point s2 and optimal clause number t2 for 1, _{s1, t1}
It can be decomposed into (s2 + 1, s1; X _{s1, t1} ). For OPTKS (s1 + 1, n; X _{n, q0} ) also s1 +
If 1 ≠ n, OPTKS (s1 + 1, n; X _{n, q0} ) = OPTKS (s1 + 1, s3; X _{s3, t3} ) using the optimal segment point s3 and optimal clause number t3 for s1 + 1, _{n, q0}
It can be decomposed as OPTKS (s3 + 1, n; X _{n, q0} ). Therefore, OPTKS (1, n; x _{n, q0} ) = (OPTKS (1, s2; x _{s2, t2} ) OPTKS
(S2 + 1, s1; X _{s1, t1} )) (OPTKS (s1 + 1, s3; X _{s3, t3} )
OPTKS (s3 + 1, n; X _{n, q0} )) Repeat this operation until ji = 0 in all of the appearing OPTKS (i, j; x _{j, q} ), and when it becomes 0, [E5]
By using (1) of (1) to replace the syntax with only one clause and perform the insertion operation by tracing the reverse of the decomposition, the optimum clause sequence and the optimum syntax on the clause sequence can be obtained at the same time.

If there are multiple optimal segment points and optimal phrase sequences, click [E
The same operation may be performed using (3) of [5].

[Operation] In the present invention, when a sequence of bunsetsu sets and a numerical value indicating the reliability of each bunsetsu in the set are given, the consistency of dependency between two bunsetsus and the reliability of each bunsetsu are determined. Originally, we select the most suitable bunsetsu example from the bunsetsu sequence that is composed by selecting one bunsetsu from each bunsetsu set in order, determine the optimum syntax for that bunsetsu string, and determine the eligibility of that syntax. In calculating the degree of, if the bunsetsu sequence that has the highest eligibility as a Japanese phrase for a given substring of a bunsetsu set and its construction, and the degree of eligibility for that construction, are calculated, By storing it and using the result of the calculation when performing a similar calculation for a longer length subsequence, it is possible to partially repeat the same calculation as in the full width search. Systematically avoid Therefore, on the given sequence of phrase sets, the Japanese phrase or the sentence with the highest qualification as a sentence,
The eligibility for the syntax can be calculated with a much smaller amount of calculation than the conventional method.

EXAMPLES The present invention will be described in detail below with reference to the drawings.

 An embodiment of the apparatus for carrying out the present invention is shown in FIG.

In FIG. 1, SC is a RAM (a holding circuit of the present invention) for storing and storing a numerical value indicating the reliability of each clause from an input terminal i1, and BUF is a RAM for retaining a clause set sequence input from a clause input terminal i2. It is a buffer memory. PEN
Is 2 of the phrase strings read from the buffer memory BUF
This is a device that calculates the degree of dependency matching between clauses.

T1 and T2 are RAMs (corresponding to the memory of the present invention) for realizing the tables TABLE1 and TABLE2 of the flowchart shown in FIGS. 2 (A) and (B).

ADD1 is an adder (a part of the arithmetic unit of the present invention) that adds TABLE1 (i, k, p) and PEN (k, p, j, q).

MIN1 is a minimum value detector (which constitutes a part of the detector of the present invention) for detecting the minimum value of the output of the adder ADD1 when the above-mentioned p is changed, and p which gives the minimum value. .

ADD2 is the output of minimum value detector MIN1 and TABLE1 (k + 1, j,
q) and (adder ADD1 constitute an arithmetic unit of the present invention).

MIN2 is the minimum value of the output of the adder ADD2 when k is changed, and the minimum value detector for detecting k that gives the minimum value (which constitutes the detector of the present invention together with the minimum value detector MIN1. ).

CONT is a control device for controlling the operation sequence of each of these parts, and is composed of, for example, a central processing unit CPU. This central processing unit CPU has a memory MEM in the form of a ROM for storing the control procedure of each unit in advance.

O1 and O2 are output terminals for outputting the results written in the RAMs T1 and T2, respectively.

2 (A) and 2 (B) show the degree of qualification of the optimum syntax on the optimum phrase sequence and the optimum degree as an example of the control procedure stored in advance in the memory MEM in the embodiment shown in FIG. It is a flowchart which shows the procedure for sequentially obtaining | requiring the optimal segmental point for defining a phrase sequence, the optimal syntax on it, and the optimal phrase number. Hereinafter, this will be described.

As shown in FIGS. 3A and 3B accompanying the flowcharts of FIGS. 2A and 2B, the number of rows and columns equal to the length of the clause column and the column number thereof are equal. Two three-dimensional upper triangular matrix tables TABLE1 (i, j, q) and TABLE2, which have as many terms as there are elements in the clause set of numbers
(I, j, q) (1ijn, 1qm) is required.
Here, n is the length of the phrase set sequence, and m is the maximum value of the original number of each phrase set. The subscripts for each table are from left to right,
Indicates the row, column, and item number.

TABLE1 (i, j, q) stores the value of OPTPS (i, j; x _{j, q} ),
TABLE2 (i, j, q) is for storing a set of optimum segment points and optimum clause numbers for i, j, q.

It is also assumed that the reliability of the p-th clause in the k-th clause set is input and held in the two-dimensional table SC (k, p).

It is assumed that the number of clauses in the i-th clause set is input and held in the one-dimensional tables N (i) and (1in).

In addition, the function to calculate PEN (X _{k, p} , X _{j, q} ) is PEN (k, p,
j, q).

In the flowchart of FIG. 2A, first, in steps S1 to S7, the value of SC (i, p) is set in TABLE1 (i, i, p) (1in, 1kN (i)). This corresponds to OPTPS (i, i; x _{i, p} ) = S _i (X _{i, p} ), according to [E4] (1).

Next, in steps S8 to S18, the column number j of each table starts from 2 and is incremented by 1 up to n, and the following process is executed for each column.

In steps S9 to S16, the following processes [F1] and [F2] are executed while decreasing the row number i of each table from j-1 to 1 by one.

The process [F1] executed in step S11 is (2) of [E4].
In addition, the processing performed by step S12 [F
[2] is an operation for obtaining the optimum segment point and the optimum phrase sequence and storing them in TABLE2.

For [F1] q = 1, ..., N (j) And store the result in TABLE1 (i, j, q).

The pair of k and p (k, p that gives the minimum value in [F2] [F1]
Store p) in TABLE2 (i, j, q).

At step S11, for k = i, ..., j−1 by [F1], the i row, k column, p term of TABLE1 and the value of k + 1 row, j column, q term, and the kth clause set The p-th clause adds the degree of consistency of being related to the q-th clause of the j-th clause set, and the minimum value (or maximum value) of k and p of the addition result is added to row i of table TABLE1. , column j, q term.

In the next step S12, the set of values k and p (k, p) that gives such a minimum value (or maximum value) is stored in the table TAB.
Store in the i-th row, j-th column, and q-term of LE2.

By repeating steps S11 and S12 by steps S9, S10 and S13 to S18, each row of tables TABLE1 and TABLE2,
The above calculation is applied to columns and terms, and the results are sequentially filled.

In step S18, j> n and table TABLE
When all rows, columns, and terms of 1 and TABLE2 are filled with the calculation results, the final eligibility degree and the last clause set in the last clause set are obtained by finding the minimum value for each term in the upper right corner of table TABLE1. The optimum clause number is obtained. Then, the entire optimum partition points and optimum clause numbers required for constructing the optimum syntax are obtained in the table TABLE2.

The shaded area in FIG. 4 indicates the area of TABLE1 that is referenced when calculating TABLE1 (i, j ,.).

Figure 5 shows the order in which TABLE1 is filled with the calculated values. Finally, TABLE1 (1, n, q) (1q
N (n)) is calculated. For TABLE2, the calculation order is the same.

When the calculation is completed, the final degree of eligibility for each term in the upper right corner of TABLE1, namely, TABLE1 (1, n, q) is OPTPS (1,
n; x _{n, q} ) is stored. Further, since the information of the optimum segment point and the optimum clause number is stored in TABLE2, the optimum clause sequence and the optimum syntax can be constructed from this information by the method described in the above item (c).

