JP2954215B2 - Language processing system - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Industrial applications]   The present invention provides a Japanese speech recognition device and a Japanese word processor.
A language processing system for post-processing the output of the processor
Especially at some point in the sequence
When multiple phrase candidates are given, certainty of those candidates
And the degree at which multiple clauses are related to another
Considering the degree of consistency of
Phrase candidates at each time so that the optimal phrase sequence is constructed
One by one, select the clause and determine its syntax.
And the Japanese phrase or sentence in the bunsetsu sequence
Language processing system that calculates eligibility
It is. [Conventional technology]   Japanese speech recognizers and Japanese word processors
Order for use in language processing for post-processing
At the next few points,
The probability of those candidates and the dependency between the two phrases
Considering the degree of consistency of
Phrase candidates at each time so that the optimal phrase sequence is constructed
And select the clauses one by one, determine its syntax,
And the resulting Japanese phrase or sentence in the bunsetsu sequence
Algorithm to efficiently calculate eligibility as
Conventionally, for example, the following has been proposed.   Kazuhiko Ozeki: "Multi-stage decision a for selecting the optimal phrase sequence
Algorithm, "IEICE Speech Research Group, SP86-32
(July 1986) [Problems to be solved by the invention]   In Japanese, word order is said to be relatively free
However, in some cases, word order is a problem. Also,
Species cases cannot be associated with one predicate. like this
Phrasal sequence, taking into account the problem of word order and case conflicts
Is eligible as a Japanese sentence or phrase
Is calculated by considering the consistency of the dependency between the two phrases.
It is not enough to take into account that multiple clauses
We must use the consistency of the two clauses.
However, in the above-mentioned conventional method, Japanese phrases or sentences are used.
Eligibility is determined by the sum of the dependencies between the two phrases.
The solution could only be found when determined. Enumeration
In principle, this problem can be solved by using the method
But very difficult to apply to practical problems in terms of complexity
Met.   Therefore, an object of the present invention is to use Japanese sentences or phrases and
Eligibility is not the consistency of the dependency between the two phrases.
And that multiple clauses affect one other clause at the same time.
Even if based on the degree, at some point in the sequence
From the given phrase set,
Choose the appropriate phrase sequence, determine the syntax above it,
And the resulting Japanese phrase or sentence in the bunsetsu sequence
Language processing system that efficiently calculates eligibility
To provide. [Means to solve the problem] (A) Japanese structure in units of phrases   Before describing the configuration of the present invention,
This section describes the structure of Japanese in terms of.   A sentence in Japanese or a set of phrases is simply called a phrase.
Thought to be based on a broader qualifying relationship between positions
Can be For example, [S1] "I go to work on the 7 o'clock train." In the Japanese sentence "I am", "7 o'clock",
"By train", "To company", "Go" are phrases
And "I", "by train" and "to the company" are all
"Go" is qualified, "7 o'clock" is qualified by "train"
By doing so, one sentence is composed.   When clause x modifies clause y, x is related to y and y
Is said to receive x. In addition, such a modification relationship
It is called receiving.   The bunsetsu sequence forms a united phrase or sentence in Japanese
To satisfy the following conditions between the clauses:
It is considered necessary to have a dependency
You. [C1] Clauses other than the last one are at the end of the sentence
Pertains to one of the clauses. [C2] The dependency between two phrases is
Do not cross with the dependency.   Conditions [C1] and [C2] are defined as follows:
Sentence ". [D1] (1) When x is a clause, [x] is “syntax”
You. (2) X₁, X_Two,…, X_mIs "syntax" and x is a clause, [X₁X_Two… X_mx] Is "syntax". [D2] Clause sequence x₁x_Two… X_nWith proper parentheses and syntax
What you did, x₁x_Two… X_nThe above syntax is called. Phrase string x₁
x_Two… X_nThe whole syntax above K (x₁x_Two… X_n) Will be expressed as follows.   Syntax [X₁X_Two… X_mx], X₁= [… X₁], X_Two= [… X_Two],…,
X_m= [… X_m], X₁, x_Two,…, X_mIs related to x
Promise to represent that in the syntax in the above sense
Satisfies the conditions [C1] and [C2], and conversely, the condition [C1]
Dependency relations in a bunsetsu sequence that satisfies
It can be represented by the syntax in the above sense.   By the way, for one phrase string, there are multiple syntaxes
Exists. For example, [[I] [at 7 o'clock] [by train] [to the company] [go
You] [[[[[[I]] 7 o'clock] by train] go to company]] [[[I] [at 7 o'clock] by train] [go to work]] [[[I] [going to the company] [by the [7:00] train]] [[I] [going to the [train]] [to the company] Are all the syntax on the phrase string of the example sentence [S1]. This
Select a well-qualified clause sequence from many syntaxes
To do this, some sort of evaluation function is needed. There
First, the degree of consistency of the dependency is determined as follows
And [C3] Clause x₁, x_Two,…, X_mIs related to clause x
Is a non-negative function PEN (x₁, x_Two,…, X_m, x) Is represented by   The value of PEN is, for example, in the range of 0 to 100.
I promise that the consistency is high. How the function pen
It is a very important issue to determine
The explanation is omitted because it is not the main point.   With the above preparation, the eligibility P (X) of the syntax X is
Is determined recursively as follows. [D3] (1) When X = [x], where x is a clause, P
(X) = 0, (2) X = [X₁X_Two…… X_mx], X₁= [… X₁], X_Two= [… X_Two],…, X_m[… X_m]When, P (X) = P (X₁) + P (X_Two) + ... + P (X_m) + PEN (x₁, X_Two,…, X_m, x)   The value of P (X) defined in this way is
It is the sum of the PEN values for loose dependencies
I have. (B) Question setting   In the following, to make long formulas easier to readInstead of   Σ (i ≦ m ≦ j) [f (m)] Notation is used. The same applies to min, argmin, ∪, etc.
Express as follows. Parentheses [,] are omitted if there is no risk of confusion.
Sometimes.   Now consider the following situation: [J1] At each time point from 1 to N, a set of clauses B₁, B
_Two,…, B_NIs given. In addition, each phrase set B_k
For a function S that takes a nonnegative real value_kIs stipulated
RU:   S_k: B_k→ [0, ∞] S_k(X) is phrase set B_kNumerical value representing the certainty of clause x in
For example, assume that the value takes a value from 0 to 100.
It is assumed that the certainty is high. Taking voice recognition as an example, S_k
(X) is the recognition device, B_kThe recognition result of x in the
It is a numerical value that indicates whether it was recognized with certainty
Some speech recognizers output such values along with the recognition results.
It is designed to help. Also, the date of the kana-kanji conversion method
Taking a Japanese word processor as an example, there are homonyms
Output multiple phrase candidates with the same reading
However, each candidate has a kanji or idiom according to the frequency of use.
A numerical value indicating certainty can be attached. Of syntax X
The certainty and the clause x₁, x_Two,…, X_mThe certainty of
Based on the following definition.   S (X) = Σ (1 ≦ k ≦ m) [S_k(X_k)]   Also, the phrase set sequence A₁, A_Two,…, A_mAgainst   KB (A₁A_Two… A_m) = ｛X | X∈K (x₁x_Two… X_m), X₁∈A₁, x_Two∈A_Two,…, X_m∈A_m｝ Notation is prepared.   Under these circumstances, the problems addressed by the present invention are as follows:
It can be stated as follows. [P1] Find the following. (1) min (X∈KB (B₁B_Two… B_N)) [P (X) + S
(X)] (2) argmin (X∈KB (B₁B_Two… B_N)) [P (X) + S
(X)]   KB (B₁B_Two… B_N) Is a finite set, so the problem is
Physically, it can be solved by the enumeration method. That is,
Any X∈KB (B₁B_Two… B_N), Based on the definition formula   P (X) + S (X) Gives the minimum value and the minimum value by calculating
X can be determined. However, KB (B₁B_Two… B_N)
Increases rapidly with N, as shown in Table 1,
Applying the enumeration method to practical problems due to the huge number
It is extremely difficult to do.   In contrast, according to the present invention, this problem is
It can be solved within the complexity.   Function PEN is   PEN (x₁, x_Two,…, X_m, x)   = PEN (x₁, x) + PEN (x_Two, x) + ... + PEN (x_m, x) And the sum of the degrees of dependency between the two phrases
In such cases, this problem can be solved by the above-mentioned conventional method.
Can be solved efficiently. In contrast, the present invention provides
For cases where the number PEN cannot be decomposed into such a sum
Things. (C) Recursive equation   Here, recursion plays a fundamental role in the present invention.
The equation is described. Embedded in some syntax
Focusing on one syntax, the clauses in it
Can have a dependency relationship with the clause of
It is only a clause. With this in mind, the following definition was made:
You. [D4] 1 ≦ i ≦ j ≦ N and each x∈B_jAgainst (1) OPT (i, j; x) = min (X∈KB (B_iB_{i + 1}… B
_j-1{X}) [P (X) + S (X)] (2) OPK (i, j; x) = argmin (X∈KB (B_iB_{i + 1}… B
_j-1{X}) [P (X) + S (X)]   The following recursive equations hold for OPT and OPK. [E1] (1) OPT (i, i; x) = S_i(X) (2) For i <j   OPT (i, j; x)   = Min (i-1 = k (0) <k (1) <k (2) <...
<K (m) = j−1, x₁∈B_{k (1)}, x_Two∈B_{k (2)},…, X_m∈
B_{k (m)}) [OPT (k (0) + 1, k (1): x₁) + OPT (k (1) +1,
k (2); x_Two) + ... + OPT (k (m-1) + 1, k (m);
x_m) + PEN (x₁, X_Two,…, X_m, x)] + S_j(X) [E2] (1) OPK (i, i; x) = [x] (2) k that gives the minimum value in [E1] for i <i
(0), k (1), k (2), ..., k (m), x₁, x_Two,…, X_mTo
(0), (1), (2), ..., (),₁,_Two,
…,_mThen   OPK (i, j; x)   = [OPK ((0) +1, (1);₁) OPK (
(1) +1, (2);₂)… OPK ((-1) +1,
();_m) X] The proof of [E1] and [E2] is easy if we pay attention to the following facts.
Details are omitted here. [E3] KB (B_iB_{i + 1}… B_j-1{X})   = ∪ (i-1 = k (0) <k (1) <k (2) <... <
k (m) = j-1, x₁∈B_{k (1)}, x_Two∈B_{k (2)},…, X_m∈B_{k (m)}) ｛[X₁X_Two… X_mx] | X₁∈KB (B_{k (o) +1}… B_{k (1)}｛X₁｝), X_Two
∈KB (B_{k (1) +1}… B_{k (2)}｛X_Two｝),…, X_m∈KB (B_{k (m-1) +1}
… B_{k (m)}｛X_m｝)   Now, the syntax [X₁X_Two… X_mx], X₁= [… X₁], X_Two= […
x_Two],…, X_m[… X_m], X₁, x_Two,…, X_mIs related to x
You. That is, x is x₁, x_Two,…, X_mTo receive
It is considered that there is a certain limit to the number of clauses received by one clause.
May be. Then, when this upper limit is L,   1 ≦ m ≦ L The solution may be obtained in the range of. Imposing such restrictions
Then, [E1] becomes as follows. [E1 '] (1) OPT (i, i; x) = S_i(X) (2) For i <j   OPT (i, j; x)   = Min (i-1 = k (0) <k (1) <k (2) <...
<K (m) = j-1,1 ≦ m ≦ L, x₁∈B_{k (1)}, x_Two∈B_{k (2)},…,
x_m∈B_{k (m)}) [OPT (k (0) = 1, k (1); x₁) + OPT
(K (1) + 1, k (2); x_Two) + ... + OPT (k (m-1)
+1, k (m); x_m) + PEN (x₁, x_Two,…, X_m, x)] + S_j(X)
If L ≧ N−1, [E1 ′] and [E1] have the same value. You
That is, [E1 '] includes [E1]. Therefore, [E
[E1 '] is used instead of [1]. [E2] this
It doesn't change even if you put restrictions like. In [E1 ']
K (1), k (2), ..., k (m), x₁, x_Two,…,
x_mIs the optimal partition point for i, j, x and their partition points
Will be called the optimal clause.   Below is B_kLet NUM (k) be the original number of and B_kB to_k=
｛X_{k, 1}, x_{k, 2},…, X_{k, NUM (k)}｝ And enumerated. (D) OPT, optimal segmentation point, optimal clause number calculation method   (2) of [E1 '] is OPT (s, t; x when i <j_{t, p})
(I ≦ s ≦ t ≦ j−1,1 ≦ p ≦ NUM (t)) has already been calculated.
OPT (i, j; x_{j, q}) Can be calculated
ing. When (1) of [E1 '] is used, i = j
Sometimes OPT (i, i; x_{i, q}) = S_i(X_{i, q}) Determines the value
So, after all, OPT (i, j; x_{j, q}) To j
-Calculate starting from the part where i is 0 and gradually increasing
And at the same time determine the optimal segmentation point and optimal clause number
You can go. (E) Optimal syntax and its eligibility calculation method   For simplicity, the optimal breakpoint and optimal clause number are uniquely determined.
Will be described. First,   min (X∈KB (B₁B_Two… B_N)) [P (X) + S (X)]   min (1 ≦ q ≦ NUM (N)) min (x∈KB) (B₁B_Two… B
_N-1｛X_{N, q}｝) [P (X) + S (X)]   = Min (1 ≦ q ≦ NUM (N))] OPT (1, N; x_{N, q})] Because   q0 = argmin (1 ≦ q ≦ NUM (N)) [OPT (1, N;
x_{N, q})] Is calculated to obtain the last clause x of the optimal clause sequence._{N, q0}
And optimality of optimal syntax OPT (1, N; x_{N, q0}) Is determined.   In addition, the optimal clause sequence and the optimal syntax   OPK (1, N; x_{N, q0}) Given by To calculate this more specifically:
You can do it.   1, N, x_{N, q0}K1 (1), k1
(2), ..., k1 (m1), and the corresponding optimal clause number
If p1 (1), p1 (2), ..., p1 (m1), [E2]
By (2)   OPK (1, N; x_{N, q0})   = [OPK (1, k1 (1); x_{k1 (1), p1 (1)}) OPK (k1
(1) + 1, k1 (2); x_{k1 (2), p1 (2)})… OPK (k1
(M-1) + 1, k1 (m); x_{k1 (m), p1 (m)})
X_{N, q0}] Holds. OPK on the right side (k1 (r-1) +1, k1 (r); x
_{k1 (r), p1 (r)}) (1 ≦ r ≦ m1), for example, OPK
(1, k1 (1); x_{k1 (1), p1 (1)}), 1 = k1
(1)   OPK (1, k1 (1); x_{k1 (1), p1 (1)})   = [X_{k1 (1), p1 (1)}] , So OPK (1, N; x_{N, q0}) Is   OPK (1, N; x_{N, q0})   = [[X_{k1 (1), p1 (1)}] OPK (k1 (1) + 1, k1
(2);   x_{k1 (2), p1 (2)}) ... OPK (k1 (m-1) + 1, k1
(M);   x_{k1 (m), p1 (m)}) X_{N, q0}] Can be rewritten. If 1 ≠ k1 (1),
1, k1 (1), x_{k1 (1), p1 (1)}K is the optimal segment point for
2 (1), k2 (2), ..., k2 (m2), corresponding optimal clause number
The symbols are p2 (1), p2 (2), ..., p2 (m2)
When,   OPK (1, k1 (1); x_{k1 (1), p1 (1)})   = [OPK (1, k2 (1); x_{k2 (1), p2 (1)}) OPK (k2
(1) + 1, k2 (2); x_{k2 (2), p2 (2)})… OPK (k2
(M-1) + 1, k2 (m); x_{k2 (m), p2 (m)}) X
_{k1 (2), p1 (2)}] , So OPK (1, N; x_{N, q0}]   OPK (1, N; x_{N, q0})   = [[OPK (1, k2 (1); x_{k2 (1), p2 (1)}) OPK (k
2 (1) + 1, k2 (2); x_{k2 (2), p2 (2)})… OPK (k2
(M-1) + 1, k2 (m); x_{k2 (m), p2 (m)}) X
_{k1 (2), p1 (2)}] OPK [k1 (1) -1, k1 (2); x
_{k1 (2), p1 (2)}]   OPK (k1 (m-1) + 1, k1 (m);
x_{k1 (m), p1 (m)}) X_{N, q0}] Can be rewritten. Such operations appear
Repeat until all OPKs have a single clause syntax
If you return, OPK (1, N; x_{N, q0}), That is, the optimal phrase sequence and its
Can be determined at the same time.   i and j are phrase set numbers, and x is B_jIf the phrase belongs to
, K (1), k (2),..., K
(M), and optimal clause x₁, x_Two,…, X_mThere are multiple pairs of
At that time, then the above
Perform the above operation and enumerate all the obtained syntaxes.   The present invention is applicable not only to Japanese
Can be described in the same way as Japanese, like a word
Needless to say, it can be applied to foreign languages that have a grammatical structure.
