JP2000298662A - Automatic information detection device - Google Patents

Abstract

(57) [Summary] [PROBLEMS] A system in which even if input information increases, the amount of memory and the amount of calculation can be reduced compared with the conventional technology, that is, conflicting propositions of improving information extraction accuracy and learning efficiency. Provided is a system that solves the problem. A database 10 having a plurality of categorized information sources and category information, and a history of data selected by the user among the data stored in the database 10 are provided. An automatic information detection unit that learns the content of interest; and an information presentation unit that provides information corresponding to data that matches the user's interest based on the learning content of the automatic information detection unit. the learning system in section 8 providing the input layer 4 _1, 4 ₂ a plurality of stages in consideration of the mutual relationship between the input layers.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an automatic information detecting apparatus for selecting an information source that meets a user's need by learning the personality of the user. And improvement of extraction accuracy.

[0002]

2. Description of the Related Art Conventionally, a system for automatically extracting interest information of a user is disclosed in (1) Japanese Patent Laid-Open No. 6-124309.
Publication (information service and broadcast receiving system) and (2)
JP-A-5-314186 (automatic information source selection device)
It has been known. In both of the prior arts (1) and (2), input information is arranged on the same layer, and information desired by the user is extracted based on the operation result.

[0003]

In order to extract the information desired by the user in more detail by the conventional method, it is necessary to increase the amount of input information. However, as the amount of input information increases, the amount of calculation and memory required for learning and calculation of a neural network (hereinafter abbreviated as “NN”) also increases, so that in practice, , Learning system processing time, memory amount,
Alternatively, in order to reduce the amount of calculation to a practical range, the amount of input information must be limited. Hereinafter, the neural network (NN) will be specifically described with reference to FIGS.
The above-mentioned problems will be described briefly with reference to FIG.

[0004] A neural network (NN) comprises synapses corresponding to nerve fibers and intersections (nodes) on the network. The minimum unit consisting of the synapse and the node is called a cell. In the present invention, a node constituting the cell 1 is referred to as a basic element 2 and a synapse is referred to as a coupling element 3 (see FIG. 27).

FIG. 27 shows a cell 1 in the case of four inputs X1 to X4. The input signal Xi is input to the basic element 2 via the coupling element 3. The coupling element 3 has a certain weight (coupling weight: Wi), and the input signal X is input to the elementary element 2 which is weighted and coupled. The weighted input values (WiXi) are output after the sum X is obtained and the input values (WiXi) are transformed by the output function f. Output value y
Is represented by the following formula: y = f (X) (where X = ΣWiXi).

The output value y represents the state of the cell 1 at that time. The weight value W takes a positive value in the case of excitatory connection, and takes a negative value in the case of suppressive connection. Further, the weight value W can be changed by learning according to the plasticity of the coupling element 3.

The load sum X of the signal Xi input to the basic element 2 is output after being transformed according to the response characteristic (f) of the unit. There are various response characteristics. Representative examples include the following threshold function (Equation 1) and sigmoid function (Equation 2). f (X) = (1 (X ≧ 0), 0 (X <0)) (Equation 1) f (X) = 1 / (1 + exp (−X)) (Equation 2)

FIG. 26 shows a basic configuration of a neural network called a three-layer perceptron using the cell 1. When an external input is given to the first layer (input layer) 4, a signal is sent to a second layer (intermediate layer) 5 which converts an input signal by various couplings with the first layer 4, and is converted there. The output signal is sent to a third layer (output layer) 6, which is composed of a single cell, and after being summed up, is output through an output function.

In order to analyze the information input to the NN in more detail, it is necessary to increase the number of intermediate layers 5 and further increase the number of basic elements in each layer. As a result, the total sum of the elements (the sum of the number of basic elements constituting each layer and the number of coupling elements coupling each layer) increases. Among them, the number of coupling elements is increased by increasing the number of intermediate layers 5 and further increasing the number of basic elements of each layer in order to couple all the basic elements 2 in the preceding stage and all the basic elements 2 in the subsequent stage. Will be considerable. The major bottlenecks in the realization of NNs on a computer are the problem of memory consumption and the problem of the amount of calculation at the time of information detection and learning. In other words, the problem of the increase in the number of elements.

Considering the amount of calculation, the calculation of NN is performed by calculating the total sum (load sum) of inputs from the coupling element 3 obtained by multiplying the input value input from the preceding stage to the basic element 2 by weighting;
Calculation of an output function for calculating a value output from the basic element 2 is divided into two.

The calculation itself of the sum of the loads is simple. But,
As the number of coupling elements increases, the amount of calculation becomes enormous. The learning of the NN is performed by changing the weight of the coupling element 3. At this time, the amount of calculation of the single coupling element 3 depends on the algorithm, but it takes a calculation time proportional to the number of coupling elements to learn all the coupling elements 3.