When actually using the present invention, in addition to the flow charts of FIGS. 2 (A) and (B), a mechanism for constructing an optimum clause sequence and an optimum syntax on it from the information of TABLE2 is required. Since the main point of the invention is to calculate the contents of TABLE1 and TABLE2, the mechanism for constructing the optimal clause sequence and the optimal syntax thereabove will be limited to the above description.

However, if the contents of TABLE1 and TABLE2 have been calculated, the most computationally expensive part of the calculation required to construct the optimal phrase sequence and the optimal syntax above it from the given set of phrases. Note that is no longer available.

There may be multiple k and p that give the minimum value in [F1]. In that case, TABLE2 (i, j, q)
It is possible to store multiple sets of numerical values in
In 2], all of those pairs may be stored in TABLE2 (i, j, q). Thus, even if the flowcharts of FIGS. 2A and 2B are changed, the calculation amount is almost unchanged.

As described above, the feature of the present invention is that the last bunsetsu is fixed in the bunsetsu string that can be selected one by one from the given set of bunsetsu B ₁ B ₂ ... B _n . In order to find the optimal phrase sequence at time, the optimal syntax above it, and the degree of eligibility for it, the subsequence B _i B _{i + 1} ... B _j of the set of phrases
Similar to the above, the optimal phrase sequence when the last phrase is fixed, the optimal syntax above it, and the degree of eligibility are sequentially obtained from those corresponding to the short subsequences, and the result is obtained. In performing the above calculation for the subsequence B _i B _{i + 1} ... B _j , the set that has already been obtained and stored for k = i, ..., j-1 is stored. Subsequence of B _i B
_{i + 1} … B _k and the subsequences B _{k + 1} B _{k + 2} … B _j , and the consistency of the fact that the clause x _{k, p} εB _k relates to the clause x _{j, q} εB _j This is where only function values are used.

In the above-described embodiment, the case of the process of obtaining the minimum value is shown, but the present invention is not limited to this case, and the process of obtaining the maximum value may be performed. However, when the minimum value is processed, the minimum value is consistently calculated, and conversely, when the maximum value is calculated, the maximum value is always calculated.

[Effects of the Invention] Regarding the function PEN indicating the suitability of dependency, once the calculation is stored, the calculation amount becomes the same between the conventional method and the present invention. Therefore, if this part is excluded, the basic operations are addition and comparison operations of real numbers. Therefore, the conventional method and the present invention will be compared below by the number of times of these operations.

The number of syntaxes on a phrase sequence of length n, that is, K (x ₁ x ₂ ...
_If the original number of x _n ) is written as L [n], then L [n] = L [n-1] .L [1] + L [n-2].
L [2] + ... + L [1] .L [n-1] holds. Using this L [n], it can be confirmed that the calculation amount is as follows. However, the number of clauses in each clause set was all M.

The present invention addition: M · (M + 1) · n · (n−1) · (n + 1) / 6 Comparison: M ² n · (n−1) · (n + 1) / 6−Mn · (n−
1) / 2 Table 2 shows the results of calculating these equations for M = 5,10, n = 5,10,15,20,25,30.

Regarding the conventional method, as described above, addition: (L [n]. (N-1) + (n-1)) ^Mn comparison: L [n] ^.Mn- 1 becomes = 5,10, n = 5,10,15,20,25,30 is the result calculated in Table 1.

As can be seen by comparing Tables 1 and 2, the larger M and n, the greater the effect of the present invention.
Number of additions when = 10 is improved to about 10 ^one-tenth, 1, compare the number of operations of the number of additions of about 10 ²⁵ minutes when the comparison operation count to 1 to about 10 ⁹ minutes, also n = 20 is approximately It is improved by a factor of 10 ²⁴ minutes.