Absent. [Action]   According to the present invention, a given phrase set sequence B₁, B_Two,…, B_N
Subsequence B of_i, B_{i + 1},…, B_jFixed the last clause for
Optimal bunsetsu sequence, optimal syntax on it, and its optimal
Degrees are determined sequentially from the ones corresponding to substrings with shorter lengths
Memorize it and include those substrings longer
Use them when calculating the same for subsequences.
Efficiency without having to repeat the same calculations
It is possible to obtain the expected result. [Example]   Hereinafter, the present invention will be described in detail with reference to the drawings.   FIG. 1 shows an embodiment of an apparatus for carrying out the present invention. No.
In FIG. 1, SC is a flowchart shown in FIG. 2 (B).
To achieve the table score in step S5 in
Buffer memory such as RAM of the
Each sentence x input from_{j, q}Of certainty S_j(X_{j, q}Hold)
It is for. BUF is input from the phrase input terminal i2.
Memory such as RAM that holds a set of clauses
It is. For example, when using the present invention for speech recognition,
Each phrase candidate output as a recognition result from the device is connected to terminal i2.
From the terminal i1 and the certainty associated with those clauses
Input from T1 and T2 are respectively shown in FIG.
Table B and table2 shown in (B) are realized.
This is a memory such as a RAM. COMBI is phrase set B₁, B
_Two,…, B_NSome phrase sets B from_{k (1)}, B_{k (2)},…, B
_{k (m)}And select it from the selected phrase set
Each phrase x_{k (1), p (1)}, x_{k (2), p (2)},…, X
_{k (m), p (m)}With a device that generates combinations to select
is there. SEL1 is the combination generator C from the buffer memory BUF
A device that selects only specific phrases specified by OMBI
is there. PE is transferred from the buffer memory BUF via the selector SEL1.
Sentence x read out₁, x_Two,…, X_m, x, PEN (x₁, x
_Two,…, X_m, x). SEL2 is memory T1
Specific information specified by the combination generator COMBI.
It is a device that selects only information. ADD1 is a PEN calculator PE
Output and read from memory T1 by selector SEL2
This is an adder for adding the calculated values. MIN occurs in combination
Adder ADD when device COMBI generates various combinations
Minimum value of output of 1 and minimum to detect the combination at that time
It is a value detector. ADD2 is connected to the output of the
This is an adder that adds a specific numerical value in the
You.   CONT is a control device for controlling the operation order of these parts.
For example, a central processing unit CPU and control means for each unit.
Memory MEM1 in the form of ROM for storing the order in advance
And a memory MEM2 in the form of a working RAM. 01 you
And 02 are the calculation results written to memories T1 and T2, respectively.
This is the output terminal that outputs the result.   2 (A) and (B) show the embodiment shown in FIG.
Control methods previously stored in the memory MEM1
Eligibility of optimal syntax on optimal clause sequence as an example of order
To determine the degree, optimal clause sequence, and optimal syntax
A procedure for sequentially finding the optimal segmentation point and optimal clause number pairs
It is a flowchart shown. This is described below.
