JPH0322140A - Knowledge refinement device - Google Patents

Abstract

PURPOSE:To reduce time required for refining knowledge by analyzing and understanding knowledge data represented by natural language by using a dictionary, retrieving a knowledge base based on this result, and executing the confirmation of the rightness of the knowledge base and the supplement of insufficient knowledge. CONSTITUTION:A word, an idiom or a part of the word (for instance, the root of a word), etc., in an object area are registered in the dictionary 4 together with the kind of the word and relation to other words, and further, restriction to the order of the kind of the word is prescribed flexibly, and the analysis and the apprehension of the data represented by the natural language are executed by using these, and the knowledge base 8 is refined based on this result. Accordingly, the meaning of the data represented by the natural language and stored in the knowledge base 8 is understood. Thus, possibility to store the data wrong in viewpoint of its contents in the knowledge base 8 is reduced, and the necessity to confirm the rightness of the knowledge base 8 decreases, and the time required for refining the knowledge base 8 can be reduced.

Description

[Detailed Description of the Invention] [Field of Industrial Application] The present invention is a knowledge refinement device that is applied to correct misdescriptions of knowledge data and add missing knowledge data in a knowledge base that stores data of natural language expression descriptions. Regarding.

[Prior Art] Conventionally, the discovery of errors in this type of knowledge data and the filling of missing knowledge data have all been left to the hands of knowledge engineers.

[Problem to be solved by the invention] When a knowledge engineer understands the knowledge extracted from an expert and builds a knowledge base, there is a possibility that data with incorrect contents may be registered in the knowledge base. Therefore, it is necessary to confirm the validity of the knowledge base directly or by looking at the inference results, and it takes a lot of time to refine the knowledge base.

The present invention has been made in view of the above points, and it reduces the possibility of erroneous registration in a knowledge base storing data of natural language expression descriptions, reduces the need to confirm the validity of the knowledge base, and improves the knowledge base. An object of the present invention is to provide a knowledge refinement device that can reduce the time required for refinement.

[Means for Solving the Problems] In other words, the knowledge refinement device according to the present invention analyzes words, idioms, or part of words (for example, word stems) in a target domain based on the type of word, the relationship with other words, etc. Furthermore, constraints on the order of word types can be flexibly defined, data on natural language expressions can be analyzed and understood using these constraints, and the knowledge base can be refined based on the results.

[Operation] According to the above configuration, the meaning of the natural language expression data stored in the knowledge base is understood. This reduces the possibility of registering data with incorrect content in the knowledge base, reduces the need to confirm the validity of the knowledge base, and
Reduces the time required to refine the knowledge base.

[Example] Hereinafter, one embodiment of the present invention will be described with reference to the drawings.

FIG. 1 shows an embodiment of the present invention. The knowledge refinement device includes a word extraction device 1, a syntax analysis device 2, a knowledge data generation device 3,
It is composed of a first memory 4, a second memory 5, a third memory 6, a working memory 7, a knowledge base 8, an input device 9, and an output device 10.

FIGS. 2, 3, and 4 show configuration diagrams of a first memory 4, a second memory 5, and a third memory 6 as one embodiment. Further, FIGS. 5 and 6 show configuration diagrams of the working memory 7 and the knowledge base 8 as one embodiment.

In the dictionary stored in the first memory 4, word data 40 is described in the format shown in FIG. The specific contents of the description will be explained below.

The word name section 41 describes a word. The synonym name section 42 is
Write a synonym for the word. The antonym name section 43 describes the antonym of the word. The higher-level word name section 44 includes a word corresponding to a physical object that physically includes the word if the word means a physical object, or a word corresponding to a physical object that physically includes the word, even if the word means an abstract concept. For example, write a word that corresponds to an abstract concept that semantically includes this concept. If the meaning of the word is a physical thing, the lower word name section 45 contains the word corresponding to the physical thing that this word physically includes.
If the meaning of the word is an abstract concept, write the word corresponding to the abstract concept that this word semantically includes, along with its word category name. The rule name section 46 describes the rule name to be evaluated when the word appears. The word category name section 47 describes the word category name of the word.

Template node 5 stored in second memory 5
0 is written in the format shown in FIG. The specific contents of the description will be explained below.

The node name section 51 describes the node name of the template node. Usually, word category names are used for this. The upstream node name section 52 describes the node name of the template node corresponding to the word category name of the word that appears before the word that has the word category name corresponding to the node name of the template node as an attribute. The downstream node name section 53 describes the node name of the template node corresponding to the word category name of the word that appears after the word that has the word category name corresponding to the node name of the template node as an attribute.

The rules stored in the third memory 6 are divided into four groups called rule sets shown in FIG. An example of the rules stored in each rule set is shown below.

