JPH02163876A - Literature retrieving method using neural network model - Google Patents

Abstract

PURPOSE:To enable a retrieving user to perform the desired retrieving jobs by using the calculation formulas which are connected to each other by securing the weighted average of the synthesization result containing functions of different types of character with use of parameters. CONSTITUTION:The calculation formulas which are connected to each other by securing the weighted average of the synthesization result containing two functions (f) and (g) with use of parameters alpha1 and alpha2 when a state r3 of a unit (j) of an output layer and a state a1 of a unit (i) of an intermediate layer are calculated. Thus the 'distance' can be shown more accurately between a key word group offered by a retrieving user and a key word group assigned previously to the literature. Then both functions (f) and (g) having difference types of character compensate the defects with each other. A retrieving standard is obtained in response to the level and the purpose of the user by changing the values of both parameters alpha1 and alpha2. Thus the user can retrieve the literature as desired.

Description

DETAILED DESCRIPTION OF THE INVENTION Field of the Invention The present invention relates to a document search method using a neural network model for document search or data search.

Conventional technology In recent years, there have been remarkable advances in information processing technology.
One of them is related to literature search. For example, “A FUZZYDOCUMENT”
TRE-RIEVAL SYSTEM AND
LEARNING METHOD BASED ON
THE DYNAMICKE'flORD CONNE
CTION METHOD (= Intelligent document retrieval system using dynamic keyword connections) (International Workshop on Applications of Fuzzy Systems 188.8.2
2 to 8.24)" (hereinafter referred to as document 1). Roughly speaking, this is the keyword group Q given by the search user and the keywords given in advance to each document i. 11 Closeness with keyword group A" = depth of relationship (file accuracy) r+.

Relevance W between keywords. (However, -jE-A + +
It is calculated by k-rI (1-X,) and searches for related documents. Relevance between keywords (
Link weight) WJk is learned by steepest descent method.

Problems to be Solved by the Invention However, the method disclosed in Document 1 has the following drawbacks in the process of calculating the file accuracy r from the degree of association WJk between keywords. first,
If only one of the keywords presented by the search user is included in the keywords assigned to a document, that document will be selected.

This happens even if the degree of relevance between all other keywords is 0, or if a search user presents a keyword that is not included in the keyword group assigned to one document but is quite similar to it, the document file The accuracy does not reflect the appropriateness of the keyword, and furthermore, the document may be selected even when the keyword group assigned to the document does not include a keyword that has a high degree of relevance to the keyword group presented by the search user. There are many.

Furthermore, a keyword group A is assigned to each document.

There is good luck in how you choose.

Means for Solving the Problems In the invention as set forth in claim 1, an input layer having a unit k corresponding to each keyword and having a keyword group specified by a search user as an input vector unit group (b,); A middle layer consisting of a unit group (a,) which has the same structure as the unit group (bk) and has a unit i corresponding to each keyword and has a real value such that O≦al≦1, and a unit j corresponding to each document j. It has a three-layer structure with an output layer, and each unit i in the intermediate layer has a link with each unit k in the input layer, and unit j in the output layer and unit i in the intermediate layer have a link. , a state a indicating the degree to which keyword i in unit i of the intermediate layer is related to the keyword group presented by the user, and a state a indicating whether document j in unit j of the output layer is detected. In a document search method using a neural network model that searches for documents by calculating a state r indicating the degree of the document.

Each unit i of the intermediate layer and each unit k of the input layer
Let Wik be the weight indicating the degree of association between keywords held by the link between the document j and the keyword held by the link between the unit j of the output layer and the unit i of the intermediate layer. Vj is the weight indicating the degree of association between
i (however, Vji is positive, O1 is a negative real value), and 3't
+ yz, h1+ hz are parameters, function g is the function f ((VJ+a+))=i n (1-Vjtaυ, then the above rj and the above a are parameters α1.α2
(However, using O≦α1≦1, 0≦α2≦1), ra=αz・g((at);(Vji)+'/zt
hJ+(1(h)f((VJta+))a+=αt”g
((bJ ; (v+mL Yz+hJ +(1-αJ
f((calculated by W+*bJ+mJ.

