JPH03269629A - Knowledge reflinement processing system using example - Google Patents

Abstract

PURPOSE:To improve the knowledge refining efficiency by tracing a decision tree and a sorting tree for inspection of the conclusion drawn with combination of conditional values and updating the knowledge after designing the undesigned conditional value or adding the conditions. CONSTITUTION:When a user 8 inputs a question to a knowledge base system 1, an inference mechanism 5 performs the inference processing based on a generated decision tree 12 and draws a conclusion. An example arrangement part 13 generates a sorting tree 19 from the examples stored in an example base 6. A knowledge refining mechanism 4 performs comparison between both trees 12 and 19, i.e., the knowledges stored in a knowledge base 3 and detects four patterns. For instance, the tree 19 contains the combination of the conditional values included in the tree 12 and at the same time the conclusion drawn by the relevant combination detects a pattern included in the tree 12. Thus a conclusion adding part 15 shows the corresponding combination of conditional values stored in the tree 19 to the user 8 to instruct the user 8 to input the information on the definition of conclusion for addition of conditions to a conclusion definition knowledge 11 and the tree 12. As a result, the knowledges are freeely extracted out of the examples.

Description

[Detailed description of the invention] [Industrial application field] The present invention relates to knowledge-based systems, and more particularly.

A knowledge base system G that expresses the knowledge of experts in a predetermined format and stores it in a knowledge base, and solves problems on behalf of the experts by making inferences using the stored knowledge. The present invention relates to a knowledge refinement processing method using examples, which makes it possible to automatically refine knowledge by using the resulting combinations of condition values as examples.

[Conventional technology]

Knowledge base systems have been developed that solve problems on behalf of experts by storing the knowledge of experts in a knowledge base and making inferences using the stored knowledge. In such a knowledge heath system, expert knowledge is input expressed in a predetermined format, refined and edited, and then stored. The stored knowledge in a predetermined form is used to make inferences that lead to conclusions.

Knowledge representation is expressed, for example, as tree-structured data as shown in FIG. The tree structure is generated from the knowledge representation as shown in FIG.

Such tree-structured knowledge data is tree-structured data indicating a bus for deriving a conclusion from a plurality of conditions, and is called a decision tree. for example.

A decision tree as shown in FIG. 6 is generated from a decision table of input knowledge by a tabular knowledge editor that expresses knowledge in a table format and inputs knowledge data.

In the decision tree shown in FIG. 6, "condition" indicates a condition judgment node, and the value of the block added under each condition judgment node indicates the condition (Ii (judgment value)) for the condition judgment. The knowledge representation in the decision tree is 9. For example, if the decision value is al for condition A of the 1-condition judgment node, it indicates that the decision tree is followed downward and the judgment of condition B of the condition judgment node is made in the first place. 3. Then, the decision tree is traced further down, and if necessary, for example, the condition C of the condition judgment node is determined.In this way, from the knowledge representation of the decision tree, the decision tree is successively traced down, and conclusions 1 to 8 are determined. In this way, by creating a decision tree as shown in Figure 6 from the input knowledge data and following the generated decision tree,
Inference execution will be performed using the knowledge data.

When different conclusions are drawn by combining multiple condition values, a common method is to create a decision tree like the one shown in Figure 6 and search this decision tree while obtaining the values of the five conditions. .

By the way, in FIG. 6, the judgment values al, a2 . a3 indicates the attribute to 9 conditions, and similarly, the judgment values bt, . . . 03, etc. indicate the attribute of 1 condition B and the attribute of 1 condition C. In a decision tree search,
First, when the value of condition A is obtained, the branch that has the attribute that matches that value is traced, and the value of condition B is manually determined first, and the technique that has the attribute that matches that value is traced. In this way, the decision tree is sequentially traced to reach a conclusion regarding one condition C in the same manner.

For example, to reach conclusion 2, there are 9 conditions A3 conditions B.

and al, b2. as attributes of condition C. cl (hereinafter (
at, b2. (denoted as cl)) is specified, and the value of the condition that matches this is derived as the obtained result. Some attributes may not be specified in the conclusion. If an attribute that is not specified is expressed as "-"9 For example, conclusion 4 has the attributes (a2, -, -) specified, and 7 just by obtaining a2 as the value of condition A, conclusion 4 is be guided. In other words, the value a2 of 1 condition A is a valid condition value for deriving conclusion 4, and the combination of condition values (a2.-,-) is a valid attribute.

In knowledge input, if the same conclusion is reached when only some condition values match and other condition values do not match, the unmatched conditions are made irrelevant to the conclusion derivation, and they are It is necessary to integrate For example, attribute 1 (at, bl, c2) is different from the one shown in Figure 6.
, attribute 2 (albl, c3), and attribute 3 (al,
bl, cl) each reach conclusion 1, the value of condition C is not related to the derivation of conclusion 1, so the value of attribute 3 should not be specified.

Attributes 1 to 3 are integrated as attributes (al, bl, ).

Furthermore, if the combinations of condition values match and different conclusions are reached, it is necessary to specify condition values or add conditions to distinguish them. For example, attribute 4 (a L, -, c2) reaches conclusion 2, and attribute 5 (al, b2.c
If 2) reaches conclusion 3, the condition value of condition B, which does not specify a value.

It is necessary to differentiate attribute 4 and attribute 5 by specifying "bl" in attribute 4. Also, when the values of condition B between attribute 4 and attribute 5 cannot be distinguished.

Alternatively, if it cannot be specified, it is necessary to differentiate attribute 4 and attribute 5 by adding a condition with a condition value that can be distinguished by one.

(Problem to be solved by the invention) By the way, as mentioned in 5 above, before inputting knowledge, it is necessary to extract attributes (combinations of condition values) that are effective for deriving a conclusion and integrate the attributes, or to specify condition values. It is necessary to refine the knowledge by distinguishing attributes by adding conditions.

Conventional knowledge refinement was performed by humans using a large number of case data.

In the conventional method, there was a problem that the amount of case data handled by five people was large, the time required to extract attributes was large, and the cost increased.

Furthermore, conventional knowledge base systems have a problem in that they cannot deal with combinations of condition values that do not exist in decision trees using knowledge stored in the knowledge base.

The present invention has been made to solve the above problems.

An object of the present invention is to provide a knowledge refinement processing method that can easily refine knowledge using the knowledge and examples stored in a knowledge base in a knowledge base.

[Means to solve the problem]

In order to achieve the above objective, a knowledge refinement processing method using an example of the present invention performs the first processing shown below, and uses the classification tree generated there and the knowledge stored in the knowledge base. A comparison is made with the decision tree, and based on the results, the second. Third. It is characterized by performing the fourth or fifth processing.

Process U: Generate a classification tree using combinations of condition values obtained as examples.

This classification tree is close to a decision tree. It also includes cases where no conclusion is reached. Ru. There is no correspondence with the conclusion. display the classification tree and provide the user with a conclusion. Have them input.

Second process: Reconfigure the knowledge based on the decision tree using the knowledge already stored in the knowledge base and the classification tree generated in the i-th process, and remove the attributes of the conditions that are irrelevant to the derivation of the conclusion. Select them and integrate their attributes.

Third process: If an attribute that does not exist in the decision tree exists in the classification tree and cannot be integrated with any attribute in the decision tree, this is displayed to the user and added to the knowledge base as new knowledge.

Fourth processing: If the combination of condition values of the two-classification tree is included in the decision tree and the conclusions are different, the existence of a condition for which two condition values are not specified can be searched for, and the condition value can be specified. In this case, specify 9 condition values.

Fifth process: If 9 condition values cannot be specified in the fourth process, or if there are no conditions for which no condition value is specified, a condition is added that can distinguish both buses.

[Operation] According to the above-mentioned means, the knowledge is refined using a combination of the decision tree 5 stored in the knowledge base and the condition values existing as examples, created by a table-format knowledge editor in the knowledge heath system. .

In this case, the first process is to integrate the combinations of condition values in the example and generate tree-structured data called classification water in which one condition corresponds to the Nord and one condition value corresponds to the branching from the node. If the combination of condition values does not correspond to the conclusion, perform inference using the knowledge-based system,
Correspond with the conclusion. If no conclusion is reached, obtain the corresponding conclusion core.

Compare the classification water obtained in the first process and the decision tree.

If the classification water has a combination of condition values that does not exist in the decision tree, and the conclusion is the same as the conclusion that exists in the decision tree, 1. The decision tree corresponding to the conclusion.

Extract the parts that have different branches from the classification tree.

Consolidate attributes of conditions. In addition, if there are trees with the same conclusion but different combinations of condition values in one classification tree, extract the parts with different branches,
If there are five defined weepings, a second process is performed to integrate the attributes of the conditions that are nodes that cause the branching.

As in step 2, follow the decision tree and classification water, examine the combination of condition values that reach the conclusion, integrate the attributes of the two conditions, and update only the attributes of the conditions that are effective for deriving the conclusion as new knowledge. .

In addition, if the 9-class water has a combination of condition values that does not exist in the decision tree, it will display to the user that it is a combination of condition values that cannot reach a conclusion with the current knowledge base, A third process of adding to the knowledge base is performed to add knowledge that has not yet been stored in step 2 to the knowledge base.

Combinations of condition values that the classified water has are present in the decision tree,
If a different conclusion is reached, a search is made for the existence of an unspecified condition in the combination of condition values of interest. An unspecified condition exists.

If it is possible to distinguish combinations of condition values by specifying the condition values, perform a fourth process of specifying two condition values,
If there is no impossible or unspecified condition, a fifth process is performed in which this fact is displayed to the user and the user is prompted to additionally input a condition that can distinguish between the two condition values.

In this way, the decision tree and the classification tree are followed, the conclusion derived from the combination of 1 condition values is inspected, unspecified condition values are designated 5 or conditions are added, and knowledge is updated 9.

〔Example〕

An embodiment of the present invention will be specifically described below with reference to the drawings.

FIG. 1 is a block diagram showing the main components of a knowledge base system according to an embodiment of the present invention.

In FIG. 1, 1 is a knowledge base system, 2 is a tabular knowledge editor, 3 is a knowledge base, 4 is a knowledge refinement mechanism, and 5 is a knowledge base system.
is a reasoning mechanism, and 6 is a case-based one.

The tabular knowledge editor 2 receives knowledge input from the expert 7 on a tabular screen and generates a decision table 9 expressing the knowledge in a tabular format. Decision table 9 is a table generated by associating condition items and conclusion items with each row and column of table 9, and specifying the condition value of each condition item corresponding to each conclusion item as an element of the table.

Knowledge expression can be unified and handled in a predetermined tabular format, so that knowledge can be easily expressed and input. Also 1 tabular knowledge editor 2.

A decision tree 12 for deriving a conclusion from the condition values of each condition item is generated from the generated decision table 9, and the generated decision tree 12 is stored in the knowledge base 3 so that it can be referenced in inference processing.

For example, when the user 8 inputs a question to the knowledge base system 1, the inference mechanism 5 performs inference processing and derives a conclusion based on the decision tree 12 generated as described above. This inference result can be stored as an example in the case base 6 together with the data used for inference (combination of condition values 18). The knowledge refinement mechanism 4 also includes a case organization section 13, a conclusion integration section 14. The conclusion addition section 151, the condition value specification section 16, and the condition addition section 17 warn each other. The case organizer 13 generates classification water 19 from the cases stored in the case base 6. If there is a bus that does not correspond to a conclusion, this fact is displayed to the user and the user is encouraged to input the corresponding conclusion.

