JPH07129593A - Text selecting device - Google Patents

Abstract

PURPOSE:To provide a text selecting device capable of selecting and providing a text which is supposed to be valuable for user. CONSTITUTION:After extracting a key word from the objective text by a key word extracting part 2 and calculating a node active value on the key word associative network of the key word by an active value calculation part 4, the device propagates the active value onto the key word associative network, calculates the score of the objective text based on the active value of all the nodes on the key word associative network after active propagation and information on weight and displays the text on a display part 7 when the calculated score is not less than a prescribed threshold value.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a text selection device, and more particularly to a text selection device for selecting a text that is of value to a user from newly added texts, for example, by using a keyword association network. Regarding

[0002]

2. Description of the Related Art In recent years, various kinds of information have been flooded, and the number of documents has increased. Many of these documents are useful to the user. However, even if one tries to use these documents, the number of documents is so large that it is difficult to accurately access the necessary documents.

Therefore, in order to select a useful one for a user from a vast amount of documents, a text selecting device for selecting using a text search means is being used. This text search method is
It is roughly divided into two types, one that uses a vector in the keyword space.

The narrowing down by keywords simply uses the keywords specified by the user as it is, or includes the keywords specified by the user and the similar keywords by using thesaurus information that defines the similarity between the keywords. A method of performing selection by group is known.

On the other hand, as a method of expressing as a vector of the keyword space, after extracting the keywords in the text used in the past in advance and creating an N-dimensional keyword space using the N extracted keywords, , There is one that makes a selection by expressing the position of the text in the created keyword space and judging whether or not this is within the range of interest of the user (Peter W. Foltz and Sus.
an T. Dumais (PERSO NALIZED Information Delivery: AN
ANALYSIS of Information Filtering Methods [COMMUNI
CATIONS OF THE ACM Dec 1992 Vo135, No.12 pp51-pp6
0])).

Further, as a method of dynamically handling a keyword space, there is a method of connecting similar keywords to each other and using a keyword connection for selecting text not only from the original keyword but also from the similar keywords (Japanese Patent Laid-Open No. Hei 10-1999). JP-A-2-42564 and JP-A-2-12536.
3 gazette). In such a document search device, the accuracy of the text in the database is calculated based on the keyword input by the user using the keyword association table, and the search results are extracted by extracting the ones that are equal to or more than the threshold value.

However, when text is selected using a method of narrowing down by a keyword or expressing as a vector of a keyword space, since it depends on a mechanical search,
There is a problem that the similarity between keywords unique to the user cannot be flexibly expressed. Moreover, in a text including many keywords of low importance, the keywords are buried in noise, and it is difficult to extract them.

Further, when a new keyword is generated, the user inputs the individual keywords one by one, so that updating the thesaurus dictionary and the keyword space requires a great deal of labor and computational resources. .

In addition, with the method using the keyword connection, it is possible to search the text which is considered to be closely related to the user from the given keyword while retaining the similarity information between the keywords unique to the user. It was impossible to determine how interested you were in a particular text.

For example, in the case of judging whether a user is interested in a certain document,
Generally, users do not read the contents one by one, but rather read the whole sentence and when very important keywords appear, or when a number of even less important keywords appear. In addition, it may take a double action pattern of reading more carefully. In the case of the former behavior pattern, the conventional search method using keywords can be sufficiently dealt with, but in the latter case, the desired text may be buried in noise.

Therefore, in order to incorporate such human judgment criteria into a computer, an attempt has been made to use a cognitive model in which human behavior is modeled as an information processing device. For example, ACT ^* (The
Architecture of Cognition: Harvard University Press
1983, ISBN 0-674-04426-6) is also known as one of such cognitive models. This theory expresses the concept of human beings as one node on the network, propagates the activity on this network, and assumes that the one with high activity is the one currently remembered. Express.

However, when a text selection device is configured as such a cognitive modeled information processing device, all keyword groups for all text groups in the database are activated and evaluated on the keyword association network. Since it must be done, so-called computational explosion occurs and is unrealistic.

[0013]

As described above, the conventional text selection has a problem that it is not possible to flexibly express the similarity between the keywords peculiar to the user because it depends on the mechanical search. In addition, it is difficult to extract a text in which many keywords of low importance appear because the keywords are buried in noise. Further, when a new keyword is generated, the user successively inputs it, which requires a great deal of labor and calculation resources for updating the thesaurus dictionary and the keyword space.

Further, in the method using the keyword connection, it is possible to search the text which seems to be closely related to the user from the given keyword while retaining the similarity information between the keywords unique to the user, but it is not specified. It was impossible to determine how much interest you had in the text.

Further, when a text search device is configured as a cognitive modeled information processing device, all keyword groups for all text groups in the database must be activated by the keyword association network for evaluation. , A computational explosion occurred and was unrealistic.

Therefore, a first object of the present invention is to provide a text selection device capable of flexibly expressing the similarity between keywords unique to a user. The second purpose is
It is an object of the present invention to provide a text selection device capable of selecting a text even for a text in which many keywords of low importance appear. Furthermore, a third object is to provide a text selection device that facilitates management of similarity information between keywords. A fourth object of the present invention is to provide a text selection device which eliminates the computational explosion that is a problem when using a keyword association network and can select a text extremely rapidly and realistically.

[0017]

In order to solve the above problems, when the present inventor considers only a new text, the keyword group is not created by the keyword association network, but the entire text is created. It was noted that if the above is evaluated and judged by a threshold value, text selection can be performed in consideration of a keyword that was conventionally cut off by setting a threshold value for each keyword.

That is, the text selecting device according to the first aspect is desired by the user by using a keyword associative network provided with nodes having information about size and weight corresponding to individual keywords and links connecting the nodes. In the text selection device for selecting the text, the means for setting the active value of all nodes on the keyword associative network to the initial state,
Keyword extracting means for extracting a predetermined keyword from the target text, judging means for judging whether or not the keyword extracted by this means exists as a node on the keyword associative network, and the extracting means by this judging means. When the determined keyword is present as a node, the active value calculating means for calculating the active value of the node corresponding to the extracted keyword based on the information regarding the size, and the active value calculating means for calculating the active value of the node Activeness propagating means for propagating an activeness value on the keyword associative network, and a score for calculating the score of the target text based on the activeness values and information on weights of all nodes on the keyword associative network after activeness propagation The calculation means and the text calculated by this means DOO scores is characterized in that a presentation unit for presenting the text to the user when the predetermined threshold value or more.