In this case, in the invention according to claim 2, the output layer unit j
The weight vjl indicating the degree of association between the document j and the keyword i, which the link between and the intermediate layer unit i, is allowed to take any real value, and the value is set by learning.

According to the invention described in claim 1, in calculating the state rj of the unit j in the output layer and the state 'MA a t of the unit i in the intermediate layer, the synthesis result by the two functions fig is
1. Since a calculation formula is used that is combined by taking a weighted average using the parameter α2, it is possible to more accurately calculate the "distance" between the keyword group presented by the search user and the keyword group assigned to the document in advance. I can express it. That is,
Function g and function f, which have different properties, complement each other's shortcomings, and depending on the values of parameters αi and α2, search criteria can be set according to the level and purpose of the search user, and search usage can be improved. The search can be performed as intended by the user.

At this time, according to the invention described in claim 2, the link weight v
Since the JI value is set by learning by presenting sample data and adjusting weights using the steepest descent method, it is possible to obtain search results that are close to what the search user has in mind.

Example An embodiment of the present invention will be described based on the drawings.

First, prior to this embodiment, the contents of the prerequisite "Intellectual Document Creation System Using Dynamic Keyword Connections" (hereinafter referred to as Document 2) of Document 1, which has already been proposed by the applicant of the present invention, will be explained. In the system of Document 2, searching for documents closely related to the keyword set specified by the search user consists of the following two steps.

Step 1: The search user selects the optimal set of keywords for the search with the support of the system.

Step 2 Calculate the "closeness" (file accuracy) to each document based on the set of search keywords selected in step 1.

Among these, the contents of step 2 are important, and only step 2 will be explained below. First, as shown in Fig. 2, the degree of association between the keywords and is set to W,1 (however, O≦W IIm≦1). Therefore, if W, , = O, the keyword Wik=1 means that there is the strongest relationship between the keywords , , and . Here, a keyword set is assigned in advance to each document i to be searched, and this keyword set is assigned to A.

In this case, if the keyword set specified by the search user is Q, the "closeness (file accuracy)" between the i-th document and Q is r 1 (Q) = eWJ*...・
・・・・・・・・・・・・・・・・・・ Calculated by (1). Here, Φ is the algebraic sum eXr=1-n (1-XJ)...
・・・・・・・・・・・・This r=( representing (2)
Looking at the calculation formula for Q), OR search of multiple keywords,
It is thought that this is achieved by utilizing the MAX operation characteristic of algebraic sums.

Also, the weight W of links between keywords. is learned using the following method. In this method, an evaluation function is determined for each document, and the steepest descent method is used by looking at the function of weights W, . That is, the teacher pattern is r' (= 3° (if appropriate), r:
= i (inappropriate case), then E 7=-L-(,, 7-r,)2 = (r-d-r+) upper unit L L-Wik ・・・・・・・・・ ・・・・・・・・(3) (However,
m EA + , n EQ , and Wik≠0
To make such an evaluation function global for all documents, E=ΣE; In the above case, Wl, ≠ 0 and Wl ,≠1, and O<W,
, <1 has been explained, but the cases where W, , ≦0 and the cases where Wゎ, ≦1 need to be considered separately. However, for the sake of simplicity, the following explanation will also be based on the assumption that o<w, , <i.

By the way, the system according to Document 2 mentioned above is a typical model of a neural network, as shown in FIG. It can be associated with a multilayer network model.

First, input layer 1 as shown in the figure. Consider a three-layer neural network model consisting of an intermediate N2 and an output layer 3.

The input layer 1 and the intermediate N2 have units k and i (indicated by O marks) corresponding to each keyword.The input layer 1 has a set of keywords specified by the search user, that is, a keyword group.
Input as a vector (bm) whose component elements are binary values of 0 or 1. Each unit group (a,) of the intermediate layer 2 is connected to the input layer 1
It has the same structure as the unit group (b2), and each unit corresponds to each keyword, except that it takes a real value of 0≦a and ≦1. Each unit i of the intermediate M2 has a link with each unit k of the input layer 1, and its weight (that is, the degree of association between keywords) is written as Wlm.