Although the amount of calculation of the output function depends on the function, the amount of calculation of the basic element alone is not so large because a complicated function is basically not required. However, since the calculation is performed in all the basic elements 2, it is necessary to suppress the increase in the number of elements in order to improve the calculation speed.

Next, consider the memory consumption. The memory is required as a space for storing the weight of the coupling element 3. When detailed information analysis is performed using the NN, as described above, it is necessary to increase the number of the intermediate layers 5 and further increase the number of basic elements in each layer, thereby increasing the number of coupling elements. Therefore, memory analysis is required for detailed information analysis.

As described above, in order to perform detailed information analysis in the NN, a memory amount and a calculation amount are required. In other words, detailed information analysis cannot be performed with a limited memory space and a limited calculation capability of the CPU.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-mentioned problems, and a system which requires less memory and calculation than conventional techniques even when input information increases, that is, information extraction. It is an object of the present invention to provide an automatic information detection device that solves the conflicting problems of improving the accuracy of learning and improving learning efficiency.

[0016]

The present invention has the following configuration to achieve the above object. The invention of claim 1 is
An automatic learning device comprising: learning means for learning information of interest to a user from a history of input information; and information providing means for automatically providing information matching the interest of the user based on the learning content of the learning means. In the information detection device, the learning means may include a plurality of input layers for inputting one or more classification input information among the classification input information obtained by classifying the input information, in a plurality of stages in a learning processing procedure direction, and perform learning of a preceding input layer. An automatic information detection apparatus characterized in that a result is input to a subsequent input layer.

According to a second aspect of the present invention, there is provided the automatic information detecting apparatus according to the first aspect, wherein the same input layer has classified input information having an independent relationship.

The invention according to claim 3 is the automatic information detecting apparatus according to claim 1 or 2, wherein the learning means is constructed by a neural network.

According to a fourth aspect of the present invention, there is provided the automatic transmission system according to the third aspect, wherein a processing result other than the neural network is used as the processing result of the input layer, instead of the neural network processing in one of the input layers. It is an information detection device.

According to a fifth aspect of the present invention, the processing of any one of the classification input information of the subsequent input layer is restricted based on the output information of the previous input layer. It is an automatic information detection device as described.

According to the first aspect of the present invention, input layers for inputting one or more classification input information among the classification input information obtained by classifying the input information in the learning means are provided in a plurality of stages in the learning processing procedure direction, By inputting and reflecting the learning result of the former input layer to the latter input layer, instead of learning all the possibilities using the input information uniformly, it processes the classification input information input to the upper input layer and Inputting and reflecting the learning result of the upper input layer to the lower input layer is repeated, and ultimately the optimum information is provided to the user through the information providing means based on the output result of the lowermost input layer.

In each input layer, processing time is short because input information is not learned uniformly for all possibilities but is learned using classification input information, and the hardware resources required for processing are large. Does not need anything. Then, by reflecting the learning result of the former stage on the input layer of the latter stage, it is possible to finally obtain a result in which all input information is considered.

According to the second aspect of the present invention, since the same input layer has classified input information having an independent relation, unnecessary learning processing between the input information can be performed for input information having no mutual relation. It can be deleted, and processing time can be reduced and processing can be performed with small hardware resources. The classified input information having an independent relationship is processed independently and in parallel in the same input layer, so that unnecessary learning processing can be deleted while reflecting the output result of the preceding stage. Also in this case, since the processing is performed for each classification input information, the search processing time is short, and the burden on hardware resources is small.

According to the third aspect of the present invention, by constructing the learning means with a neural network, it is possible to efficiently learn information of interest to a user from the history of input information.

According to the fourth aspect of the present invention, a processing result other than the neural network is used as the processing result of the input layer instead of the neural network processing in any of the input layers. The processing efficiency can be improved by substituting for other means.

According to the fifth aspect of the present invention, the processing of one of the classification input information in the subsequent input layer is restricted based on the output information of the preceding input layer, thereby eliminating the apparently unnecessary learning processing. And the burden on hardware resources can be reduced.

[0027]

DESCRIPTION OF THE PREFERRED EMBODIMENTS In an automatic information retrieval apparatus (system) according to an embodiment of the present invention, input information is divided into upper and lower information, and the lower information is classified into categories according to the higher information. As a result, it is possible to reduce the number of coupling elements without impairing the information retrieval ability even when compared with input information processing in the same hierarchy. In other words, detailed information search can be performed even with a limited memory space and CPU operation capacity. In addition, by applying the automatic information retrieval system having these features to communication information, for example, detection of programs such as television and radio, and information detection of the Internet (e-mail system, etc.), expert systems, etc. Access to programs and information is facilitated. Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. still,
The same components as those described above are denoted by the same reference numerals and description thereof will be omitted.

[First Embodiment] FIG. 1 shows a first embodiment of the present invention.
1 is a block diagram of an automatic information detection device according to an embodiment of the present invention, which is schematically composed of an information selection unit 7, an automatic information detection unit 8, an access history management unit 9, an information database 10, and an information presentation unit 11.

The information selecting section 7 is an input means for accessing information desired by the user.
Tuners, keyboards, and the like are equivalent.

The automatic information detecting section 8 is a means for detecting information desired by the user by performing a learning processing operation, and based on information from the information selecting section 7, the access history managing section 9, and the information database 10. Do the learning. The processing contents will be described later.

[0031] The access history management unit 9 stores the interests of the user,
This system manages the storage, storage, and updating of the system usage status so that the type of hobbies and the like can be weighted. The selection information or the like specified by the user through the information selection unit 7 is used as input information. Weighting is applied to rules, for example, neural networks. Therefore, the access history management unit 9 returns the corresponding access information when there is an information request from the automatic information detection unit 8 and enables the automatic information detection unit 8 to perform timely learning processing. It is.

The information database 10 is a database for storing detailed information of information provided to the user, and when an information request is received from the automatic information detection unit 8,
The requested information is output and contributes to a learning process suited to the user's interests and hobbies.

The information presenting section 11 is an output means for providing the user with the information selected based on the result of the learning process of the automatic information detecting section 8, such as a television screen for providing an optimal television image, an optimal music, And a radio that provides broadcasting of radio stations, a computer screen that displays desired optimal information from the Internet, and displays an output of a predetermined expert system.

In the above configuration, the user selects the information selecting unit 7
When the user accesses various kinds of information by using the information, basically, the selected information (what information is needed) is sent to the automatic information detection unit 8, and the automatic information detection unit 8 detects the information that interests the user. For this purpose, a learning process is performed by extracting optimal information from the access history management unit 9 and the information database 10. The extracted information, which is the result of the learning process, is transmitted to the information presentation unit 11
, The extracted information content is displayed on the information presenting unit 11 as a response to the user's access.

The user who sees the contents of the response information presented to the information presenting section 11 can repeat the access to the information selecting section 7 again when the user is not satisfied with the information or desires a different one. Optimal information can be selected. It is also conceivable to directly access the information database 10 and directly output data to the information presenting unit 11.
Then, the behavior of the user in the above-mentioned information selecting unit 7 is sent to the access history managing unit 9 and is accumulated and managed as user information.

In the prior art, when the access history management unit 9 and the automatic information detection unit 8 shown in FIG. 1 use an automatic learning unit as a unit for extracting optimum information, all input information is treated as input information of the same level. However, in the present embodiment, the input information is classified and, after considering the mutual relation between the classified input information, the input layers are configured in a plurality of stages based on the mutual relation, and the input layers are processed in order from the upper input layer. By reducing the unnecessary processing by associating the processing result with the related lower input layer and performing the processing, it is possible to obtain the same processing result as the conventional parallel processing. . According to this configuration, it is possible to derive the contradictory effect of detecting more detailed information while suppressing the system size of the automatic information detection device.

In the following, a neural network NN, which is considered to be more effective if the input layer is configured in a plurality of stages and the lower input layer is classified for each category, is applied to the automatic learning means for extracting optimal information in the automatic information detection device. Will be described. As an example, the number of coupling elements in the conventional method (FIG. 2) when the number of inputs is the same and the method of this embodiment (FIG. 3) are simply calculated. In the calculation, the number of basic elements in each intermediate layer was the same as that of the input layer, and the number of stages in the intermediate layer was half the number of input elements. In FIG. 3, the input layer 4 ₁ , the intermediate layer 5 ₁ , and the output layer 6 ₁ of the NN on the upstream side (upper side) in the information processing direction are respectively provided with a subscript 1 and the downstream side (in the information processing direction). The lower NNs are distinguished by subscripts 2 respectively.

The number of coupling elements in the conventional method shown in FIG. 2 is (the number of inputs × the number of basic elements of each intermediate layer × the number of stages of the intermediate layer) =
(22 × 22 × 11) = 5324. On the other hand, the number of coupling elements in the present embodiment shown in FIG. 3 is (6 × 6 × 3 + 4 × 4 × 2 + 4 × 4 × 2 + 4 × 4 × 2 +
4 × 4 × 2 = 236). Above but in the calculation is that omitting the number of binding elements according to the output layer 6,6 _1, 6 _2, the difference between 5324 and 236 have been evident, the coupling element number as further increasing the NN The difference stands out.

Next, a case in which program information obtained by television broadcasting is used as an information source to automatically detect which program is most preferred by a user will be described with reference to FIGS. Here, words representing the features of the program in the program information are used as the input information. For example, a drama is a suspense, a love story, a comedy, and a music program is an artist name.

[0040] As shown in FIG. 6, the input layer 4 ₂ is another program category (drama, music, variety, news, etc.) to N
N is divided to input program information C (for example, suspense, love story, comedy for a drama, artist name for a music program, etc.). By the input layer 4 ₂ in this configuration, you learn interesting information 6 ₁ from the top of the NN in time information T for the program category, so that the further to learn more about each program category by the program information C. With this configuration, even if the input information X is the same as the conventional method, a small NN is added, so that the number of coupling elements is significantly reduced as compared with the conventional method, and the learning time is considerably reduced accordingly.

Here, the program information C from the information database 10 and the date and time information T from the clock (not shown) are acquired by using a selection signal from the remote controller (information selection unit 7) of the television, and the automatic information detection unit 8 performs learning. In the conventional method, as shown in FIG. 4, in the conventional method, a process of inputting to the same input layer 4 without considering the mutual relationship between the program information C and the date and time information T and connecting all nodes to each other in the intermediate layer 5 is performed. .

On the other hand, in this embodiment, enter the date and time information T to the input layer 4 _1, as shown in FIG. 5, it thus drama, learns to classify the program category of the variety such as lower falls input layer are entering program information C 4 ₂ in the form that corresponds to that category. Information in this embodiment as described above, by the input layer 4 ₁ higher always perform lower input layer 4 _second category classification, unchanged as parallel processing in the input layer 4 of single layer shown in FIG. 4 Detectability is possible.

FIG. 6 shows FIG. 5 more specifically, and the flow of the above-described category classification processing will be described with reference to FIG. The upper layer has date and time information T as input,
Here extracts a category for each of the interest degree of the lower layer side of the category classified program information C (the input layer 4 ₂₎ (output layer _61), input the degree of interest (output layer ₆₁₎ of the input layer 4 ₂ Each category is multiplied as a weight.

As described above, when the conventional input layer has one layer, the relationship between the respective pieces of input information which had an influence on the intermediate layer 5 is described. In the present embodiment, the input is performed for each small NN for each category in the lower layer. On the upper layer side, the substitution process is realized by extracting the degree of interest for each category of the lower layer NN from the input information.

As shown in FIG. 6, the above-mentioned processing includes the program category (drama, music, news, variety, etc.) and program information (suspense, love story, etc. for drama, singer name for music, etc.) If the relationship is known, that is, if the input information X can be classified and the inclusion relationship is clear, the multi-stage configuration of the input layer according to the present embodiment is effective for downsizing the automatic information detection device, improving the processing efficiency, and the like. It is.

Next, a case where the user selects a channel at 21:00 will be described as an example. FIG. 7 shows Y / Z / X2
It is a part of the program guide after 1:00, and in fact, the program guide of 100 or more channels is stored in the electronic program guide EPG (El
ectric Program Guide).

First, when the user has determined a program to be watched, a channel may be selected by inputting a channel number. In this case, the automatic detection system is not activated. If the user wants to select channels in the order of channel numbers, the channels may be sent in sequence using the channel UP button shown in FIG. 8, and in this case, the automatic detection system is not activated.

On the other hand, when the user does not determine the program to be viewed and wants to select programs in the order in which the user is interested, the automatic detection system according to the present embodiment is started. In an actual system, the automatic detection system is activated by the automatic button 13 or the like of the remote controller 12 corresponding to the information selection unit 7 (FIG. 8).

[0049] When the system is started, the input to the input layer 4 ₁ of the automatic information detection unit 8 is date and time information day, from the input layer 4 to the ₂ categories details information database 10 for each (this program if) Is done. Here, Y month, Z date, W day of the week, given four parameters of 21:00 as the input layer 4 _1, the input layer 4 ₂ give detailed information for each program.

[0050] The input layer 4 _2, the channel CH of the program guide
When detecting the degree of interest of one song program, the type of the program (countdown, MTV, variety, etc.) and the names of the performers are given as input information. Similarly, channel CH2
For drama, the drama genre (love story, suspense, comedy, etc.) and the name of the performer. For channel CH3 news, the type of news (regular, special (military, politics, economy, sports, etc.) input information) is applied to the input layer 4 _2, respectively. Thus, input information to the input layer 4 ₂ for each category are different.

[0051] First of all, from the input information to the input layer 4 _1, Y
On the output layer 6 ₁ , the weight (real value of 0 to ₁ ) of the genre that the user may be interested in at 21:00 on the month, Z day, W day, and 21:00
Is output to Here, for example, a song program (0.87),
It is assumed that drama (0.43), news (0.10), movie (0.36), and animation (0.15) are obtained. NN internal operation here is an operation for converting the output value weighted value of each coupling element 3 to the input information of the input layer 4 _of the basic element 2.

Next, the program information C of each channel CH is applied to the input layer 4 _2, in the case of channel CH1 is countdown as the type of information of the program, Kiroro as input performers, Takako Matsu given, Channel C
In the case of H2, suspense is given as program type information, and Takuya Kimura and Miho Nakayama are given as input of performers. Programs of the same genre (channel CH1, channel CH
5) Even if the program information is different, a different output appears.