It is to be noted that the activity amount calculation means calculates the activity value of a keyword extracted from the text for a partial network of a predetermined area centered on each keyword on the keyword associative network. Is significantly reduced, which is preferable. The partial network may be determined by the distance on the keyword associative network centered on each keyword,
It may be determined by the strength of the link on the keyword associative network centering on each keyword.

The presenting means updates the information about the size of the node corresponding to the extracted keyword based on the evaluation value input by this means, and the means for inputting the evaluation value for the presented text. It is preferable to include a means for doing so.

Further, it is preferable that the presenting means includes an evaluating means for tracking the movement of the line of sight when the user reads the presented text and evaluating the value of the text from the tracking information.

A text selecting apparatus according to a seventh aspect extracts text information that is considered useful to the user from the input text information, predicts the degree of importance, and displays and presents the text information determined to be important to the user. The main body of the text selection device tracks the movement of the line of sight when the user reads the displayed text information, and the degree of interest of the user with respect to the displayed text information is evaluated from the tracking information to determine the main body of the text selection device. Evaluation means for feedback,
There is provided learning means for learning the coefficient used in the evaluation by the evaluation means by using the evaluation value obtained by the evaluation means and the predicted value used in the text selection device body.

[0023]

According to the text selecting apparatus of the first aspect, first, a predetermined keyword is extracted from the newly targeted text. If the extracted keyword exists on the keyword association network, that is, if the node corresponding to each keyword exists on the keyword association network showing the connection relation between keywords, it is determined by the size of the node. The initial value is the node activation value. After obtaining the active value of the node for all the keywords in which the node exists, the active value is propagated on the keyword associative network by the method of ACT ^* described later,
The final activity value of the nodes in the entire keyword association network is calculated.

Next, the score of each node is obtained by a function considering the activity values of all the nodes on the keyword associative network and the weight representing the importance of the node, and the scores of all the nodes are summed up. , Find a text score that represents the value to an individual user.

When the score of the text thus obtained exceeds the threshold value set by the user, this text is presented to the user. The user makes a judgment on this text and gives the judgment result to the system. As a result, the size of the nodes corresponding to all the keywords extracted from the text is increased by the score of the text, and the nodes corresponding to the keywords extracted from the text have links. The keyword associative network is learned as described above.

On the other hand, when the extracted keyword does not exist as a keyword on the keyword association network, that is, when the node corresponding to the keyword does not exist on the keyword association network, after selecting the text, the keyword association is performed. Learn about networks. That is, the node is added to the keyword association network, the keyword is displayed to the user to prompt the user to input the weight, and this input value is set as the weight of the corresponding node. Similarly, increase the size of the nodes corresponding to all the keywords extracted from the text by the evaluation value of the text, and provide links between the nodes corresponding to the keywords extracted from the presented text. Learn keyword associative networks.

Therefore, according to the text selection device of the present invention, a new keyword can be flexibly taken in, and similarity information between keywords unique to the user can be obtained.

In the text selection device of the first aspect, the ACT ^* logic proposed by Anderson is used as a method of propagating the activity on the keyword association network. The method of propagating the activity will be described below. The amount of change in activity at time t at node i is

[0029]

[Equation 1] Represented by

The first term on the right side of the equation (1) is the time t of the node i.
Represents the total amount of activity acquired in
Represents the total amount of activity lost at time t. The amount of change can be obtained from the difference between the two. Thus, the total inflow of activity at node i at time t is

[0031]

[Equation 2] Represented by

The first term on the right side of the equation (2) represents the amount of activity directly supplied from the system at the time t of the node i, and the second term is the outflow of the activity of the node i to other nodes at the time t. And inflow. That is, since there are only two types of active flows, the sum of the two becomes the total active inflow amount n _i (t) at the time t of the node i.

By the way, although the expression (1) is an expression for each node, the vector A of active values of all nodes, the vector N of total inflow active of all nodes, and the amount of change in activity are expressed as follows. Like

[0034]

[Equation 3] Similarly, when the matrix R representing the strength of the link between the nodes and the matrix C representing all the active supply amount are introduced, the equation (2) becomes as follows.

[0035]

[Equation 4]

The expression (3) is an expression for obtaining the change amount of the activity, and what we desire is to know the distribution of the activity of the network in the steady state. Since the amount of change in the activity value in the steady state is 0, the left side of equation (3) is 0.
Can be regarded as It takes an infinitely long time to reach a steady state in a pure mathematical sense, but it is sufficient to approximate 0 here. Therefore, equation (5) is obtained from equations (3) and (4).

[0037]

[Equation 5] Finally, the following equation is obtained by modifying this equation (5).

[0038]

[Equation 6]

According to Anderson, ρ ^* = 1, B = 0.
It is supposed to be 8. Therefore, the final value of each node activity on the keyword association network is

[0040]

[Equation 7] Can be approximated by

According to the text selection device of claim 7,
Since the movement of the line of sight when the user reads the displayed text information is tracked and the degree of interest of the user with respect to the displayed text information is evaluated from this tracking information, the degree of interest of the user is evaluated. It becomes possible for the user to learn the text selection device without inputting explicitly.

[0042]

Embodiments will be described below with reference to the drawings. FIG. 1 is a block diagram of a text selection device according to an embodiment of the present invention.

This apparatus is divided into a part for selecting a text and a part for performing learning for reflecting the selected text on the keyword association network.

In the figure, the document storage unit 1 is for temporarily storing a new target text. The keyword extraction unit 2 extracts a predetermined keyword from the text stored in the document storage unit 1.

The extracted keywords are stored in the keyword storage unit 3. With respect to the keyword stored in the keyword storage unit 3, the active value calculation unit 4 obtains the active value of the keyword. Then, the obtained activity value is propagated on the keyword association network stored in the keyword association network storage unit (hereinafter referred to as “network storage unit”) 5. The score of the text is calculated based on this propagation result. The threshold determination unit 6 determines whether the calculated score of the text is equal to or higher than a predetermined threshold, and if the score is equal to or higher than the threshold, the threshold determination unit 6 sends the text to the display unit 7 for display. Further, the display unit 7 appropriately displays data, commands, etc. input from the input unit 8. The keyword association network updating unit 9 updates the keyword association network based on an evaluation value, which will be described later, input by the input unit 8.