This weight W is a real number satisfying O≦Wik≦1.

vl: W, , is fixed as ≦1. Each unit of output layer 3
corresponds to each document j, and the state rj of the unit j
(However, a real value of O≦r, ≦1) represents the strength with which document j is selected by the search. Unit j of the output layer 3 and unit i of the intermediate layer 2 are linked, and the link Let vJl be the strength of keyword i in the literature, that is, the strength associated with keyword i in the literature, this vJI is generally positive, 0. Takes a negative real value.

Therefore, when searching for one document, the value of the file accuracy rj is calculated, but first, according to the above-mentioned document 2, the value of (Vji) is O or 1, which is fixed for each document. There is. Because, if.

iεA, then V. ≦1, and if every i is A J, then V J t = O. Therefore, the file accuracy rj is r4=1 n (IVJta+)...
・・・・・・・・・・・・・・・(4) a r= 1 r
I (I Wtkb m) ・・・・・・・・・・・・
It can be considered that it is calculated according to (5). This calculation side has multiple evidences (y,) (however,
To reach a certain conclusion from 0≦y,≦1), Max (
M, AX) Arithmetic function f((y+))≦1−n (1
-y+) can be considered to be used as the synthesis side. but,
When viewed as an inference method, it is a fairly ``optimistic'' inference method.

This is because, as mentioned above in the item of 1 assignment.

If there is any one piece of positive evidence (Y+:)'+#1), even if all the other evidence is negative (Vj, +: yI':O), the conclusion is affirmative (i.e. , f ((y i)
)≠1) It will be done. Also, all the evidence must be negative to refute a conclusion. From the perspective of literature search, it may be desirable to select all documents that contain at least one keyword that is highly related to the set of keywords specified by the search user, but in reality, there are the following drawbacks: be.

Now, suppose there are n pieces of evidence around 0.5, and other pieces of evidence are 0.
shall be. Then.

f((y+))41(1-0,5)'=1-Ln, but when n=6, f((y+))white 0,98.

This means that there are 6 pieces of uncertain information of 0.5, and all other information is completely negated (i.e., y
Even though r = O), a rather abject conclusion of 0.98 is reached. Furthermore, when it is desired to quantitatively evaluate the final "closeness (r, ) + t", all of the values are close to 1, and this synthesis is inappropriate.

On the other hand, the composition f ((y +)) by such an algebraic sum
In contrast, the synthesis method used in the multilayer network model is a 5-sigmoid function g ((y,)).

That is, in the case of the multilayer network model shown in FIG.

It can be expressed as -"ja*V, y2.ho, h, phi are parameters, respectively.

Now, consider the meaning of the sigmoid function in equation (6). When the parameter y0 is made small, as shown in Fig. 3,
'! of the curve representing g ((y +)). -ho The rise near o becomes sharp and approaches a step function, therefore.

When the parameter y0 is small, g ((y +)) has the characteristics of a threshold function. In this case, if ha~1 is set, how many links close to 1 exist among (W, ,) Even if the parameter y0 is large, the composition result g((yd) becomes close to 1 and resembles the property of an algebraic sum.On the other hand, when the parameter y0 is large, , and is suitable for quantitatively evaluating file accuracy. However, the "M A Can not.

Therefore, the characteristics of the above two synthesis methods for calculating file accuracy can be summarized as follows.

■ Literature 2 method % formula %) Not suitable for this quantitative evaluation Even if there is only one book, related documents will be searched ■ Neural network method Weighted sum ten sigmoid function Features This book is suitable for quantitative evaluation (especially w ,,y, are not in the vicinity of O and 1) This MAX operation is suitable for calculating the closeness between the entire book keyword group, which cannot be expressed, and the literature.However, in this example, these two ideas are This paper proposes a combined synthesis method.

That is, by superimposing r, = (Xi' g
((a +) ; (VjiL 3'll hJ +
(1as) f ((V, a 1))・・・・・・・・・
・・・・・・・・・・・・・・・・・・(9) al=
ax'g((bJ:(Vlb)+V, t h□)+(
1(!,)f((W+kbyamayu,)・・・・・・・・・
By (10), the state r of unit j in the output layer 3 and the state a1 of unit i in the intermediate layer 2 are calculated. be. here-
Both rj+a+ are 0≦r.

Take a real value that satisfies ≦1.0≦a, ≦1, and input N1
Nosu takes two values (0, 1). Also, the weighting coefficient Wl
k takes a real value satisfying O≦WLk≦1, and vJI generally takes an arbitrary real value.

Furthermore, the synthesis function g used in these equations (9) and (10) is a sigmoid function often used in multilayer neural network models, as described above.

・・・・・・・・・・・・・・・(11)* Y
The value of the parameter zs h□ is determined later by learning. On the other hand, the function f is a function according to Document 2.

f((VJ+ai))=I II(I VJ+a+
) ・------...(12).

In this way, calculations using equations (9) and (10) in this example are as follows:
The calculation is performed using a function that is a linear combination of the function g and the function f. The function g is a function that maps the weighted sum of inputs between (0, 1), and is suitable for comprehensively evaluating the entire input keyword group to determine the state.Furthermore, the function f has the character of MAX operation and is a function g
This is difficult to express. Therefore, (9) (10)
The equation complementarily compensates for the shortcomings of the functions f and g, and the parameters α0. Search users can choose a strategy depending on how α2 is taken. That is, α1. For the value of α2, two strategies can be selected by selecting 0 or 1, the one-unit multi-phase method and the Sumiko sigmoid method, and furthermore, α□
, α2 between O and 1, an intermediate synthesis method is also possible.

For example, if you set αi Cape 01 α2≠O, (9
) (10) is an algebraic sum method using the function f, and one of the keywords assigned to the target document is
If there is even one item that is closely related to a document, that document is searched. This is considered effective when search users are unable to select appropriate keywords.

α□Valve 1. When α2 is set to 1, equations (9) and (10) become a summative sigmoid method using the function g, which indicates the "overall closeness" between the keyword group presented by the search user and the keyword group assigned to the document. It is considered to be effective when it is necessary to quantitatively compare ``. Therefore, it is considered to be effective when the search user is sufficiently confident in selecting keywords.

Furthermore, α1 press 1. If α2to0 is set, Q and ゛′
A comprehensively similar group of keywords is selected, and all documents that have a high degree of relevance to the keywords included therein are searched.

Also, if αX ”F Ol α2to1 is set, Q
Searches for documents that are ``comprehensively similar'' to a keyword group that has a high degree of relevance to keywords belonging to .

In this way, the parameter α1. By setting the value of α2,
Since functions gyf with two different properties can be selected, the search criteria can be changed according to the level and purpose of the search user.

FIG. 4 shows the processing procedure of the system at the time of search.

By the way, according to the system of this embodiment, it is possible to set the link weight coefficients (VJl) and (v+m) to optimal values through learning (however, V+:WI
(fix t=1). The learning method is to apply the steepest descent method to the teacher's input and output pattern c<b: >)ccr: )), and calculate the weighting coefficients (vjt), (w+, ), etc.
Parameters hi, h2. It is only necessary to change the y-force 13'2. FIG. 5 shows the processing procedure during such learning.

More specifically, the following processing is performed. First, let the output pattern for the teacher be (r,), define an evaluation function as a measure matching the output pattern, and use the steepest descent method for this evaluation function.

The result is as follows.