The knowledge refinement mechanism 4 compares the decision tree 12, which is the knowledge stored in the knowledge base 3, with the classification water 19 stored in the case 6, and detects four patterns of the first order.

Pattern 1: Classified water has a combination of condition values that is partially different from the combination of condition values in the decision tree and that leads to the same conclusion.

Pattern 2: The classification water has a combination of condition values that does not exist in the decision tree, and the conclusion derived from it does not exist in the decision tree.

Pattern 3: The classification water has a combination of condition values that is the same as or has an inclusive relationship with the combination of condition values that the decision tree has, and that leads to a different conclusion, and there is a condition for which no condition value is specified.

Pattern 4: The classified water has a combination of condition values that is the same as the combination of condition values that the decision tree has, and that leads to a different conclusion, and there is no condition for which no condition value is specified.

When pattern 1 is detected, the 5-conclusion area combining section 14 performs.

Different condition values are extracted and one condition definition knowledge IO is used to check whether the different condition values are all the defined condition values. If all t of condition values exist in a definition, the condition value designation of that condition is canceled and the decision tree 12 and classification water 19 are updated. Also,
If all of the defined condition values do not exist, nine condition values are integrated.

When pattern 2 is detected, the conclusion addition unit 15.

User 8 selects the combination of applicable condition values in classified water 19.
, and let you enter information about the conclusion definition, and check the conclusion definition knowledge! 1m 11 and addition to the decision tree 12.

When pattern 3 is detected, the 2-condition value specifying unit 16 performs the following.

This information and the combinations of condition values for each of the corresponding decision trees and classification trees are displayed to the user 8, and the user 8 is prompted to input the condition values for distinguishing between the combinations of one condition value, and the decision tree 12 is updated. If the user 8 is unable to input a condition value for distinguishing between combinations of 1 condition value, it is assumed that the condition value is equivalent to pattern 4, and the 9 condition addition unit 17 performs the process.

When pattern 4 is detected, the 1-condition addition unit 17 performs the following.

Display that fact and the combination of condition values of the corresponding decision tree 1 classification tree to the user 8, and let the user 8 add a new condition 9 Addition to condition definition knowledge 11i 10 6 and decision tree 12
Update.

FIG. 2 is a diagram showing a specific example of processing performed by the conclusion integration unit.

The attributes (al, bl) that the decision tree has lead to conclusion 1.
Attributes (al, b2) that the classified water has.

(at, b3) derives conclusion 1, and if it is defined as condition definition knowledge that the possible values of condition B are bl, b2゜b3, then the value of condition A is used to derive conclusion l. a
It is related to the fact that l.

Since the values of condition B are irrelevant, the attributes (al, bl), (al, b2), (at) that derive the conclusion l.

b3) is integrated into (al,-) to generate a decision tree “
If the value of condition A is al and the value of condition B is bl, then the conclusion is 1.'' and ``If the value of condition A is 31 and the value of condition B is b2, then the conclusion is 1.'' The three knowledges ``If the value of condition A is al and the value of condition B is b3, then the conclusion l is'' are ``If the value of condition A is al, then the conclusion l.'' ” will be updated to the knowledge.

Also, if the classification tree has only attributes (a 1. b2), it is integrated using (al, bl or b2).

The user is asked about the uninput attributes (al, b3), and if information is input that this cannot occur or that conclusion 1 is derived, it is integrated with (al).

FIG. 3 is a diagram showing a specific example of the processing performed by the conclusion addition section.

When the conclusion addition unit detects that classification water having an attribute (alb2) that does not exist in the decision tree derives a conclusion 2 that does not exist in the decision tree, the conclusion addition unit adds the attribute (alb2).

b2) is added to the decision tree by displaying to the user that it does not exist in the decision tree.

FIG. 4 is a diagram showing a specific example of processing performed by the 1-condition value designation section.