Next, each section of the text selection device according to this embodiment will be described in detail. The new text temporarily stored in the document storage unit 1 is a known keyword extraction means (for example, Tetsuya Ishikawa: “Subject index creation support system based on textual analysis processing”, IPSJ Journal Vo132, No. twenty one
Keywords are extracted by the keyword extracting means used in 991). The extracted keywords are stored in the keyword storage unit 3.

The active value calculation unit 4 first gives initial values to all nodes stored in the network storage unit 5 and corresponding to each keyword on the network. Then, when the processing for all the keywords in which the corresponding node exists is completed, the active value calculation unit 4 propagates the active value on the network to obtain the final active value. Furthermore, the score of each node is obtained using a function that takes into account the activation value and weight of each node. In this embodiment, the score of a node is the product of the active value of the node and the weight.

Here, the keyword association network used in this embodiment will be described. FIG. 2 is a schematic diagram showing an example of a network.

In the node 21 corresponding to each keyword,
A “size” that represents the cumulative total of the user's scores and a “weight” that represents the degree of importance to the user are provided. There are cases where links exist between the nodes and cases where links do not exist. When there is a link, the "link strength" from the node A to the node B is the value obtained by dividing the "size" of the node B by the sum of the "sizes" of all the nodes to which the node A has a direct link. Determined by That is, the sum of the strengths of all links from a certain node A is always 1.

FIG. 3 is a diagram showing an example of a state in which the node information on the network shown in FIG. 2 is stored in the network storage unit 5. The attributes of the node include "node name", "size", "weight", "active value" and "score". In this embodiment, the active values are all 0 in the initial state.

FIG. 4 is a diagram showing an example of a state in which the link information between the nodes on the network shown in FIG. 2 is stored in the network storage unit 5. The matrix index (i / j) corresponds to the node number in FIG. When there is a link between nodes, the corresponding element of the matrix is represented by 1, and when there is no link, it is represented by 0. The node has no link to itself and will always be 0. Also, if there is a link from node i to node j,
At the same time, it is assumed that there is a link from node j to node i. Therefore, the matrix representing the inter-node link information is always a symmetric matrix. The link information does not necessarily have to be bidirectional, and may be a triangular matrix that describes only one direction.

FIG. 5 is a diagram showing an example of a state in which information about the “link strength” between the nodes is stored in the network storage unit 5. The strength of the link between the nodes represents the strength of the link from the node i to the node j, and is obtained from the node information and the link information between the nodes. For example, the strength of the link between the node “user” and the node “interface”, that is, the node number i = 1
From the node number j = 2 to the link number = 90 / (80 + 90) = 0.53.

Next, the activity value calculation unit 4 will be described.
The "active value" attribute of the node information stores the active values of all the nodes on the network. In this embodiment, all initial states are 0. The activity value calculation unit 4 reads the keyword group from the keyword storage unit 3, and when the node corresponding to each keyword exists on the keyword associative network, the activity value of the node is set as the “size” of the node. Set to the same value. When the processing is completed for all the keywords, activity propagation is calculated.
The activity of a certain node is distributed by the value determined by the attenuation rate ρ to the adjacent node having the link according to the strength of the link. The function of the active value calculation unit 4 will be described with reference to FIG.

FIG. 6 is a flow chart for explaining the function of the active value calculation unit 4. It is assumed that the keyword storage unit 3 stores the extracted keywords as shown in FIG. 7 as a keyword list. The active value calculation unit 4 sets the index i for referring to this keyword list to 1 which is the head of the keyword list (step 6).
1).

Next, it is searched whether or not the i-th keyword is present in the keyword list, that is, whether or not there is a keyword that has not been read in the keyword list (step 62). Is read (step 63). When the read keyword is stored in the network storage unit 5 (step 64), the active value of the corresponding node is set to the same value as the size of the node (step 65). Then, the index is advanced by one (step 66), and the series of processes is performed for all the keywords in the keyword list.

When the processing for all the keywords in the keyword list is completed (step 62), the active value calculation unit 4 obtains the propagation of the active value (step 67). The final liveness value of each node on the network is approximated by Eq. (7), as described above. R in the equation (7) is the “link strength” between the nodes shown in FIG. 5, and C ^* is the result of activation of the node corresponding to the keyword extracted from the text by the system, that is, is supplied to the network. Represents the amount of activity value. The amount of this active value has already been assigned to the “active value” attribute in step 65. Therefore, if equation (7) is calculated and sequentially written in the “active value” attribute shown in FIG. 3, it is equivalent to obtaining the propagation of the active value of the network. In this example, (7)
SOR (Successive O
ver-Relaxation) method is used, but it is not limited to this.

Next, the active value is normalized in order to avoid overflow (inflation) of the active value due to the enlargement of the nodes accompanying the learning of the keyword association network. Specifically, the sum of the activity of all nodes is calculated, this is divided by the value of the activity of each node, and this is stored again in the term of the activity value of FIG. 5 (step 68).

Next, the activity value calculation unit 4 obtains the score of each node by a function considering the activity value and the weight. Here, it is assumed that the score is obtained by the product of the active value and the weight. The obtained score of the node is written in the "score" attribute of the node in FIG. 5 (step 69). The text score is then calculated by summing the scores of all nodes (step 70).

For example, when the network shown in FIG. 2 and the keyword list shown in FIG. 7A are given to the active value calculation unit 4, the active value is propagated as shown in FIG. The score is 0.15 × 4 ＋ 0.23 × 4 ＋ 0.17 × 0 ＋ 0.14 × 3 ＋ 0.16 × 2 ＋ 0.15 × 1 ＝ 2.41.

On the other hand, when the same network and the keyword list of FIG. 7 (b) are given to the active value calculation unit 4, the active value is propagated as shown in FIG. 8 (b), and the score of the text is 0.13 × 4 + 0.20 x 4 + 0.30 x 0 + 0.12 x 3 + 0.14 x 2 + 0.10 x 1 = 2.06.

Next, the threshold value judging section 6 will be described. The threshold determination unit 6 sends a control command to the display unit 7 to display the text when the score of the text is equal to or higher than the threshold. Therefore, the text is displayed on the display unit 7.
If the score of the text is less than or equal to the threshold, no display is made.

In this way, the display unit 7 displays the text stored in the document storage unit 1 according to the control command of the threshold value determination unit 6. Therefore, in the above example, when the threshold value held in the threshold value judgment unit 6 is between 2.40 and 2.06, for example, 2.20, the text that generated the keyword list of FIG. Is displayed to the user. On the other hand, the text that generated the keyword list in FIG. 7B will not be displayed to the user.

That is, focusing on the network,
It turns out that the user is interested in the keyword "interface". However, this means a “user interface” and not a “communication interface” between hardware, so that it matches the original intention of the user.

In the above example, both are "interfaces".
However, even if the node size and the node weight are small, a plurality of nodes are included even if the node size is large and the node weight is small (0 in this example). When the node supports it, the text score is higher.

Therefore, even when the keyword list extracted from the text includes a large number of nodes having small sizes and weights, it is possible to obtain a large text score. Moreover, in such a case, the user often does not simply know a keyword having a large weight. Therefore, according to such a method, it is possible to discover a new keyword. On the contrary, when only one keyword having a very large weight is included in the keyword group, it can be achieved by the ordinary keyword search means, and therefore it may be used together with this.

With respect to the text displayed on the display unit 7 in this way, the user judges the text and inputs the judgment result for the user from the input unit 8 as an evaluation value. In the present embodiment, a five-level evaluation value of a number from 0 to 4 is used, and the higher the number, the more important the text to the user. Further, in the present embodiment, the extracted keywords that are not registered on the network are sequentially displayed on the display unit 7, and the weights of these keywords are also evaluated by the user from the input unit 8 by the five-level evaluation. I am trying to encourage you to type by.

The determination result input from the input unit 8 is sent to the keyword association network updating unit 9. The keyword associative network update unit 9 updates the network stored in the network storage unit 5 based on the keyword stored in the keyword storage unit 3 and the input determination result.

Next, the function of the keyword association network updating unit 9 will be described. The keyword association network updating unit 9 updates the state of the network when the text displayed on the display unit 7 receives the determination result from the input unit 8.

9 to 11 are flowcharts for explaining the function of the keyword association network updating unit 9.
In the process shown in FIG. 9, registration is performed for a keyword that is not registered on the network, and the size of the node corresponding to the keyword newly extracted by the keyword extracting unit 2 this time is changed.

That is, the index i for referring to the keyword list stored in the keyword storage unit 5 is set to 1 which is the head of the keyword list (step 91). Next, it is confirmed whether or not the i-th keyword is present in the keyword list (step 92), and if it is present, the corresponding keyword is read (step 9).
3). Next, when the read keyword is not stored in the node information area of the network storage unit 5 (step 94), the corresponding node is added to the node information and the keyword is registered in the "node name" attribute. , "Size", "Weight", "Activity" and "Score" of node
0 is assigned to the attribute (step 95), and the keyword is displayed on the display unit 7 (step 96).

The user evaluates the weight of the keyword in five levels and inputs the determination result from the input unit 8 (step 97). The determination result is represented by a numerical value of 0 to 4, and the larger the numerical value, the more important the user is.

Next, the judgment result is substituted into the "weight" attribute of the corresponding node of the node information (step 98). Next, the user writes the judgment result of the text evaluated by the user in the "size" attribute of the corresponding node (step 99). Then, the keyword index i is advanced to the next (step 100). This series of processes is performed for all the keywords in the keyword list (step 92), and the process proceeds to the next process shown in FIG.

In the process shown in FIG. 10, all the keywords newly extracted this time are provided with links. That is,
The i-th keyword in the keyword list is read (step 102), the node number of the read keyword is retrieved from the node information area of the keyword storage unit 5,
The variable a is substituted (step 103). Next, let j be the index of the keyword of the other party to which the link is to be made, this index j points to the next of the index i (step 104), and it is confirmed whether the jth keyword is present in the keyword list, and it is present. In this case, the jth keyword in the keyword list is read (step 105). The node number of the read keyword is extracted from the node information area and assigned to the variable b (step 106).

Next, the elements (a, b) and (b, a) of the link information between the nodes represented by the matrix in the link information area of the keyword storage unit 5 are set to 1 (step 107), The index j of the keyword list is advanced (step 108). When the process for j is completed for all the keywords in the keyword list (step 105), the keyword index i is advanced to the next (step 109). In this way
A series of processes is performed for all the keywords in the keyword list (step 102), and the process proceeds to the next process shown in FIG.

In the processing shown in FIG. 11, the strength between the links shown in FIG. 5 is changed in accordance with the change in the links between the nodes and the size of the nodes. That is, the index i for referring to the node information is set to 1 at the beginning of the node information.
(Step 111). Next, i
Check if the th node exists (step 11)
2). Since the strength of the link differs depending on its direction, the node i here represents the starting point of the link.
Next, the index j for referring to the node information is set to 1 which is the head of the node information, and the variable "total" for obtaining the total size of the nodes is set to 0 for initialization (step 113).

Next, it is confirmed whether or not the jth node exists in the node information (step 114). Where node j
Represents the destination of the link. In the link information area of the keyword storage unit 5, each element of the matrix representing the presence or absence of a link is represented by "link [i, j]", and the link [i,
By adding [j] × size [j] to the variable “total”, the size of the connected node can be added if a link exists (step 115).
Next, the index j for the node information is advanced (step 116). When all the nodes for the index j have been processed (step 114), the index j is returned to 1 which is the head of the node information again (step 117). At this point, the total size of all the nodes connected to the node i is obtained.

Next, it is confirmed whether or not the i-th node exists in the node information (step 118), and if the node exists, the element "strength [i [i , J] ”is substituted with the value obtained by dividing the sum of the values of the size [j]. However, by multiplying the link [i, j], the strength is set to 0 when the link does not exist (step 119).

Next, the index j of the node information is advanced (step 120). When all the nodes of the node information have been processed for j (step 118), the index i of the node information is advanced (step 12).
1). When all the nodes in the node information have been processed for i (step 112), the process ends.

FIG. 12 shows the update result obtained by adding a new keyword “man-machine” having a score of 4 and a weight of 4 to the keyword associative network shown in FIG. 2 for generating the keyword group of FIG. 7 (a). Shown in. Newly added links are shown by solid lines, and existing links are shown by dotted lines.

By the way, in the above example, in the active value calculation unit 4,
The activity value is propagated using the ACT ^* method, and the final activity value on the keyword association network is obtained. If the method of ACT ^* is applied as it is, it is necessary to solve simultaneous linear equations of the same dimension as the number of nodes existing on the keyword association network. Solving simultaneous linear equations increases the amount of calculation in proportion to the cube of the dimension of the equations, which makes it difficult to use for a large-scale keyword association network.

Therefore, here, two types of active propagation methods will be described which can obtain a steady state of active propagation without causing a computational explosion even in a large-scale keyword associative network.

(1) Method Using Simple Distance In this method, each keyword extracted from the text can be reached within a predetermined number of steps by following the link of the keyword association network centering on each keyword. The activity value is calculated using the partial network and the ACT ^* formula, which is obtained by extracting only the node and the link. Then, the activation value of each node of the obtained partial network is added to the corresponding node of the keyword associative network, and this is performed for all the keywords to obtain the activation values of all the nodes of the keyword associative network.

Hereinafter, description will be made with reference to the flow of processing shown in FIG. First, the index i for referring to the keyword list as shown in FIG. 7 stored in the keyword storage device 3 of FIG. 1 is set to 1 which is the head of the keyword list (step 131). Next, it is confirmed whether or not the i-th keyword is present in the keyword list (step 132).
The th keyword is read (step 133). If the read keyword is stored in the keyword association network as shown in FIG. 3 (step 134),
The variable "starting point" is set to the read keyword, and the variable "distance" is set to the predetermined size of the partial network (5 in this example) (step 135).

The partial network shown in FIG. 17 is initialized (step 136). However, no node is registered in the partial network at the initialization stage. Arguments to the partial network generator (starting point, distance)
Is passed, the link is traced around the starting point, and a partial network including up to five adjacent nodes is created (step 137).

The active value of the starting point in the obtained partial network is used as the keyword association network (entire network).
Set to the size of the node at the starting point in (Step 13
8). Link strength information between nodes included in the partial network is extracted from the link information as shown in FIG. 5 (step 139), and propagation of the activity is obtained (step 14).
0). The liveness value of each node of the partial network is added to the liveness value of each node of the corresponding whole network (step 141). Advance index i by 1 (step 14
2) The process proceeds to the next keyword. When the processing of all the keywords is completed (step 132), the processing is completed.

Next, the operation of the partial network generator will be described with reference to FIG. The partial network generation unit recursively expands the partial network while the argument (distance) is positive. First, the number in the whole network of the starting point received as an argument is registered in the partial network, and its active value is set to 0 (step 141). Next, the distance received as an argument is decremented by 1, and the value is stored in the variable "next distance" (step 142). If the next distance is 0, the system terminates and returns to the caller (step 143). If the next distance is not 0, the variable i serving as an index is initialized to 0 (step 14).
4). By referring to the link information as shown in FIG. 4, the node adjacent to the origin node and the adjacent node [i] are searched, and when all the adjacent nodes have been searched, the process ends (step 145).

If the adjacent node [i] is not registered in the partial network (step 146), the partial network generator is recursively called with the adjacent node [i] as an argument (step 48). If the adjacent node [i] is registered in the partial network (step 146), the distance between the adjacent node [i] and the next distance are compared. If the next distance is larger (step 147), the adjacent node The partial network generation unit is recursively called with [i] and the next distance as arguments (step 148). The index for searching for the adjacent node is advanced by one, and the next adjacent node is searched for (step 149). If the distance of the adjacent node [i] is equal to or greater than the next distance, the index for searching for the adjacent node is increased by one, and the next index is searched for. Search for an adjacent node (step 149).

In this way, a partial network is created by tracing all the links and extracting all the nodes existing in the reachable range by the number of times defined by the initial distance.
Next, a method of creating information on the strength of links between nodes corresponding to a partial network will be described with reference to FIG.

First, the index i is initialized to 1 (step 151). When the index i (the number of nodes included in the partial network) is N or less, the process is performed on the node [i], and when the process reaches N, that is, when the process is completed for all the nodes included in the partial network, the process ends. (Step 152).

The index j is initialized to 1 (step 153). When the index j (the number of nodes included in the partial network) is N or less, the process is performed on the node [j] (step 154), and N is reached, that is, the process is completed for all the nodes included in the partial network. In this case, the index i is incremented by 1, and the process is performed again for all j (step 155).

The variable a stores the number of the entire network of node [i], and the variable b stores the number of the entire network of node [i] (step 156). Strength from node [i] to node [j]: Since the strength [i, j] of the sub-network is the strength [a, b], this value is stored (step 157). Index 1
Next (step 158), the process continues for the next node.

When the processing is performed for all the nodes in this way, the extraction of the necessary link information is completed. (2) Method that uses distance considering link strength With this method, for each keyword extracted from text, centering on each keyword, the link of the keyword association network is used by the marker passing method, and the link strength is determined. The value of the marker zorch is decreased according to, and only the nodes and links up to the range reached by the zorch are finally extracted, the partial network and ACT ^*
The activity value is calculated using the equation and. The obtained activation value of each node of the partial network is added to the corresponding node of the keyword association network. By performing this for all the keywords, the active values of all the nodes in the keyword associative network are obtained.

That is, in the above example (1), the distance is reduced by one each time the adjacent node advances, but since the propagation of the activity depends on the strength of the link, in this method the strength of the link is The distance is calculated using a function that takes into consideration. This can be realized by replacing the partial network creation unit in the example (1) described above.

FIG. 16 shows an example of the creating unit of the partial network.
Will be explained. The process is performed recursively as in the partial network creation unit in the above example (1). The starting point given as an argument is registered in the partial network as shown in FIG. 17 (step 161). The index i is initialized to 0 (step 162). If there is a node adjacent to the unprocessed starting point, the process is performed (step 163).

The distance between the adjacent node and the starting point is obtained using the distance function. In this example, as an example, the reciprocal of the strength of the link from the origin node to the adjacent node is used as the distance. That is, 1 / strength [starting point, adjacent node] is the distance from the starting point to the adjacent node. The value obtained by subtracting the value of the distance from the starting point to the adjacent node from the variable "distance" is stored as the variable in the next distance (step 164). If the next distance is not negative, the process is performed (step 165). If the adjacent node is already registered as a partial network (step 166), the partial network creating unit is recursively called only when the next distance is larger than the distance when it was registered (step 167). Is updated (step 168). If the adjacent node is not registered (step 166), the partial network creation unit is recursively called to perform registration (step 168). Advance index i by 1 (step 1
69), and process the next adjacent node.

In addition to the above method, rather than having the active value information as a part of the entire network,
The activity value information table as shown in FIG. 18 is independently provided, and only the activity values of the nodes whose activity values have changed are stored, so that the activity remains in the initial state (0) when the score of the document is calculated. It is not necessary to consider the node of, and it becomes possible to speed up the score calculation of the document.

That is, by registering the active value information of each node obtained in the partial network in the active node table as shown in FIG. 18, it is possible to create a list of only nodes having a positive active value.

That is, in step 140 of FIG. 13, when the partial network is added to the entire network, the node number and the active value are registered in the active node table as shown in FIG. When the node is already registered, the activation value newly obtained in the partial network is added.
When obtaining the score of the document, the active node table is referred to, the node numbers are obtained in order from the head node, and the weights of the corresponding nodes are obtained from the entire network using the node numbers. Next, the activation value is obtained, and the score of the node is obtained from these two values. By performing the above process for all nodes registered in the active node table,
You can get the score of the document.

In the embodiment shown in FIG. 1, the display unit 7
A method is adopted in which the user himself / herself evaluates the selected result while reading the text displayed on the screen in five levels, and the user directly inputs the evaluation value via the input unit 8. However, with such an evaluation input method, it is difficult to satisfy promptness and accuracy. Therefore, the emergence of a method for easily and accurately evaluating the selection result is desired.

The method for evaluating the presented text is as follows:
A method of determining that the user is interested in information on which an operation such as a search is performed using a history of search and operation, as seen in the personal electronic newspaper system disclosed in JP-A-4-192775, A method of calculating the interest of the user from the presentation time per information unit based on the presentation amount and presentation time of information, as seen in the information providing method of the information providing system disclosed in Japanese Patent Laid-Open No. 4-77866. Are known.

However, in the method using the history of search and operation, the degree of interest of the user can be obtained only indirectly, and although he is interested in information, he is satisfied without performing further search or operation. It is impossible to distinguish information that has been lost from information that was not of interest. In addition, in the method of calculating the user's interest from the presentation time per information unit based on the presentation amount and presentation time of information, if the information is left in the presented state, it is thoughtful. In such a case, it may be determined that there was a serious interest even if the interest was not.

Therefore, here, it is presented by tracking the viewpoint of the user who sees the text selected and displayed by the text selection device shown in FIG. 1, that is, the user, and analyzing the movement characteristics thereof. An evaluation device for quickly and accurately evaluating a text without the user explicitly inputting the degree of interest the user has will be described.

That is, this evaluation device is a line-of-sight tracking device which is determined to be important by the text selection device shown in FIG. 1 and tracks the movement of the line of sight when the user reads the displayed text. A coordinate conversion device that converts the coordinate information obtained by the eye gaze tracking device into position information in the text, an evaluation value calculation device that evaluates the degree of interest of the user using the converted position information, and is used at the time of evaluation An evaluation value storage device that stores a coefficient, a past prediction value, and an evaluation value, and a learning device that learns the coefficient by using the stored prediction value and the evaluation value are provided.

In this evaluation device, when the user views the displayed text, the Mo of Applied Science Laboratories is used.
A line-of-sight (viewpoint) tracking device such as del 3250R is used to constantly record the position of the viewpoint on the display device. Human beings have different ways of carrying the viewpoint on the display device depending on their interests. That is, when displaying text information, if the information you are interested in,
There is little disturbance in the movement of the viewpoint, and the characters are followed sequentially from the beginning.
On the other hand, if the information is not of interest, skip it while reading,
Disturbances such as stopping on the way or moving the viewpoint outside the display device occur. This evaluation device detects these disturbances and evaluates the degree of interest in the presented text based on the degree of the disturbances.

The measured degree of interest is passed to the input unit 8 of the text selection device shown in FIG. 1 and managed and used as user information. The degree of interest measured by this evaluation device is not limited to the text selection device shown in FIG.
Another type of text selection device, for example, Japanese Patent Laid-Open No. 2-1
It can also be used for the keyword connection as seen in the document search device of Japanese Patent No. 25363. In this document search device, the strength of connection between keywords is expressed by a network, a document search is performed based on the strength of connection from the search condition, the user judges / inputs the search result, and the judgment result is used. Then, the connection between keywords is learned, and the behavior is changed to be closer to the user's intention from the next search.

The carrying of viewpoints gradually changes due to physical deterioration, learning, and training. Therefore, a mechanism for gradually changing and learning the evaluation value corresponding to the movement of the viewpoint is required.
The text selection device shown in FIG. 1 is used not only for presenting information but also for filing, information retrieval, kana-kanji conversion dictionary, etc., and it is also possible to learn with different feedback information. If the prediction prediction function and the actual evaluation value are continuously different from each other, it is determined that the evaluation device has deviated from reality and the coefficient used for evaluation of the evaluation device is adjusted to the reality. Need to learn.

FIG. 19 shows an example of the evaluation device. Here, the evaluation value obtained from the evaluation device is passed to the text selection device via the input unit 8 of FIG. Then, in FIG. 19, the text selection device shown in FIG.
It is indicated by 2.

The text information input from the document storage section 31 is judged by the text selection device 32 to what degree the text information is important to the user, and only those which are judged to be important are scored. It is stored in the device 33.

Based on the score added by the text selection device 32, the control device 34 takes out text information that is considered useful to the user from the information recording device 33, and displays it on the display device 36 in response to a request from the input device 35. Display textual information.

The line-of-sight tracking device 37 tracks the movement of the line of sight of the user who reads the text information displayed on the display device 36, and uses the position of the user's viewpoint on the display device 36 as coordinate information in the coordinate conversion device 38. hand over.

The coordinate conversion device 38 acquires the text information displayed on the display device 36 from the information storage device 33 and the control device 34, and obtains it from the eye gaze tracking device 37.
The position of the user's viewpoint on 6 is converted into the coordinates of the text information.

When the next document display command or the end command is input from the input device 35, the converted coordinate information is passed to the evaluation value calculation device 39 together with the text information. The evaluation value calculation device 39 reads the coefficient used for evaluation from the evaluation value storage device 40, determines the degree of interest of the user from the information, and passes the calculated evaluation value to the text selection device 32 as feedback information. The text selection device 32 performs learning based on this feedback information.

The learning device 41 uses the score added by the text selection device 32 from the information storage device 33 and the evaluation value calculation device 3
The result of the evaluation performed by 9 is received and stored in the evaluation value storage device 40. If the values are significantly different, the evaluation value calculation device 39 is learned, and the coefficient after learning is stored in the evaluation value storage device 40. .

Next, referring to FIG. 20, the control device 34
Explain the function of. The document information and the score added by the text selection device 32 are read from the information storage device 33, and the number of the read document is notified to the coordinate conversion device 38. (Step 211). A display start line which is a variable is initialized to 1 (step 212). Document information for N lines from the display start line is displayed on the display device 36 (step 213). However, it is assumed that N is the number of lines that can be displayed on the display device 36 at one time.

The display start line is notified to the coordinate conversion device 38 (step 214) and the input of an instruction from the input device 35 is waited for (step 215). If there is an input, if the input is a next page display command (step 216), the display start line is advanced by N and the process returns to step 213 (step 21).
7). If the input is a next document display command (step 21)
8) The evaluation start is notified to the coordinate conversion device 38 (step 219), and the process returns to step 211. If the input is an end command (step 220), the start of evaluation is notified to the coordinate conversion 38 (step 221), and the process ends.

Next, the operation of the coordinate conversion device 38 will be described with reference to FIG. The document number of the currently displayed document is received from the control device 34 (step 231). Based on the received document number, the relevant document is read from the information storage device 33 (step 232). A display start line is received from the control device 34 (step 233).

When there is no command from the control device 34 (step 234), the coordinate information of the viewpoint of the user on the display device 36 is received from the eye gaze tracking device 37 (step 23).
5) Based on the display start line information, it is determined which position in the document is being read, the coordinates are converted into coordinates on the document in the format shown in FIG. 24 (step 236), and the process returns to step 234.

If there is a command from the control unit 34 (step 234), and if the command is a command to change the display start line (step 237), the display start line is changed (step 238) and the process returns to step 234. . If the command is not a display start line change command (step 237), and if the command is not an evaluation start command (step 23).
9) and returns to step 234. If the command is an evaluation start command (step 239), the evaluation value calculation device 3
8 is started (step 240) and the process ends.

Next, an example of coordinate conversion will be described using an actual example. It is assumed that a part of the document as shown in FIG. 22 is displayed on the display device 36. Here, since the display is started from the 6th line, the display start line is 6. The number of rows (N) that can be displayed is 12. From the line-of-sight tracking device 37 to the display device 3
The coordinates with the upper left of (6, 0) as (0, 0) are input to the coordinate conversion device 38 at regular intervals. Here, the display device 36
The coordinates with the upper left corner of (0,0) and the lower right corner of (100,100) are passed, and are ignored when the user looks at a place other than the display device 36.

Now, it is assumed that there is an input from the line-of-sight tracking device 37 as shown in FIG. In this case, the position of the viewpoint on the display device 36 is the position indicated by "x" in FIG. 25, and it is considered that the viewpoint has moved as shown by the solid line in FIG.

The coordinate conversion device 38 converts such coordinate information into document position information. In the above example, the coordinates (4,
2) is converted into [6,1]. However, in [line, number of characters], [6, 1] means the first character in the sixth line. Similarly, all the coordinates are converted and passed to the evaluation value calculation device 39 as shown in FIG.

The evaluation value calculation device 39 uses the coordinate information after conversion to evaluate how much the user is interested in the display document. There are various possible methods for this evaluation. For example, if the user scans too quickly,
It judges that it does not understand the content, calculates the probability that the speed of processing characters is within a certain speed, and uses this as the evaluation value, and based on how thoroughly the area of the document is seen Using a ratio as an evaluation value, a number of times the viewpoint moves outside the area of the display device 36 while reading the document, a method of evaluating that the interest is less when the number of times is larger, and a movement of the viewpoint is a character. There are many methods that can be considered by using the number of times that goes backwards and evaluating that it is less interesting if the number of times increases, and by combining these. It is also possible to use a neural network for these evaluations.

Here, as an example, the ratio is used as the evaluation value based on how much the document area is thoroughly viewed. FIG. 27 shows the flow of processing for calculating the evaluation value from the above viewpoint.

First, the current document is read from the information storage device 33 (step 331) and the total number of characters is counted (step 332). The number of processed characters, which is a variable for the user to count the number of characters read by the user, is initialized to 0 (step 333), and the previous coordinate, which is a variable for storing the previous coordinate, is set. It is initialized to [0,0] (step 334).

If there are still input coordinates (step 3
35), one next coordinate is read (step 336).
If the read coordinates are on the same line as the previous coordinates (step 337), it is considered that the characters between these two coordinates have been read, and the number of characters between them is added to the number of processed characters (step 338).

The read coordinates are substituted for the previous coordinates which are variables (step 339) and the process returns to step 335. If the next coordinate does not exist in step 335, the coefficient is read from the evaluation value storage device 37 (step 340) and the evaluation value is obtained (step 341). The obtained evaluation value is output to the learning device 38 and the text selection device 32 (step 342), and the process ends.

For example, in a document in which the total number of characters is 100,
When the line of sight moves as shown by the solid line in 8, the number of processed characters is 90, and the evaluation value is (90/100) ×
It becomes a coefficient. For example, if the coefficient is 5, the evaluation value is 4.5. On the other hand, when the line of sight moves in the same document as shown by the solid line in FIG. 29, the number of processed characters is 50.
And the evaluation value is (50/100) × coefficient. If the coefficient is 5, the evaluation value will be 2.5 and the user will be shown in FIG.
It is evaluated that he was not interested in the case.

The learning device 41 holds the predicted value from the user model 32 and the evaluation value from the evaluation value calculation device 39 a predetermined number of times (H times), and the difference between these values is continuously set to a constant threshold value. If it exceeds, the coefficient is learned and the coefficient is set so that the difference becomes half at each time. That is, a coefficient is newly selected such that the difference between the predicted value and the evaluation value is half.

