JPH0258166A - Knowledge search method - Google Patents

Abstract

PURPOSE:To easily constitute a retrieval system of a knowledge data base based on meaning contents of knowledge by preliminarily giving one or plural meaning structure data at the time of registering knowledge data to the knowledge data base. CONSTITUTION:One or plural meaning structure data 12 expressing representative meanings of contents of knowledge data 10 are preliminarily given to knowledge data 10 in the knowledge data base at the time of registering this data 10 to the knowledge data base. At the time of knowledge retrieval, an inputted question is analyzed and meaning structure data 12 corresponding to meaning contents of the question is extracted, and meaning structure data 12 given to respective knowledge data 10 in the knowledge data base are collated with meaning structure data 12 of the question one by one, and meaning structure data 12 whose collation results in success is selected and outputted. Thus, contents are efficiently retrieved with a high precision.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a method for searching knowledge data, and particularly to an apparatus for searching knowledge data regarding the semantic content of the knowledge data.

[Conventional technology]

Conventionally, in information retrieval systems related to documents, etc., keywords included in documents, etc., or logical combinations thereof (AND, ○■, etc.) have been used as search means. However, with this method, information about the semantic relationship between keywords is not reflected in the search key, so ambiguity remains, and search results often contain noise.

For example, when searching for information about news articles,
The keywords for the question “What are the trade liberalization measures that the United States has requested from Japan?” are “United States” and “Japan.”
"Trade liberalization policy" is used, but this is indistinguishable from the question "What is the content of the trade liberalization policy that Japan has requested from the United States?"

For example, in a content search for an online manual for programming R language such as LISP, the question 'C1: ``What is the function that converts a character into its corresponding numeric code?''
``What is the function that changes a numeric code to the corresponding letter?'' It is indistinguishable from the question 'J'.

An additional problem with information retrieval using keywords is that keywords may be mismatched, leading to incomplete searches. This is a mismatch caused by a discrepancy between the keywords assigned to each piece of knowledge data in the database of the information retrieval system and the keywords that the user comes up with when searching. To prevent this,
Although keywords that can be assigned to knowledge data are restricted, this creates a burden on users who have to search for keywords related to the information they want to search for each time they perform a search.

Attempts have been made to solve the above problems by using natural shield words.

For example, Information Processing Society of Japan Natural Language Processing Research Group Materials 58-
8 (Sugiyama et al., "Information Retrieval System IRISJ Based on Natural Language Understanding," November 22, 1986) accepts question sentences in natural i language such as Japanese, and searches for keywords contained in the question sentences and the keywords between them. A method of searching using relational information is discussed.

In this system, knowledge data is searched by the procedure described below.

First, an input question sentence is analyzed using a so-called natural i-word analysis technique (morphological analysis and syntactic-semantic analysis) to extract the semantic structure of the question sentence. Next, from this semantic structure, the conceptual keywords of the question and the search expression (and
and Or) is generated. Using this conceptual keyword search formula, the text data 9 of the information retrieval system using conventional keywords is searched and corresponds to the search formula! , distortion, obtain text data. At this stage, there is no difference from conventional information retrieval systems, except that the burden on the user is reduced because the question text becomes natural blind text, eliminating the need to create a search formula.

However, after this, the obtained text data is further subjected to natural i-word analysis to create a semantic structure corresponding to the text data. Then, by comparing the semantic structure of which question text with the semantic structure of this text data, a search result with less ambiguity of the content requested by the question text is obtained.

[Problem to be solved by the invention]

In the conventional system described above, a search formula using keywords is first generated and used to access the information retrieval system, but this is not an essential process.

First of all, data that is ultimately unnecessary is picked up as candidates during a keyword search, and extra data must be checked in the final semantic structure comparison, and the keyword search formula must be created in the first place. This creates extra work. In addition, natural five-word analysis of the searched text data is performed at the final stage, but this work would be unnecessary if it had been performed in advance at the stage of creating the database of the information retrieval system.

An object of the present invention is to solve the problems of the conventional system and technology, and to provide an efficient and accurate content retrieval means.

[Means to solve the problem]

In order to achieve the above object, the present invention uses the method shown below.

(1) One or more semantic structure data consisting of one or more semantic components expressing the meaning representative of the content is assigned to the knowledge data in the knowledge database in advance when the knowledge data is registered in the knowledge database. death.

When searching for knowledge, the input question text is analyzed and semantic structure data corresponding to the semantic content of the question text is extracted.
A knowledge retrieval method characterized in that the semantic structure data assigned to each piece of knowledge data in the knowledge database is compared one by one with the semantic structure data of the question sentence, and those that are successfully matched are selected and output.

(2) The semantic structure data includes words that make up the meaning, that is, semantics, case indicators that express the relationship between the semantics to express a complex meaning by linking the semantics, and the semantics and case indicators. A knowledge retrieval method characterized in that it consists of a tree structure composed of pairs of components as components.

(3) The above question analysis process includes morphological analysis that divides the question into words or morphemes, syntactic-semantic analysis that generates an expression using semantic structure data of the question based on the results of the morphological analysis, and syntactic-semantic analysis of the results of the syntactic-semantic analysis. A knowledge retrieval method characterized by comprising a semantic standardization process for generating standard semantic structure data by absorbing redundancy in semantic expressions present in the semantic structure data.

(4) The semantic standardization process includes a matching pattern section for matching the whole or part of the semantic structure data including redundancy in expression, a condition section for checking various conditions during matching, and a condition section for checking various conditions at the time of matching, and A knowledge retrieval method characterized by using a semantic standardization rule consisting of a result pattern part generated as a result.

(5) A knowledge retrieval method characterized in that when a search of a knowledge database using semantic structure data fails, a part of the semantic structure data of the question sentence is replaced with other related data and the comparison is attempted again.

[Effect]

The above (1) is the main method for realizing the knowledge retrieval method according to the present invention. This makes it possible to input a question regarding the meaning and content of knowledge data and search for knowledge corresponding to the question.

(2) shows the specific format of the semantic structure data to be added to each piece of knowledge data in the knowledge database when creating the target knowledge database. By expressing the semantic structure data of question sentences and knowledge data as a tree structure consisting of pairs of case indicators and semantics, it is possible to express the meaning of any question sentence or knowledge data, and the matching of both can be unified. And it is easily realized.

Methods (3) and (4) are used to deal with the diversity of input question sentences, and apply semantic standardization rules to the semantic structure data of question sentences to obtain a standard format of semantic expression. This is effective in reducing the required amount of semantic structure data to be added to the knowledge data in the knowledge database, lightening the burden of verification processing, and increasing the accuracy rate of answers to question sentences. Note that, as a matter of course, each knowledge data is given semantic structure data after the above semantic standardization process.

(6) provides a method of searching for other answer possibilities by partially changing the semantic structure data of the question text when matching fails. Similar to (3) and (4), this has the effect of increasing the precision rate of one answer.

〔Example〕

Hereinafter, the present invention will be explained in detail using Examples.

In the following explanation, LISP i is used as knowledge data.
Although knowledge of the language, particularly the Common LIS Pi language will be taken as an example, the present invention is not limited thereto, as can be easily understood.

FIG. 2 shows an example of the hardware configuration of a device that implements the knowledge retrieval method according to the present invention. