JPH05120250A - Relational knowledge learning method - Google Patents

Abstract

PURPOSE:To easily specify and store a demand and learn relational knowledge so that the stored demand is approached by storing the demand for the correcting direction of the input/output result of a system and learning the relational knowledge from the stored demand. CONSTITUTION:A relational knowledge utilization result display part 105 accepts an input from a user, performs an inference process by utilizing the relational knowledge 104, and shows the result to the user. Then a demand input correspondence process part 106 accepts demands for the utilization results of the relational knowledge 104 from the user and stores them in demand items 107. A learning part 108 for the relational knowledge extracts input/output relations corresponding to individual demands described in the demand items 107 from the relational knowledge 104, adds input/output relations corresponding to demand items which can not be extracted to the relational knowledge 104, and alters the extracted relational knowledge 104 as the user demands to learn the knowledge.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a learning method of relational knowledge between input items and output items (hereinafter, simply referred to as "relational knowledge"). The present invention relates to a related knowledge learning method that is adapted to a user's request.

[0002]

2. Description of the Related Art As a technique for learning a correspondence relationship between an input and an output, a method using learning of a neural network is known. For example, DE Rumelh
"Learning internal expressions by error propagation, PD
According to "P model cognitive science and search for neuron network" (Sangyo Tosho pp. 321-334), this method is applied to a unit of an input layer and an output layer of a layered network which is one of neural networks. A unit is associated with each output item, a set of input and output to be learned is given as a teacher signal, and learning is performed using a learning algorithm such as an error backpropagation algorithm.
Further, as a technique for obtaining a causal relationship between events having many chains, for example, Japanese Patent Laid-Open No. 59-167 is available.
There is a causal knowledge acquisition device disclosed in Japanese Patent No. 771. This device utilizes an already systematized causal relationship when asking questions about cases experienced by experts, tracing the causal chain one after another according to the expert's response from the final event to the ultimate result. A function to create a question to find out the cause, a function to add or delete a causal relationship in the process, a function to present the result of comparison with the causal knowledge already accumulated at that time to the expert, and a question to the expert And a means for inputting and outputting causal relationships. Further, as a technique of gradually adapting the related knowledge to the user's request while the user uses it, there is a learning function of homophones and acronyms in the kana-kanji conversion of the Japanese word processor and the Japanese front-end processor. This is a technique for learning the relationship between the input "kana character string" and the output "converted character string" so that the conversion that matches the conversion characteristics of the user is performed as it is used. For example, according to Yasuo Ogawa's "Details of Grammar Processing for Kana-Kanji Conversion" (Separate Volume Transistor Technology Computer, July 1990, pp. 71-78),
The above-described learning of homophones can be roughly divided into two methods: usage frequency learning and latest usage priority. In the frequency-of-use learning method, the frequency of the homophones used is counted up each time it is used, and when the same homophone is converted next time, candidates are displayed in descending order of frequency. Also, in the latest usage priority method, each time it is used, the words used are rearranged at the beginning in the homophone group and registered in the dictionary, and when the same "kana character string" is converted, the converted character string candidates , The most recently converted "conversion character string" is displayed first.

[0003]

In the first of the above-mentioned prior arts, a set of input / output patterns is given to the network as teacher data for learning related knowledge. As a result, there was a problem that a correct input / output correspondence should be given. In the second conventional technique, the expert needs to answer questions for tracing the causal relationship one after another from the final event,
There was a problem that the burden on the expert would be heavy. When the third conventional technique is used for learning related knowledge, the frequency-of-use learning method only increases the number of points for determining the priority order of the used related knowledge, and the latest usage priority method is adopted. Since only the first priority is given to the character string, there is a problem that it is not possible to input a request for an inappropriate one or input a plurality of requests for one output. Furthermore, there is a problem in that the input item (kana character string) and the output item (converted character string) are directly linked and only the relational knowledge can be handled, and the relational knowledge between items having many chains cannot be handled. The present invention has been made in view of the above circumstances, and an object of the present invention is to solve the above problems in the related art, hold related knowledge, and use the related knowledge to identify items related to input items. It is an object of the present invention to provide a relational knowledge learning method in which a request for input / output results of the system can be easily specified / stored in a system for searching / outputting, and related knowledge is learned so as to approach the stored demand. In other words, one of the objects of the present invention is that it is not necessary to give a correct input / output relationship as teacher data to be given at the time of learning, and a user can learn related knowledge simply by inputting a request regarding output results. One is to provide a knowledge learning method, and the other is a relationship that enables the learning of related knowledge without the user answering each question to trace the causal relationship chain from the output item to the input item. To provide knowledge learning methods, and
One is to provide a relational knowledge learning method that allows a user to specify a plurality of requests for output results at the same time, making it easy to specify the desired result of the user.

[0004]

The above-mentioned object of the present invention is to retain relational knowledge indicating a relation between items, and use the relational knowledge to retrieve / output an item related to the input item from an input item. This is achieved by a relational knowledge learning method characterized by accumulating requests regarding correction directions for input / output results of the system in a processing system, and learning related knowledge from the accumulated requests.

[0005]

In the relational knowledge learning method according to the present invention, the user can easily specify / accumulate the request for the input / output result of the system by the requirement input / accumulation function, and by the relational knowledge extraction function and the relational knowledge change function. , It becomes possible to learn related knowledge so as to approach the accumulated desires. Hereinafter, it will be described more specifically. When the related knowledge is used, the result at that time is presented to the user by the related knowledge use result display function, and at the same time, the user inputs the result by the request input / accumulation function to determine whether or not the result is the intended result. Just enter the request,
Requirements for relational knowledge can be recorded, and it becomes unnecessary to input an accurate input / output correspondence relationship for learning the relational knowledge. The relational knowledge extraction function automatically searches for the relational knowledge corresponding to the input requirements, so you have to answer the questions to trace the causal relations one after another from the final event. The burden on the specialist can be reduced. In addition, when the user provides information that expresses the request regarding the output result by the request input / accumulation function, the command corresponding to the case of "We want to weaken the relationship" is prepared. Not only that, you can also specify the output items you want to weaken the relationship. In addition, commands are provided for learning the relationship between new input items and output items, so even if there is no relationship knowledge between the input items and the items for which output is desired can do. In addition,
For the requirements for which the related knowledge extraction function could not extract the corresponding related knowledge with the related knowledge addition function, a new route is established by connecting the input item and the output item in the related knowledge Related knowledge can be added. Moreover, by changing the relational knowledge, the relational knowledge extracted by the relational knowledge extracting function is increased / decreased at a fixed rate according to the degree of connection,
You can learn to strengthen or weaken the relevant relationship knowledge. As a result, after that, the output when the same input is made can be brought closer to the result intended by the user, and the related knowledge can be adapted to the user's request as it is used.

[0006]

Embodiments of the present invention will now be described in detail with reference to the drawings. FIG. 2 shows a device configuration of an information processing system which is an embodiment of the present invention. In the figure, 204 is operated by a built-in program 205, and the keyboard 20
2 is a processing device that receives inputs from the mouse 203 and the mouse 203, performs calculation / inference processing using the knowledge base 207 stored in the external storage device 206, and outputs the result to the display 201. As described above, the keyboard 202 and the mouse 203 each have a function of receiving information regarding the input of characters from the user and the position on the display, and transmitting the information to the processing device 204. Further, the display 201 has a function of receiving an output from the processing device 204 and displaying it on the screen. External storage device 206
Stores a program and data necessary for executing the program and a knowledge base 207. Knowledge base 2
The related knowledge and the like necessary for the inference processing performed by the processing device 204 are stored in 07. FIG. 1B is a diagram showing the system configuration of this embodiment. In the figure, reference numeral 104 indicates relational knowledge. 105 receives an input from the user, performs inference processing using the above-mentioned relational knowledge 104,
The related knowledge utilization result display part which displays a result to a user is shown. Reference numeral 106 denotes a request input handling processing unit that receives a request for a related knowledge use result from a user and accumulates it in a request item 107. Further, 108 extracts the input / output relationship corresponding to each request described in the above-mentioned request item 107 from the above-mentioned relationship knowledge 104, and adds the input / output relationship corresponding to the request item that could not be extracted to the relationship knowledge 104. Then, the learning section of the relational knowledge for learning is shown by changing the extracted relational knowledge so as to meet the user's request. FIG. 3 shows an example of relational knowledge handled in this embodiment. The relational knowledge shown in this figure represents the relation between documents and keywords for searching them. In the figure, 301 to 304 are nodes representing keywords, and 3
Reference numerals 05 to 307 are nodes representing documents. 308-3
Arrows having weights (subscripts) such as 13 represent connections indicating relationships between nodes. In this case, for example, if inference is executed by inputting the keyword 301, the document 30
5, the document 306 will be obtained as output. Next, the processing flow in this embodiment will be described with reference to FIG. In step 101, the inference process is performed using the related knowledge based on the input from the user, and the result of using the related knowledge is displayed to the user. Here, although the example of the related knowledge to be used is shown in FIG. 3, as an example of the usage result display, the document “No. 1_1 Depositor's inheritance when the keyword“ deposit ”is input with the related knowledge of FIG. To ... "
The screen output of is shown in item 4. Next, in step 102, a request from the user regarding the use result of the displayed related knowledge is received as an input, and this is stored as a request item, for example, as shown in FIG. 5, and the relationship to the stored request item is stored. Knowledge is extracted from the current relational knowledge as shown in FIG. 6, and when it cannot be extracted, new relational knowledge is added as shown in FIG. , Learning is performed in the direction of reducing the difference between the result of using the related knowledge and the user's request.