You.   Attached to the flow charts of FIGS. 2 (A) and (B)
As shown in FIGS. 3A and 3B,
Number of rows and columns equal to the length of the clause column N, and the j-th column
With a term equal to the original number NUM (j) of the clause set
Two three-dimensional tables table1 (i, j, q) and table
2 (i, j, q) (1 ≦ i ≦ j ≦ N, 1 ≦ q ≦ NUM (j)) is required
It is. Subscripts in each table indicate rows, columns, and terms in order from the left.
You.   table1 (i, j, q) is OPT (i, j; x_{j, q}) Value and also tabl
e2 (i, j, q) is i, j, x_{j, q}Segmentation points and optimal clauses for
This is for storing a set of numbers. Kth clause set B_kof
The original number NUM (k) is entered into the one-dimensional table num (k).
Is retained. Kth clause set B_kThe p-th clause x in_{k, p}Sure
Reality is input and stored in a two-dimensional table score (k, p)
It is. Also, PEN (x_{k1, p1}, x_{k2, p2},…, X_{km, pm},
x_{j, q}) Is calculated by pen (k1, p1), (k2, p2),
..., (km, pm), (i, q)).   In the flowcharts of FIGS. 2 (A) and (B)
In steps S1 to S13, each table
Column number j from 1 is increased by 1 to N,
Perform the following operations on the column:   Rows of each table in steps S2 to S11
Starting with j, decrease the number i by 1 to 1 and add
Then, the following processing is performed.   Item number of each table from step S3 to step S9
Start q from 1 and increase by 1 to num (j)
The following processing is performed on the q term. (1) If it is determined in step S4 that i ≠ j,
If so, proceed to step S5 to execute the next [F1], and then
[F2] is executed in step S6. [F1] table1 (i, j, q)   : min (i ≦ k₁<K_Two<… <K_m= J-1,1 ≦ m ≦ L, 1 ≦ p₁≤
num (1), 1 ≦ p_Two≤num (2), ..., 1≤p_m≦ num (m))
[Table1 (i, k₁, p₁) + Table1 (k₁+1 and k_Two, p_Two) + ... + ta
ble1 (k (m-1) + 1, k_m, p_m) + Pen ((k₁, p₁), (K
_Two, p_Two),…, (_m, p_m), (J, q))] + score (j, q) [F2] table2 (i, j, q)   : Give the minimum value in [F1] (k₁, p₁), (K_Two,
p_Two),…, (K_m, p_m) (2) If i = j is determined in step S4
For example, the process proceeds to S7, where the next [F3] is executed. [F3] table1 (i, j, q): = score (i, q)   The occurrence of the combination in [F2] is the combination shown in FIG.
It is performed by the generator COMBI. About their combination
The minimum value and the combination that gives the minimum value are detected as the minimum value.
It is performed at the container MIN. Calculate PEN with PEN calculator PE
The selection of the clauses necessary for the selection is made by the selector SEL1.
And table1 (i, k₁, p₁), Table1 (k₁+1 and k_Two, p_Two),…, Ta
ble1 (k_m-1) + 1, k_m, p_mThe value of () is read out by the selection device SEL.
Done by two. Also, the value of num (k) is
It is kept in F. With the above processing, table1 and table2
Apply the above calculation to each row, column, and term in
Write to ble1 and table2.   The calculation ends when j> N in step S13.
OPT (1, N, x) is stored in table1 (1, N, q)_{N, q}), (1 ≦ q
≤ NUM (N)). Also, table2 has
Since the information of the appropriate segmentation point and the optimal clause number is stored,
(4) The optimal sentence is obtained from this information by the method described in (e).
Clause sequences and optimal syntax can be constructed.   When actually using the present invention, the apparatus shown in FIG.
And the flowchart shown in FIGS. 2 (A) and (B)
In addition to the table, the optimal phrase sequence and the
Although a mechanism for constructing a proper syntax is necessary, the main point of the present invention is
The point is where the contents of table1 and table2 are calculated.
From this information, the optimal phrase sequence and
The mechanism for constructing a simple syntax is only described above.
You.   However, if the contents of table1 and table2 can be calculated
If the optimal phrase sequence and its
Of the calculations required to construct the optimal syntax above,
Note that the computationally intensive part is no longer done
Keep it.   In [F1], a pair of numerical pairs giving the minimum value ((k₁, p
₁), (K_Two, p_Two),…, (K_m, p_m))
At that time, in table2 (i, j, q)
So that you can memorize the pairs of numerical values of
And store them all in table2 (i, j, q)
do it. Thus, the flow charts shown in FIGS.
Even if you change the chart, the amount of calculation hardly changes
No.   In the above-described embodiment, the processing for finding the minimum value
But these are S_kThe smaller the value of (x), the more clauses
The higher the certainty of x and the smaller the value of PEN, the better the dependency
This is because the degree was high. If S_kThe value of (x) is large
The higher the certainty, the higher the value of PEN
If the degree of consistency is high, the process of finding the optimum value instead of the minimum value
Just do it. [The invention's effect]   As mentioned above, according to the present invention, a given sentence
Clause set B₁, B_Two,…, B_NSubsequence B of_i, B_{i + 1},…, B_jThe most
The optimal phrase sequence when the later phrase is fixed and the optimal
The syntax and its eligibility correspond to short-length substrings
Find them sequentially from the thing and memorize it, and those parts
Calculate similar for longer subsequences including columns
Sometimes by using them, we can perform the same calculations
The desired result can be obtained efficiently without returning.   Multiple clauses x₁, x_Two,…, X_mAt the same time, one phrase x
The degree of consistency PEN (x₁, x_Two,…, X_m, x)
The attributes of the words that make up and the dependencies that appear in the actual sentence
Can be calculated based on statistical information such as
You. The amount of calculation also depends on the language dictionary construction method, etc.
However, as a guide, the present invention
And the number of addition and comparison operations in the enumeration method
Value. (1) To calculate PEN (x, y), a total of J times
Requires arithmetic. (2) PEN (x₁, x_Two,…, X_m, x) to calculate PEN
(X₁, x) + PEN (x_Two, x) + ... + PEN (x_m, x)
The same amount of calculation as, ie, addition (J + 1) · m-1 times
Requires the computational complexity of   Further assume that the original number of clause sets is all equal to M.
You. Then, a parameter that determines the size of the problem to be solved
The data is that, in addition to J,
You.   M: number of elements in each clause set   N: Length of phrase string   L: Upper limit of the number of phrases that can be related to one phrase at the same time   Under the above assumptions, the amount of calculation is as follows. (A) The present invention (1) Addition   _iC_jIs a binomial coefficient, and the function f (n) is If you define (2) Comparison   Function g (n) If you define (B) The enumeration method   knum (n, L) in the dependency structure on a clause sequence of length n
And the number of clauses pertaining to one clause at the same time is L or less
Assuming that (1) Addition   Total number of additions = ｛knum (N, L) + (J + 1) · (N−
1) + (N-1)｝ · M^N (2) Comparison   Total number of comparisons = knum (N, L) M^N-1   knum (n, L) can be calculated using the following recurrence formula:
Wear.  J = 1 for the total number of additions and the total number of comparisons
The calculated values for some M, N, and L are listed in Tables 2 and 3.
I can.   According to these tables, the effect of the present invention is clear,
For example, when M = 5, N = 20, L = 5, the calculation amount is enumerated.
About 10^FifteenIt is reduced by a factor of one.

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a block diagram showing an embodiment of an apparatus for carrying out the present invention, FIGS. 2 (A) and (B) are flowcharts showing an example of the control procedure, FIG. FIGS. 3A and 3B are table structure diagrams showing an example of a table required when executing the flowchart of FIG. SC: RAM for holding phrase certainty, BUF: RAM for holding phrase set, T1: RAM for table1, T2: RAM for table2, SEL1: Data selector, SEL2 ... Data selector, PE ... Dependency matching degree calculator, COMBI …… Combination generator, ADD1 …… Adder, ADD2 …… Adder, MIN …… Minimum value detector, CPU …… Central processing unit, MEM1 …… ROM for storing control procedures , MEM2: CPU working RAM, CONT: Control device for controlling the operation order of each part, i1: ... clause certainty input terminal, i2 ...: clause input terminal, 01: output of results obtained in memory T1 Terminal, 02 ... Output terminal of the result obtained in memory T2.

Continuation of front page       (56) References IEICE Technical Report, Vol.               86, no. 93 (1986. 7.16), p. 41               -48 (SP86-32)                 IEICE Technical Report, Vol.               86, no. 221 (November 20, 1986), p.               47-57 (COMP86-47)

Claims (1)

(57) [Claims] Given a sequence of clause sets and a numerical value representing the certainty of each clause in the clause set string, the sum of the consistency of the fact that a plurality of clauses simultaneously belong to one clause and the certainty of each clause. By minimizing or maximizing, the optimal clause sequence under the condition that one clause is selected from each clause set, the optimal syntax on the clause sequence, and the degree of eligibility are determined. In the language processing system, a first buffer memory means for holding the certainty of each clause in each of the input phrase sets, a second buffer memory means for holding each of the input phrase sets, There is a first and second table in the form of a two-dimensional upper triangular matrix having a number of rows and a number of columns equal to the column length N, wherein the first table and the second table are Each cell is defined as the element of the phrase set with the number equal to the column number j. A memory configured by dividing the first table and the second table into three dimensions by dividing the term into a number NUM (j), and an i-th row, an i-column, and an r term of each of the first table Means for storing the certainty factor of the r-th clause in the i-th clause set, a predetermined integer L smaller than N, and 1i <jN
Integer m that satisfies 1 mL for integer i, j that satisfies
And i-1 = k (0) <k (1) <k (2) <... <
integers k (0), k (1), k satisfying k (m) = j-1
(2),..., K (m) and an integer p that satisfies 1p (n) NUM (k (n)) for each n of 1 nm
(N) (m, k (0), k (1), k (2), ..., k
(M), p (1), p (2),..., P (m)), and a set of integers (m, k) generated by the combination generating means.
(0), k (1), k (2), ..., k (m), p (1), p
(2),..., P (m)), for each integer n that satisfies 1 nm, the k (n) -th clause set among the clause sets held in the second buffer memory means. First selecting means for selecting the p (n) -th clause of the set, and a set of integers (m, k) generated by the combination generating means
(0), k (1), k (2), ..., k (m), p (1), p
(2),..., P (m)), for each integer n satisfying 1 nm, k (n−1) +
A second selecting means for selecting a p (n) term in one row and a k (n) th column, and a p (n) clause in each k (n) clause set selected by the first selecting means Is calculated by the calculation means for calculating the degree of consistency of simultaneously relating to the q-th clause in the j-th clause set held in the second buffer memory means, and the output of the calculation means and the second selection means are selected. And the k-th (n-1) + 1-th row and the k-th
(N) a first adding means for calculating the sum of the numerical values held in the p (n) term of the column, and a set of integers (m , k (0), k (1), k
(2), ..., k (m), p (1), p (2), ..., p
(M)) a set of minimum values for all of the above and an integer giving the minimum value A minimum value detecting means for calculating the minimum value, and a first value which is a first output of the minimum value detecting means.
Q-th in the j-th clause set held in the buffer memory means
Second adding means for adding the certainty factor of the clause, and outputting the output of the second adding means to the i-th row, j-th row of the first table.
A set of means for storing in a column and a q-th term, and a set of integers as a second output of the minimum value detecting means In the i-th row, the j-th column, and the q-th term of the second table; and filling each row, column, and term of the first and second tables with sequentially calculated values. Calculation order control means for controlling the values of the integers i, j, and q in order to go on; when the first table and the second table are all filled with the calculated values, the first By calculating the minimum value in each term in the first row and the Nth column of the table and the term number giving the minimum value, the degree of eligibility of the optimal syntax on the optimal clause column and the final clause set in the final clause set are determined. The means to obtain the optimal clause number and the set of integers required to construct the optimal syntax Means for reading from each row, column, and term of the second table.