The input value of each piece of information given to the input layer 4 ₂ is 1, and a value obtained by multiplying the input value by the weight extracted from the input layer 4 ₁ is the actual value of each category to the input layer 4 ₂ . It becomes the input value for each. The input value information in the present embodiment applied to the input layer 4 _second channel CH1, channel CH5 as described above is 0.87 (Popular Song Program), and in the case of channel CH2 and 0.43 (drama). Thus the input data to an input layer 4 ₂ of the system, carried out for all the programs, and extracts the user's interest degree 6 ₂ for each program. Similar to the case of calculating the NN work output 6 of the _internal herein is an operation for converting the output value weighted value of each coupling element 3 to the input information of the input layer 4 ₂ in the basic element 2.

Next, for the user to select a channel,
Sort program to be output to the information presentation unit 11 based on the output 6 ₂ in order of the degree of interest is performed. If the degree of interest is the same, the higher the larger the value of the output 6 _1, which is the degree of interest for each of the categories. If the categories are the same, they are arranged in channel CH order. The currently displayed channel CH (before selection) is placed next to the lowest-order program, separately from the channel CHs arranged in order of interest. This allows
The currently selected channel CH15 appears twice in the interest list (selection order 98, 101) (see FIG. 9).

As described above, when the user wants to search for an automatic program, the user first presses the automatic selection button 13 and then operates the up / down key 14 to select programs in descending order of interest and display information. It can be displayed on the section 11. In this operation, either the up key or the down key 14 is assigned to the descending order or the ascending order button. It is also conceivable that the one that starts pressing either the up key or the down key 14 is a descending order button, and the reverse is the ascending order button. In other words, at the beginning of pressing, the most interesting program is always displayed, and by repeatedly pressing the same button, the programs having the second and third interest levels are selected in descending order. By pressing the button on the opposite side, the programs are selected in ascending order. Such an operation with the up / down key 14 includes 1
There is an implicit understanding that if you go forward and return one, you will always return to the same place, so the currently selected channel C
H must be at the bottom.

In channel selection, when the highest order is selected, pressing the ascending button displays the lowest order channel, and pressing the descending button when selecting the lowest order displays the highest order channel. Is expected, the configuration is as shown in FIG.

Next, a learning procedure of the automatic information detection system will be described. The learning system basically performs learning in proportion to the length of the program viewing time. In this system, supervised learning is performed. The supervised learning is learning that gives whether the learning is correct or wrong when the NN performs learning.
The system performs learning at regular time intervals, and gives a teacher signal indicating that the selected program is correct, and gives a teacher signal indicating that all other programs are incorrect. This avoids the situation that, if there is no learning in the suppression direction, the learning of the program is accumulated as the time elapses, and in the worst case, the interest levels of all the programs may be 1. It is a measure to do.

[0058] Since all of the basic element 2 in the intermediate layer in the conventional system are attached to one another, for very even time learning according to the input layer 4 ₂ or less in the present system the genre NN Since there are a plurality of NN small-scale systems, the number of coupling elements is greatly reduced as compared with a conventional input layer 4 having a single-layer structure. Therefore, the learning speed is greatly increased.

Further, it is necessary to change the learning coefficient (coefficient indicating how much the learning is reflected on the NN) in this learning depending on whether the answer is correct or incorrect. Because 1
This is because there is one viewing channel out of 00 channels. When the learning coefficients are the same, it is easily expected that the weight difference between the coupling element 3 of the selected channel and the non-selected channel will be widened. This can result in a very oscillatory system (the channel of interest changes every time). The conventional system can sufficiently cope with changing the learning coefficient between the correct answer and the incorrect answer. However, in terms of say change of the learning coefficient is input layer 4 ₂ or less that constitutes a genre NN, NN small system of a plurality of the system configuration produces significant benefits. This is because the learning coefficient can be changed for each genre.

When the automatic information detection system is applied to television broadcasting as in the present embodiment, if a learning system is constructed with a conventional configuration, it is impossible to change the learning coefficient of a song program and news. If the same learning coefficient is used for a type of broadcast that has a change in taste following a trend such as a song program, and a type of broadcast that has a fixed degree of taste such as a news program or drama, a learning coefficient that matches the trend can be adopted. In addition, the degree of interest is changed every moment to a program having a fixed preference, such as a news program. Conversely, if a learning coefficient is adopted that matches a program whose taste is fixed to some extent, the degree of interest in a program that must follow the trend, such as a song program, will also be fixed, and the result will not be able to reflect user preferences. turn into. As described above, when the conventional method is applied, there is a possibility that the system may not be compatible with all programs. In the present embodiment, since a small NN group is formed for each genre, it is extremely easy to change the learning coefficient for each genre. Therefore, optimal learning can be performed for all genres.

In addition, with respect to input information, input information can be easily added but not easily deleted in the conventional format. On the other hand, in the present embodiment, input information management is performed for each category so that only input information can be added. Therefore, it is easy to delete the input information. Also, the present embodiment has a smaller effect on the NN when deleting the input information.