Next, the step 102 in FIG. 1 will be described in detail with reference to FIGS. Step 80 of FIG.
In 1, the user inputs a request for the related knowledge utilization result currently displayed on the display,
Also, a window for entering commands for the working file is displayed. An example of the request input screen at this time is shown in FIG. The processing corresponding to the user input command will be described below. (1) When a “strength” command is input For example, the input items input by the user are A1 and A.
2, ... Aα (α ≧ 1), assuming that the output item displayed on the display is B, file processing 1 is executed in step 804, and “A1, A2, ... A
One record of “α → B up” (501 in FIG. 5) is entered in the work file (step 901 in FIG. 9). (2) When the "weak" command is input For example, the input items input by the user are A1 and A.
2, ... Aβ (β ≧ 1), assuming that the output item displayed on the display is C, file processing 2 is executed in step 805, and “A1, A2, ... A
One record of “β → C down” (502 in FIG. 5) is entered in the work file (step 100 in FIG. 10).
1). (3) When the "register" command is input For example, the input items input by the user are A1 and A.
2, ... Aγ (γ ≧ 1), step 8
In 06, the file processing 3 is executed, the output item desired by the user is accepted as a keyboard input, and it is D
(Step 1101 in FIG. 11), "A1, A2, ...
..One record of “Aγ → D add” (503 in FIG. 5)
In the work file (step 110 in FIG. 11)
2). (4) When the "Cancel" command is input File processing 4 is executed in step 807 to delete the latest one record in the work file (Fig. 12).
Step 1201). (5) When the “Cancel All” command is input File processing 5 is executed in step 808 to delete all records in the work file (FIG. 13).
Step 1301). (6) When the "End" command is input In step 802, file processing 0 is executed, the contents of the work file are accumulated as request items (step 1401 in FIG. 14), and a request input window is opened in step 803. Erase from the display.

Next, referring to FIG. 15, step 1 of FIG.
03 will be described in detail. First, in step 1501, the causal chain of relational knowledge is traced backward from the node corresponding to the output item in one record of the desired item, and the nodes that have passed to the end or the nodes located at the end are recorded. In the example of FIG. 6, since the node corresponding to the output item is E, the nodes recorded in this step are A, B, C, D, and F. In step 1502, the node recorded in step 1501 is determined as the node corresponding to the input item in one record of the same request. Taking the example of FIG. 6 as before, since the nodes corresponding to the input items are B and C, the nodes obtained in this step are
B and C. Steps 1501 and 15 described above
In the process of 02, it is checked whether or not there is a corresponding one in the request knowledge entered by the user using the "register" command in step 102. In the process of step 1502, when one or more corresponding items are obtained, in step 1503, the strength of the connection that is the route from the obtained node to the output item is set at a constant ratio according to the requirement. (For example, if you want to strengthen the relationship, it is 1.1 times, and if you want to weaken the relationship, it is 0.9 times). Further, when the corresponding node is not obtained in the process of step 1502, in step 1504, relational knowledge that links the input item and the output item of the desired item targeted at that time is added. The requirements to be covered in this step are the requirements entered by the user using the “register” command in step 102 of FIG. 1, and the steps 1501 and 1 in FIG.
In 502, the corresponding relational knowledge could not be extracted. In the case of FIG. 7 as an example, the connection from the node F corresponding to the input item to the node E corresponding to the output item corresponds to the relational knowledge added in this step. The above-described processing is repeated for each of the demands accumulated in step 102 of FIG. According to the above-described embodiment, the user can easily specify / store the request for the input / output result of the system by inputting / storing the desired items, and by the relational knowledge extracting function and the relational knowledge changing function, It becomes possible to learn related knowledge so as to approach the demands made. Furthermore, for the requirements for which the related knowledge extraction function could not extract the corresponding related knowledge by the related knowledge addition function, a new route for connecting the input item and the output item in the related knowledge is provided. It will be possible to add new relationship knowledge. In addition, the relational knowledge change function increases or decreases the degree of connection of the relational knowledge extracted by the relational knowledge extraction function at a fixed rate according to the request, thereby weakening the relevant relational knowledge. You can also learn in the direction. It is needless to say that the above-mentioned embodiment shows an example of the present invention, and the present invention is not limited to this.