=0 (when m = n) aE9 Δy・“−aYx aE′ Δy゛”]π x z(ΣVjl””/x−(ΣVJ+a+ ho)
)YzraVe 1 1 r aa+ + (1-α2suTryodo AJI-) However, L = -Hedge Rainbow.

aYI Vsahx As described above, according to this processing method, the degree of association between keywords (W+m) - the degree of association between keywords and documents (vj
I) and other parameters (hl, h2゜y1+
3') can be automatically set to an appropriate value through learning, and search results that are close to what the search user is thinking can be obtained. Furthermore, at this time, (v
Since Jl) can also take a negative real value, it can also express the exclusivity between a certain keyword and a document (specifying a certain keyword makes it difficult to search for a certain document). That is, in the case of the document 2 method, the weight of the link between one document and the keyword group added to it was fixed to the binary value of (0, 1), but according to the method of this embodiment, (vjI) is a real number. Since it has been extended to , it can also express suppressive effects, making it a more dynamic model.

Effects of the Invention Since the present invention is configured as described above, according to the invention described in claim 1, the state r of the output layer unit, the state a of the intermediate layer unit, and In the calculation of α1. Since a calculation formula that is combined by taking a weighted average using the parameter α2 is used, the functions g and f, which have different properties, complement each other's shortcomings, and the search user's presentation The "distance" between the keyword group and the keyword group assigned to the document in advance can be expressed more accurately, and the parameter α0. Depending on how the value of α2 is taken, it is possible to set the search criteria according to the level and purpose of the search user and to perform the search as intended by the search user.In this case, according to the invention as claimed in claim 2, Link weight vJ,
The value is set by learning by presenting sample data and adjusting the weights using steepest descent method, so it is possible to obtain search results close to what the search user is thinking. In order to obtain this information, it is also possible to express the exclusivity between a certain keyword and a document.

J, 1 Diagram (o, 1 -A Z diagram

[Brief explanation of the drawing]

The drawings show an embodiment of the present invention, and FIG. 1 is a structural diagram of a neural network model, FIG. 2 is a structural diagram of the method of document 1, and FIG. 3 is a g(y) characteristic diagram. FIG. 4 is a flowchart showing the search process, and FIG. 5 is a flowchart showing the learning process. 1... Input layer, 2... Intermediate layer, 3... Output layer (ΣW lacquer n)

Claims (2)

[Claims] 1. Input the keyword group specified by the search user with unit k corresponding to each keyword.
b_k}, an intermediate layer consisting of a unit group {a_i} which has the same structure as the input vector unit group {b_k}, has a unit i corresponding to each keyword, and has a real value such that 0≦a_i≦1; It has a three-layer structure with an output layer having a unit j corresponding to each document j, each unit i in the intermediate layer has a link with each unit k in the input layer, and a unit j in the output layer has a link with each unit k in the input layer. A state a_i indicating the degree to which the keyword i in the intermediate layer unit i has a link with the keyword group presented by the user, and a state a_i in the output layer unit j. In a literature search method using a neural network model in which a document is searched by calculating a state r_j indicating the extent to which document j is detected, each unit i in the intermediate layer and each unit in the input layer are Let W_i be the weight indicating the degree of association between keywords of the link between
_k (where 0≦W_i_k≦1), and the weight indicating the degree of association between the document j and the keyword i held by the link between the unit j of the output layer and the unit i of the intermediate layer is V_
j_i (where V_j_i is a positive, 0, or negative real value), y_1, y_2, h_1, h_2 are parameters, the function g is ▲There are mathematical formulas, chemical formulas, tables, etc.▼, and the function f is ▲mathematical formulas, chemical formulas, There is a table etc. ▼ When we set the above r_j and the above a_i to the parameter α
_1, α_2 (however, 0≦α_1≦1, 0≦α_2≦1
), r_j=α_2・g({a_i};{V_j_i},y
_2, h_2)+(1-α_2)f({V_j_ia_
i})a_i=α_1・g({b_k};{V_i_k
},y_1,h_1)+(1-α_1)f({W_i_
1. A literature search method using a neural network model, characterized in that calculation is performed by (kb_k}_i_■_k). 2. The weight V_j_i, which indicates the degree of association between document j and keyword i in the link between unit j in the output layer and unit i in the intermediate layer, is allowed to take any real value, and the value is set by learning. 2. A literature search method using a neural network model according to claim 1.