First Rule Set 61 Φ Nore Nore A Sentences with conjunction-like words violate the word order regulations constrained by template nodes before and after the conjunction-like word, but this is tacitly accepted.

●Rule B: Quantity words must appear at the beginning of a sentence, or after the subject and before adverbs, verbs, and adjectives.

Second rule set 62 ●Rule C Words used as conjunctions such as “and”, “or”, etc.
Words that are considered to be omitted from the natural language expression data are complemented by referring to the word category names before and after the words.

・Rule D: Two words are originally connected by a conjunction, and a word that can be seen as a simplification (for example, 1 entrance/exit)
may be considered "exits and inlets") are returned to their original unsimplified representations.

●Rule E “Word 1 + negative word” is replaced with an antonym of word 1.

Third rule set 63 ●Rule F The word of word category name 1 must have word 1 or word 2 as the word located in template node name 1.

Conflict Nore Nore G A word with word category name 1 must not have a word with word category name 2 as a word located in template node name 1.

Fourth rule set 64 ◆Rule H Data containing word 1 (for example, "or")' has no meaning in natural language even if the longest word string data that is located before and after it and satisfies the word order template is exchanged. is no different.

●Rules! Quantitative words have the same meaning in natural language whether they appear at the beginning of a sentence, or after the subject but before adverbs, verbs, or adjectives.

Next, a case will be explained in which the processing procedure of the embodiment of the present invention is applied to a production rule using terminology used in a PWR plant as a specific example.

It is assumed that the knowledge base 8 has already stored production rules "Rule 1" and "Rule 2" consisting of a condition part and a consequence part of natural language expression (see FIG. 7). At this time, in the embodiment of the present invention, processing is performed in the following steps.

When the knowledge refinement device is activated, (1) the knowledge data generation device 3 generates natural language expression data stored in the condition section 82 of the first knowledge data 80 among the knowledge data 80 stored in the knowledge base 8; The “rapid increase in pressurizer pressure and temperature” is extracted and stored in the natural language data section 71 of the working memory 7 (No. 8
(see figure).

(2) The word extracting device 1 sequentially extracts words from the word name section 41 stored in the first memory 4 (see FIG. 9), and these words are stored in the natural language data section 71 of the working memory 7. Evaluate whether it is included in the data "rapid increase in pressurizer pressure and temperature". If it is included, this word is stored at the beginning of the keyword section 72 of the working memory 7.

If not included, extract the synonym stored in the synonym name field 42 of the word data 40 and evaluate whether it is included in the data stored in the natural language data field 71 of the working memory 7. do. If this synonym is included, the word (the word stored in the word name field 41 of the word data 40 in which this synonym was registered in the synonym name field 42) is stored in the keyword field of the working memory 7. Stored at the beginning of 72. In this case, since the word "sudden" includes the synonym "sudden", the working memory becomes as shown in FIG. 10.

(3) If the word extraction device 1 is unable to extract any word through the process in (2), it jumps to (6). If even one can be extracted, process (4) is performed. In this case, (4
] will be processed.

(4) The syntax analysis device 2 first extracts the word "temperature" in the data "pressurizer pressure and sudden rise in temperature" stored in the natural language data section 71 of the working memory 7 in the process of section (2). , “and”, “rise”, “suddenly”,
The order of words stored in the keyword section 72 of the working memory 7 is changed in accordance with the order in which "pressure" and "pressurizer" appear. (See Figure 11). Next, the words “pressurizer”, “pressure”, “and”, “temperature”, and “sudden” stored in the keyword section 72 of the working memory 7 are
, for "rise", the rule name part 4 of the corresponding word data 4o stored in the first memory 4 is sequentially changed.
6 and the second ruleset 62 (
If the rule name of a rule stored in (see FIG. 12) is written, the rule is extracted and fired. In this case, the word data "and" is included. The rule "Rule C" is fired, and the word is completed as shown in FIG.

Furthermore, for the words stored in the working section 72 of the working memory 7, the word category name section 47 of the corresponding word data 40 stored in the first memory 4 is checked, and Evaluate whether the order can be realized by the link information described in the template node 5o stored in the second memory 5 (see FIG. 14).
.

During this evaluation, the first rule set 6 of the third memory 6
If there is a word to which a rule stored in 1 is applied, that rule is taken into consideration. If it is possible, the syntax analysis device 2 outputs the analysis result from the output device 10 and performs the process (5). If it is not feasible, jump to (6). In this case, the rule "Rule A" is considered for the word "and" and determined to be feasible.

<<5>> The syntactic analysis device 2 determines whether there is a word to which the rule stored in the third rule set 63 of the third memory 6 is applied to the word stored in the keyword section 72 of the working memory 7. If so, evaluate the rule. If there is no word that does not satisfy the rule, jump to (8). If there is even one word that does not satisfy the rules, the process (6) is performed. In this case, as a specific example of "Rule F" and "Rule G", the word with the word category name "Measurement" is treated as a word belonging to the template node name "Verb".
It must have words such as "fluctuation", "increase", "decrease", "normal", etc. The word “and” word category name “Measurement amount” is the template node name r verb JI: The word to which it belongs is “
Must not have words such as "open" or "closed." ° is evaluated, but since both are satisfied, we will jump to (8).