[0009]

As described above in detail, according to the present invention, the relational knowledge indicating the relation between items is held, and the relational knowledge is used to search for an item related to the input item from the input item. In the information processing system for outputting, by storing the request concerning the correction direction for the input / output result of the system and learning the related knowledge from the stored request, the related knowledge is held and the related knowledge is used. In a system that searches and outputs items related to input items, it is possible to easily specify and store requests for the input and output results of the system, and realize a relational knowledge learning method that learns related knowledge so as to approach the accumulated demands. It has a remarkable effect.

[0010]

[Brief description of drawings]

FIG. 1 is a diagram showing a system configuration and an operation flow of an information processing system which is an embodiment of the present invention.

FIG. 2 is a diagram illustrating a device configuration of an information processing system according to an embodiment.

FIG. 3 is a diagram showing an example of relational knowledge handled in the present embodiment.

FIG. 4 is a diagram showing an example of a request input screen in the embodiment.

FIG. 5 is a diagram showing an example of accumulated requirements in the embodiment.

FIG. 6 is a diagram illustrating extraction of relational knowledge according to an embodiment.

FIG. 7 is a diagram illustrating addition of related knowledge in an example.

FIG. 8 is a flowchart showing details of the operation in the embodiment.

FIG. 9 is a flowchart showing details of the operation in the embodiment.

FIG. 10 is a flowchart showing details of the operation in the embodiment.

FIG. 11 is a flowchart showing details of the operation in the embodiment.

FIG. 12 is a flowchart showing details of the operation in the embodiment.

FIG. 13 is a flowchart showing details of the operation in the embodiment.

FIG. 14 is a flowchart showing details of the operation in the embodiment.

FIG. 15 is a flowchart showing details of the operation in the embodiment.

[Explanation of symbols]

 104 Relational Knowledge 105 Relational Knowledge Usage Result Display Unit 106 Request Input Corresponding Processing Unit 107 Request Items 108 Relational Knowledge Learning Unit 201 Display 202 Keyboard 203 Mouse 204 Processing Device 205 Program 206 External Storage Device 207 Knowledge Base

─────────────────────────────────────────────────── ───

[Procedure amendment]

[Submission date] November 20, 1992

[Procedure Amendment 1]

[Document name to be corrected] Drawing

[Correction target item name] All drawings

[Correction method] Change

[Correction content]

[Figure 1]

[Fig. 2]

[Figure 3]

[Figure 4]

[Figure 5]

[Figure 6]

[Figure 7]

[Figure 8]

[Figure 9]

[Figure 10]

FIG. 11

[Fig. 12]

[Fig. 13]

FIG. 14

FIG. 15

 ─────────────────────────────────────────────────── ─── Continuation of front page (72) Inventor Masahiro Hirose 1099, Ozenji, Aso-ku, Kawasaki-shi, Kanagawa Ltd. System Development Laboratory, Hitachi, Ltd.

Claims (5)

[Claims] 1. Holding relational knowledge indicating a relation between items,
In an information processing system for searching and outputting an item related to an input item from an input item using the related knowledge, a request regarding a correction direction for an input / output result of the system is accumulated, and learning of the related knowledge is performed from the accumulated request. A method for learning relational knowledge, characterized by performing. 2. The learning of the relational knowledge extracts the relational knowledge by tracing the input items from the output items representing the result, changes the strength of the extracted relational knowledge, and newly adds the relational knowledge when the relational knowledge cannot be extracted. The method for learning relational knowledge according to claim 1, wherein 3. The relational knowledge learning method according to claim 1, wherein it is possible to specify, as a request regarding the correction direction, two directions, a direction for strengthening an input-output relationship and a direction for weakening the input-output relationship. 4. When a plurality of items are output as a result of using the relational knowledge, it is possible to specify a request regarding the correction direction for each of the individual output items. Or the related knowledge learning method described in 3. 5. The information processing system is a system for retrieving a document from a keyword, and learning of the relationship knowledge is for learning a relationship between a keyword and a document and a relationship between a keyword and a keyword. The relational knowledge learning method according to claim 1.