[Second Embodiment] In the first embodiment, the input layers 4 ₁ and 4 ₂ and the intermediate layer 5 are used for both the upper and lower NNs.
₁ , 5 ₂ , and the output layers 6 ₁ , 6 ₂ are provided. However, the efficiency of the NN processing can be improved by pre-processing the learning processing relating to any one of the input layers in a separate processing step. Therefore, a second embodiment will be described below. The same components as those described above are denoted by the same reference numerals and the description thereof will be omitted, and the following description will focus on the differences from the first embodiment.

FIGS. 10 and 11 show the configuration of the neural network NN according to the present embodiment. Figure 10 is one attached to the input layer 4 ₂ as an input layer 4 ₁ input layer 4 ₃ preprocess up output layer 6 ₁ from FIG. 3, and FIG. 11,
Preprocessing the input layer 4 ₃ of FIG. 10 are those provided between the intermediate layer 5 ₂ and the output layer 6 _2.

The pre-processing means uses, for example, a usage status table (see FIG. 16) in which the access status management section 9 stores the usage status of the user every week in a table according to a predetermined rule. For example, on a particular day time, in which any genre program outputs the automatic information detecting unit 8 whether the interest of the user as the input layer 4 ₃ is taken out from the usage table.

For example, from the input of the broadcast date and time 15 (input layer 4 ₁ ) shown in FIG.
(Output layer 6 ₁ ) is substituted by using the above-mentioned usage status table. As a result, the information detection system becomes genre 17
b, program information 18 (input layer 4 ₂ ), intermediate layer 19 (5 ₂ ),
And the degree of interest 20 (output layer 6 ₂ ) (FIG. 13). By using the pre-processing in this way, it is possible to reduce the load on the learning and information extraction processing in the automatic information detection unit 8 without reducing the width of the input information.

Next, as a system requiring preprocessing, a system for processing information as to which channel CH is most selected in the time zone is considered. Such a system is one of the very difficult to process by NN. It is meaningless to give the channel CH number as one input information as the input information. That is, FIG.
In the case of 1, 1 to 100
Will be entered. This configuration is effective when adjacent channels are broadcasting related programs, but otherwise it is meaningless. Because, when inputting a value of 1 to 100, for example, 50
This is because it is very difficult to distinguish between 49 and 51.

Therefore, in order to make the input of the channel CH valid, it is necessary to prepare inputs for the number of channels as shown in FIG. However, with this configuration, it is meaningful to use the channel CH as input data, but this is not practical because the system becomes too large.

[0068] Therefore, if you give the channel selection information as the input layer 4 ₃ pretreated in such a case, it is possible to be effective treatment also in terms say increasing amount of information in system size basis.

As shown in FIG. 16, a table of the day of the week and time is created, and the access frequency management unit 9 stores and manages the selection frequency CF of the channel CH there. Then, 17 that input as the input 4 _3, given as Figure 18. These neural network system, the date and time information of the input layer 4 _1, and the weighting for the lower layer output layer 6
_{Get one} . The output layer ₆₁ has a weight for each program genre in the lower of the input layer 4 _2. Therefore, input layer 4
Extracting channel viewing frequency CF using _one of the input information, the input 4 _3. In Figure 17, the weight for genre multiplying the input 4 ₃ and the output layer _61, FIG. 18
In reflects the channel information by applying the input 4 ₃ to the final output. With the above configuration, an optimal system can be created, and a large amount of information can be handled without constructing the system on a large scale.

[0070] In the Third Embodiment The first embodiment, as shown in FIG. 3, the input layer 4 of _the input from the output layer 6 ₁
There are determined, the output layer ₆₁ results had an output layer 6 ₂ linked to each of the lower input layer 4 _2. While reducing the memory size to be used for more neural networks in the third embodiment, and selectively carried to the intermediate layer 4 ₂ of the lower layer in order to perform the improvement of the processing speed.

[0071] The intermediate layer 4 ₂ selection method of the lower layer is not performed, the value for the input from each basic element 2 of the upper output layer ₆₁ to be connected to the intermediate layer 4 ₂ - what is included in the constant range shall elect to limit the lower falls input of the input layer 4 ₂ of the basic element 2 issued該選. For example, if the output value of the output layer ₆₁ are binary (0,1), the input layer 4 ₂ corresponding to the lower base element 2 of the output layer ₆₁ output value is 0 as such does not perform an input Apply to

FIG. 19 shows an example in which the above selection means is applied. The degree of interest in the genre 17 is derived from the input of the broadcast date and time 15, and the program information 18 corresponding to each genre 17 is input to an input layer of the program information 18 below the genre 17. The element having the largest value is selected from the basic element values corresponding to the genre 17 to be output. Then, the input at the subsequent stage is limited to only the program information 18 corresponding to the genre 17. If the genre selected in genre 17 is a drama, the program information 18 below will use love stories, suspense, comedy, etc. as input information. If the genre selected in genre 17 is news, economics, military, politics, etc. The information 18 is used as input information. The broadcast date and time 15 and the program information 1
8, the degree of interest 20 of the program is output from the two-stage input layer. Further, in this case, the degree of interest 20 is obtained in an upper stage than in FIG.

As described above, according to the present embodiment,
The amount of calculation at the time of information detection can be reduced. For this reason,
It is possible to detect information for a short time. The above method is particularly effective in the case of a pyramid type configuration in which input information increases as going to lower layers (FIG. 20).

In the configuration shown in FIG. 21, the output value (output 6 ₁ ) of the genre 17 from the broadcast date and time 15 which is the input layer is binary (1, 0). All may be 0. At this time, with respect to the genre having an output value of 0, the search in the subsequent stage is not executed, thereby limiting the search area.

In FIG. 21, the output 61 corresponding to the genre 17 is ₁ only for the drama of the lower layer, and all other genres are 0. As a result, only the operation for the drama needs to be performed for the second and lower layers. In the example of FIG. 21, there is little merit because the input has only two layers, but FIG. 22 shows an example in which the input is a three-layer input.

[0076] By giving the date and time information 15 in FIG. 22 in FIG. 21 similarly to the input layer 4 _1, to obtain an output layer 6 ₁ a is genre 17. Here, in the genre 17, the input of the lower layer is 1 only for music, and all other genres are 0.
It has become. As a result, only the operation for music is performed in the second and lower layers.

[0077] Then obtain an output layer 6 ₂ for the sub-category from the output layer ₆₁ to the input layer 4 _2. In FIG. 22, only the output corresponding to “MTV” which corresponds to the sub-category of the music program is “1”, and the other outputs are “0”. In this way, it can be seen that the three-layer configuration significantly reduces the amount of calculation due to the limitation of the search area. This becomes a more effective system as the number of input stages increases (FIG. 20).

Although the second and third embodiments are improvements of the first embodiment, faster information detection can be achieved by applying the third embodiment to the second embodiment. This will be described below as a fourth embodiment.

[Fourth Embodiment] An information retrieval system combining a system characterized by limiting a search area at the time of information retrieval with a system characterized by performing preprocessing on input information is provided by a TV broadcast program. The case where the present invention is applied to detection will be described with reference to FIGS. Note that the same parts as those described above are denoted by the same reference numerals, and description thereof will be omitted. FIG.
As shown in 3, 24, and outputs an output layer ₆₁ corresponding the broadcasting date and time information 15 to the lower genre 17 in binary.

[0080] In addition, it extracts the channel viewing frequency CF from the broadcast date and time information 15, performs weighting to the output _61. For example, at 21:00 on Monday, when examining the degree of interest in the song program of the channel, the viewing frequency CF of the channel is extracted from the date and time information 15. Using the example in FIG.
At 21:00 on Monday, the viewing frequency CF of the channel CH1 is 0.15. Viewing frequency CF of this channel CH1
Is weighted to the basic element 2 of the genre 17 (output layer 6 ₁ ) (see FIG. 23) or to the final output (FIG. 24) to extract the information of the viewing frequency CF. Can be reflected in This makes it possible to reflect the channel information that NN is not good at in a system in which the number of input layers increases as the number of input layers decreases as compared with the conventional system.

Although the preferred embodiment of the present invention has been described in the above embodiment, it is needless to say that the present invention is not limited to this.

[0082]

As described above, according to the present invention, the increase in the amount of calculation accompanying the increase in the amount of input information is suppressed, and the resources and processing time required for learning and calculation processing are reduced. For this reason, it has become possible to satisfy two conflicting problems of learning efficiency and information extraction accuracy.

According to the first aspect of the present invention, in each input layer,
The processing time is short and the hardware resources such as memory and CPU required for the processing are large because the input information is not learned uniformly for all possibilities but is learned using the classification input information. Do not need. Then, by reflecting the learning result of the former stage on the input layer of the latter stage, it is possible to finally obtain a result in which all input information is considered. Therefore, detailed learning and automatic detection can be quickly performed even for devices with limited capabilities such as a television receiver and a set-top box.

According to the second aspect of the present invention, since the same input layer has classified input information having an independent relationship, unnecessary learning processing between the input information can be performed for input information having no mutual relation. It can be deleted, and processing time can be reduced and processing can be performed with small hardware resources. The classified input information having an independent relationship is processed independently and in parallel in the same input layer, so that unnecessary learning processing can be deleted while reflecting the output result of the preceding stage. Also in this case, since the processing is performed for each classification input information, the search processing time is short, and the load on hardware resources such as a memory and a CPU can be reduced.

According to the third aspect of the present invention, by constructing the learning means with a neural network, it is possible to efficiently learn information that the user is interested in from the history of the input information.