(6) The output device 1o outputs that the word order and words used in the natural language expression data stored in the condition part of the knowledge data are incorrect, and inputs the data from the input device 9. Outputs a message to correct it.

(7) When data is input from the input device 9, the knowledge data generation device 3 converts the data into the corresponding knowledge data 8 of the knowledge base.
After storing in the condition section 82 of o and the natural language data section 71 of the working memory 7, and clearing the keyword section 72 of the working memory 7, the process jumps to (2).

(8) The knowledge data generation device 3 generates the word string data stored in the keyword section 72 of the working memory 7 or some words in the word string data stored in the keyword section 72 of the working memory 7. Considering the rules stored in the third rule set 63 of the third memory 6, the word stored in the upper word name section 44 of the word data 40 or the lower word name section 4
The knowledge data 8o of the natural language expression description, which contains the word string data replaced with the words stored in 5 in this order in the consequent section 83, is stored in the fourth rule set 6 of the third memory 6.
All knowledge bases 8 are searched taking into account the rules stored in 4.

If knowledge data of such a natural language expression description exists, the searched knowledge data 8o is created using the data stored in the natural language data section 71 of the working memory 7 as a consequent section.
New knowledge data having the data stored in the consequent section 83 in the condition section is generated, and this is outputted from the output device 10, stored at the end of the knowledge base 8, and the working memory 7 is cleared. In this case, the conversion of the word "rise" to its higher-ranking word "fluctuation" was recognized as a result of the evaluation of "Rule F" and "Rule G", and furthermore, the rule "Rule H" was considered for the word "and". As a result, "rapid fluctuations in pressurizer temperature and pressure" was searched, and knowledge base 8
It is updated as shown in Figure 5.

(9) Each device performs the processes (1) to (8) regarding the second and subsequent knowledge data 80 among the knowledge data stored in the knowledge base 8. All knowledge data 8
When the above processing is completed for 0, the knowledge refinement device stops.

[Effects of the Invention] As described above, according to the present invention, words in the target area,
Idioms or parts of words (e.g. stems) can be registered in a dictionary along with word types and relationships with other words, etc., and constraints regarding the order of word types can be flexibly defined and input using these. analysis of natural language expression knowledge data,
By understanding the information, searching the knowledge base based on the results, confirming the validity of the knowledge data, and adding missing knowledge, the possibility of registering incorrect data in the knowledge base is reduced. This reduces the need to confirm the validity of the knowledge base and reduces the time required to refine the knowledge base.

[Brief explanation of the drawing]

The drawings are diagrams for explaining one embodiment of the present invention; FIG. 1 is a block diagram showing the configuration of the knowledge refinement device, FIG. 2 is a diagram showing the configuration of the first memory, and FIG. 3 is a block diagram showing the configuration of the knowledge refinement device. A diagram showing the configuration of the second memory, FIG. 4 is a diagram showing the configuration of the third memory,
FIG. 5 is a diagram showing the structure of the working memory, FIG. 6 is a diagram showing the structure of the knowledge base, and FIGS. 7 to 15 are diagrams for explaining the processing procedure of the same embodiment. 1...Word extraction device, 2...Syntax analysis device, 3...
・Knowledge data generation device, 4...first memory, 5...
- Second memory, 6... Third memory, 7... Working memory, 8... Knowledge base, 9... Input device, 10... Output device, 40... Word data, 45・
...Lower word name part, 50...Template node, 8
0...Knowledge day evening.

Claims (1)

[Claims] In a device having a memory for storing data described in natural language expression, a first memory for storing a dictionary in which important words in a target area are registered, and a template node for flexibly specifying the order of words. a second memory for storing words, rules for allowing exceptional situations such as word order regulations, rules for completing, replacing, and changing the order of words, and rules for restricting word substitutions. a third memory for storing rules or rules for flexibly searching already existing knowledge data;
A knowledge base for storing knowledge data that describes information in the target domain using natural language expressions, and words that are extracted from the data that are described using natural language expressions.
a word extraction device that replaces words with synonyms as necessary;
Check the order of words extracted by this word extraction device.
A syntactic analysis device that corrects and replaces words with the same meaning and completes omitted words, and collates the output of this syntactic analysis device with knowledge data of natural language expressions stored in the knowledge base, It is characterized by comprising a knowledge data generation device for generating missing knowledge data, an input device for inputting from the outside using natural language, etc., and an output device for displaying the generated knowledge data. A knowledge refinement device.