As is clear from the above, according to the present invention, when a sequence of bunsetsu sets and a numerical value indicating the reliability of each bunsetsu in the set are given, the consistency of dependency between two bunsetsus, Based on the reliability of each bunsetsu, select an optimal bunsetsu string from the bunsetsu string that is constructed by selecting one bunsetsu from each bunsetsu set in order, and then determine the optimum syntax for that bunsetsu string. Moreover, in calculating the degree of eligibility of the syntax, the bunsetsu sequence most highly qualified as a Japanese phrase for a given subsequence of a bunsetsu set, the syntax for it, and the degree of eligibility for that syntax Once it has been calculated, it is stored and the result of that calculation is used to perform a similar calculation on a subsequence with a longer length, so that the same partial calculation as in full-width search is performed. Is repeated Therefore, by applying the present invention, the syntax that is most qualified as a Japanese phrase or sentence on a given bunsetsu set sequence and the suitability for that syntax are It can be calculated with a much smaller amount of calculation than the method.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of an apparatus for carrying out the present invention, FIGS. 2 (A) and 2 (B) are flow charts showing an example of the control procedure, and FIGS. 3 (A) and 3 (B) are Fig. 4 is a structural diagram showing an example of a table required when executing the flowchart of Fig. 2. Fig. 4 is the area of TABLE1 that is referenced when calculating the first two subscripts i, j of TABLE1. FIG. 5 is an explanatory diagram showing the order in which TABLE1 is filled with the calculated values when the flowcharts of FIGS. 2A and 2B are executed. SC: Numerical value holding RAM showing the reliability of each clause, BUF ... Clause set sequence buffer memory, PEN ... Consistency calculation device, T1 ... TABLE1 RAM, T2 ... TABLE2 RAM, ADD1 ... Adder, MIN1 … Minimum value detector, ADD2… Adder, MIN2… Minimum value detector, CONT… Control device for controlling the operation sequence of each part, ME
M ... ROM for storing control procedure, CPU ... Central processing unit, i1 ... Phrase reliability input terminal, i2 ... Phrase string input terminal, O1 ... Output terminal of result obtained at T1, O2 ... Of result obtained at T2 Output terminal.

 ─────────────────────────────────────────────────── ─── Continuation of front page (56) Reference JP-A-57-14971 (JP, A) JP-A-60-22275 (JP, A) JP-A-60-157659 (JP, A) JP-A-58- 208847 (JP, A) A. V. Eiho, Translated by Nozaki and Noshita, "Science Library Information Computer = 35 Algorithm Design and Analysis ▲ I ▼" (Published October 5, 1977), Science Co., P. 59-61 IPSJ Journal, Vol. 27, No. 7 (Jully 1986) P.I. 679-687

Claims (1)

(57) [Claims] 1. When a column of a bunsetsu set and a numerical value indicating the reliability of each bunsetsu in the bunsetsu set sequence are given, the sum of degree of dependency matching between two bunsetsus and the reliability of each bunsetsu The optimal clause sequence under the condition that one clause is selected from each clause set by minimizing or maximizing the sum of the numbers that represent, the optimal syntax on it, and the degree of its eligibility. In the language processing method of the language processing apparatus, the two-dimensional upper triangular matrix-shaped first and second tables having the number of rows and columns equal to the length of the clause set column are prepared in the memory. Then, each grid of the first table and the second table is divided into as many terms as the original number of the clause set having a number equal to the column number, and the first table and the second table are preliminarily three-dimensionally divided. In the first table, the i-th clause, the i-th column, and the p-th term in the p-th clause of the i-th clause set A numerical value representing the reliability of the clause is set in the holding circuit, and for k = i, ..., j−1, each item in the i-th row and k-th column in the memory and k + 1-th row, j-th column, q-th term. The calculated value is stored in, and if the storage is done, the calculated value of the i-row, k-column, p-term of the first table in the memory and k + 1-row, j-column, q
The value of the term and the degree of consistency that the p-th clause of the k-th clause set relates to the q-th clause of the j-th clause set are added in the computing unit, and k and p of the addition result are added. The minimum value or the maximum value of the above and the values of k and p that give the minimum value or the maximum value are detected by the detector, and the detected minimum value or the maximum value is detected in i on the memory.
Stored in rows, j columns, and q terms, a set of values of k, which is the optimum segment point that gives the minimum value or maximum value, and p, which is the optimum clause number, is stored in the i of Table 2 on the memory. Stored in row, j column, q term, fill the first table and the second table in the memory with sequentially calculated values, and the first table and the second table in the memory are all calculated When it is filled with the value of, the final degree of eligibility and the optimum clause number in the last clause set are obtained by finding the minimum or maximum value among the terms in the upper right corner of Table 1. At the same time, the entire optimal partition points and optimal clause numbers required for constructing the optimal syntax are obtained in the second table on the memory, and the language processing method of the language processing device.