This is obtained by new coefficient = {current coefficient (predicted value-evaluation value) / 2 evaluation value} + current coefficient.
This is obtained for all sets of the predicted value and the evaluation value H times, and the average thereof is used as a new coefficient. For example, assuming that the coefficient is 5, H is 3, and the threshold value is 2, • a combination of predicted value 5, evaluation value 2 • predicted value 4, evaluation value 1 • predicted value 4, evaluation value 2 is continuously recorded , Applying the above equation and taking the average, the new coefficient is 18.75,
The same input is modified to give a larger rating. The evaluation device described here can be used not only for text information but also for evaluation of figures, pictures with bitmaps, and the like.

[0130]

According to the present invention, when selecting only a useful text for a user from a huge amount of text,
The keywords are not so important in the conventional text selection device, but when many appear, they can be extracted as the text required by the user. That is, it is possible to flexibly express the similarity between keywords that are different for each user, and select only a desired text.

[Brief description of drawings]

FIG. 1 is a configuration diagram of a text selection device according to an embodiment of the present invention.

FIG. 2 is a schematic diagram showing an example of a keyword association network stored in a keyword association network storage unit of the device.

FIG. 3 is a diagram showing an example of node information stored in a keyword association network storage unit of the device.

FIG. 4 is a diagram showing an example of link information between nodes stored in a keyword association network storage unit of the device.

FIG. 5 is a diagram showing an example of information regarding the strength of links between nodes stored in a keyword association network storage unit of the device.

FIG. 6 is a flowchart for explaining the function of the activity value calculation unit of the device.

FIG. 7 is a diagram showing an example of keywords stored in a keyword storage unit of the apparatus.

FIG. 8 is a schematic diagram showing an example of an activated keyword association network.

FIG. 9 is a flowchart for explaining the function of a network updating unit of the same device.

FIG. 10 is a flowchart for explaining the function of the network updating unit of the same device.

FIG. 11 is a flowchart for explaining the function of the network update unit of the same device.

FIG. 12 is a schematic diagram showing an example of a keyword association network updated by a network updating unit of the same device.

FIG. 13 is a flowchart of processing in a position example in which the activity value is calculated only in a partial network.

FIG. 14 is a flowchart of processing when creating the partial network.

FIG. 15 is a flowchart of processing when creating strength information of a link connecting nodes in the same partial network.

FIG. 16 is a flowchart of processing when creating another partial network.

FIG. 17 is a diagram showing an example of registration of a partial network.

FIG. 18 is a diagram showing an example of an active node table corresponding to a partial network.

FIG. 19 is a configuration diagram showing an example of an evaluation device.

FIG. 20 is a flowchart showing the operation of the control device in the evaluation device.

FIG. 21 is a flowchart showing the operation of the coordinate conversion device in the evaluation device.

FIG. 22 is a diagram showing a display example on a display device.

FIG. 23 is a diagram showing an example of input coordinates to the coordinate conversion device.

FIG. 24 is a diagram showing an example of change output in the coordinate conversion device.

FIG. 25 is a diagram for explaining a viewpoint position on a display device.

FIG. 26 is a diagram for explaining movement of a viewpoint position on a display device.

FIG. 27 is a flowchart of processing for obtaining an evaluation value.

FIG. 28 is a diagram for explaining the movement of the line of sight when the user is interested in a document.

FIG. 29 is a diagram for explaining the movement of the line of sight when the user is not interested in the document.

[Explanation of symbols]

1 ... Document storage unit 2 ... Keyword extraction unit 3 ... Keyword storage unit 4 ... Activity value calculation unit 5 ... Keyword association network storage unit 6 ... Threshold determination unit 7 ... Display unit 8 ... Input unit 9 ... Keyword association network update unit 21 ... Node 31 ... Information input device 32 ... Text selection device 33 ... Information storage device 34 ... Control device 35 ... Input device 36 ... Display device 37 ... Eye tracking device 38 ... Coordinate conversion device 39 ... Evaluation value calculation device 40 ... Evaluation value storage device 41 ... Learning device

Claims (7)

[Claims] 1. Text selection for selecting a text desired by a user using a keyword associative network provided with nodes having information regarding size and weight corresponding to individual keywords and links connecting the nodes. In the apparatus, means for setting active values of all nodes on the keyword associative network to an initial state, keyword extracting means for extracting a predetermined keyword from the target text, and keywords extracted by this means are associated with the keyword associative network. Determining means for determining whether or not the node exists on the network, and when the determining means determines that the extracted keyword exists as a node, the extracted keyword is added to the extracted keyword based on the information on the size. Liveness value of the corresponding node Active value calculating means for calculating the active value of the node calculated by this means, and active propagation means for propagating the active value of the node on the keyword associative network, and active values and weights of all nodes on the keyword associative network after active propagation. And a presenting means for presenting the text to the user when the score of the text calculated by this means is greater than or equal to a predetermined threshold value. A text selection device characterized by the following. 2. The activity value calculating means obtains an activity value for a keyword extracted from the text for a partial network of a predetermined area centered on each keyword on the keyword associative network. The text selection device according to claim 1. 3. The partial network is defined by a distance on the keyword associative network centered on each keyword.
The text selection device described in. 4. The text selection device according to claim 2, wherein the partial network is determined by the strength of a link on the keyword associative network centered on each keyword. 5. The presenting means updates the information about the size of the node corresponding to the extracted keyword based on the evaluation value input by this means and the evaluation value input by the presenting means. 2. The text selection device according to claim 1, further comprising: 6. The presenting means further comprises an evaluating means for tracking the movement of the line of sight when the user reads the presented text, and for evaluating the value of the text from the tracking information. Item 5. The text selection device according to item 5. 7. A text selection device main body for extracting text information considered to be useful for a user from the input text information, predicting the degree of importance, and displaying and presenting the text information determined to be important to the user. An evaluator that tracks the movement of the line of sight when the user reads the displayed text information, evaluates the degree of interest of the user with respect to the displayed text information from the tracking information, and feeds it back to the main body of the text selection device; It is characterized by comprising learning means for learning the coefficient used in the evaluation by the evaluation means by using the evaluation value obtained by the evaluation means and the predicted value used in the text selection device body. Text selection device.