According to the fourth aspect of the present invention, a processing result other than the neural network is used as the processing result of the input layer in place of the neural network processing in any of the input layers. The processing efficiency can be improved by substituting for other means.

According to the fifth aspect of the present invention, the processing of one of the classification input information in the subsequent input layer is restricted based on the output information in the previous input layer, thereby eliminating the apparently unnecessary learning processing. And load on hardware resources such as a memory and a CPU can be reduced.

[Brief description of the drawings]

FIG. 1 is a block diagram of an automatic information detection device according to a first embodiment of the present invention.

FIG. 2 is an operational explanatory diagram of a conventional basic neural network process.

FIG. 3 is an operational explanatory diagram of neural network processing of the automatic information detection device according to the first embodiment of the present invention.

FIG. 4 is an explanatory diagram in the case where the processing in FIG. 2 is applied to television program detection.

FIG. 5 is an explanatory diagram in the case where the processing in FIG. 3 is applied to television program detection.

FIG. 6 is an explanatory diagram in the case where the processing of FIG. 3 is applied to television program detection.

FIG. 7 is an explanatory diagram of a channel configuration according to the first embodiment of the present invention.

FIG. 8 is an explanatory diagram of a remote control device according to the first embodiment of the present invention.

FIG. 9 is an explanatory diagram of a channel arrangement according to the first embodiment of the present invention.

FIG. 10 is an operational explanatory diagram of the neural network processing of the automatic information detection device according to the second embodiment of the present invention.

FIG. 11 is an operational explanatory diagram of the neural network processing of the automatic information detection device according to the second embodiment of the present invention.

FIG. 12 is an operational explanatory diagram of the neural network processing of the automatic information detection device according to the first embodiment of the present invention.

FIG. 13 is an operational explanatory diagram of the neural network processing in which the preprocessing is performed in FIG.

FIG. 14 shows one channel cell C in the upper input layer of FIG.
It is an operational explanatory view of the neural network processing provided with H.

FIG. 15 is an operational explanatory diagram of a neural network process in which 100 channel cells are provided in the upper input layer of FIG. 12; FIG. 2 is a diagram showing cells of a neural network.

FIG. 16 is a partial explanatory diagram of a channel CH selection frequency management table.

FIG. 17 is an operational explanatory diagram of the neural network processing of the automatic information detection device according to the second embodiment of the present invention.

FIG. 18 is an operational explanatory diagram of the neural network processing of the automatic information detection device according to the second embodiment of the present invention.

FIG. 19 is an operational explanatory diagram of the neural network processing of the automatic information detection device according to the third embodiment of the present invention.

FIG. 20 is a schematic functional explanatory diagram of a neural network process of the automatic information detection device according to the third embodiment of the present invention.

FIG. 21 is an operational explanatory diagram of the neural network processing of the automatic information detection device according to the third embodiment of the present invention.

FIG. 22 is an operational explanatory diagram of the neural network processing of the automatic information detection device according to the third embodiment of the present invention.

FIG. 23 is an operational explanatory diagram of the neural network processing of the automatic information detection device according to the fourth embodiment of the present invention.

FIG. 24 is an operational explanatory diagram of the neural network processing of the automatic information detection device according to the fourth embodiment of the present invention.

FIG. 25 is an explanatory diagram of an operation of a cell of the neural network.

FIG. 26 is an explanatory diagram of a three-layer perceptron.

FIG. 27 is an explanatory diagram of a schematic configuration of a conventional neural network.

[Explanation of symbols]

1 cell 2 basic elements 3 coupled device 4,4 _1, 4 ₂ input layers 5 _1, 5 ₂ intermediate layer 6,6 _1, 6 ₂ output layer 7 information selecting unit 8 automatic information detector 9 access history management unit 10 Information database 11 Information presentation section CH channel CF Viewing frequency 4 ₃ Viewing frequency input NN neural network

Claims (5)

[Claims] 1. A learning means for learning information of interest to a user from a history of input information, and an information providing means for automatically providing information matching the user's interest based on the learning content of the learning means. In the automatic information detection device comprising: the learning means, provided in a plurality of stages in the learning processing procedure direction, an input layer for inputting one or two or more classification input information among the classification input information obtained by classifying the input information, An automatic information detecting apparatus, wherein a learning result of a preceding input layer is input to a succeeding input layer. 2. The automatic information detecting apparatus according to claim 1, wherein the same input layer has classified input information in an independent relationship. 3. The automatic information detecting apparatus according to claim 1, wherein said learning means is constructed by a neural network. 4. The automatic information detecting apparatus according to claim 3, wherein a processing result other than the neural network is used as a processing result of the input layer, instead of the neural network processing in one of the input layers. 5. The automatic information detection according to claim 1, wherein the processing of one of the classification input information of the subsequent input layer is restricted based on the output information of the previous input layer. apparatus.