The condition value specifying part is the attribute (a I) in the decision tree.

bl) derives conclusion 1, and the attribute (a 1
．． -), if it is detected that conclusion 2 is derived, the user is informed that the combination of 1 condition values is in an inclusive relationship, but the conclusions are different.

and combinations of condition values are displayed, prompting the user to specify unspecified condition values, and when the combinations of conditions can be distinguished based on the input condition values, they are added to the decision tree.

Similarly, when it is detected that the attribute (al, -) in the decision tree derives conclusion 1, and the classified water with the same attribute (al, -) derives conclusion 2, the user Displays a message indicating that the combination of condition values is the same but the conclusion is different and the combination of condition values, prompts the user to specify unspecified condition values, and distinguishes between condition combinations based on the input condition values. Additions to the decision tree are made when possible.

FIG. 5 is a diagram showing a specific example of processing performed by the three-condition adding section.

The condition addition part derives conclusion 1 for attributes (al, bl) in the decision tree, and calculates the same attributes (at, bl) for 1-class water.
) is detected to derive conclusion 2, the user is informed that the combination of condition values is the same but the conclusion is different, and the combination of condition values is displayed, and the addition of 5 conditions and the addition of the added condition are displayed. Input the value specification and add to the condition definition knowledge and update the decision tree.

Although the present invention has been specifically described above based on the embodiments, it goes without saying that the present invention is not limited to the embodiments described above, and can be modified in various ways without departing from the gist thereof.

〔Effect of the invention〕

As described above, according to the present invention, it is possible to automatically update already stored knowledge or extract knowledge from cases using cases.

This has the effect of greatly improving the efficiency of knowledge refinement.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing the components of the knowledge heath system according to an embodiment of the present invention, and FIG. 2 is the knowledge refinement I! Figure 3 shows an example of conclusion integration processing by the conclusion integration unit of the structure.
The figure shows an example of the conclusion addition process by the conclusion addition unit of the knowledge refinement mechanism, Figure 4 shows an example of the conclusion integration process by the condition value specification unit of the knowledge refinement mechanism, and Figure 5 shows the conditions of the knowledge refinement mechanism. Diagram 6 showing an example of conclusion addition processing by the addition section.
The figure shows an example of a decision tree generated from the decision table of knowledge representation, and FIG. 7 shows an example of knowledge representation. In the figure, 1... knowledge base system, 2... tabular knowledge editor, 3... knowledge base, 4... knowledge refinement mechanism. 5... Reasoning mechanism, 6... Case-based.

Claims (1)

[Claims] Using an input screen in the form of a table in which conditions and conclusions are arranged in rows and columns, respectively, condition values are specified in columns at intersections of the table, and a conclusion is derived from a combination of these condition values. In a knowledge base system that has a tabular knowledge editor that generates a decision tree of knowledge data, a knowledge base that stores the decision tree, and an inference mechanism that makes inferences using the decision tree, combinations of condition values input as examples are integrated. do,
A case organization method that creates and displays a classification tree in which conditions correspond to nodes and condition values correspond to branches from nodes, and a classification tree that compares the classification tree with a decision tree stored in a knowledge base. If there is a combination of condition values that does not exist in the decision tree, and the conclusion is the same as the conclusion that exists in the decision tree, extract the part where the decision tree corresponding to the conclusion and the classification tree have different branches. and a conclusion integration means for integrating the classification tree into a decision tree corresponding to the conclusion by integrating the condition values; and the classification tree has a combination of condition values that does not exist in the decision tree, and If this does not exist in the decision tree, it is considered as knowledge that has not been stored in the knowledge base, and a conclusion addition means is provided to add new knowledge to the decision tree, so that the combination of condition values in the classification tree is If they are included in a decision tree and their conclusions are different, search for the existence of a condition for which no condition value is specified, and if the condition value can be specified, specify the condition value to connect both buses. A condition value specifying means to be satisfied, and a condition adding means for adding a condition for distinguishing both buses when a condition value cannot be specified or a condition for which a condition value is not specified does not exist, A knowledge refinement processing method using examples, which is characterized by modifying and adding to decision trees by adding examples.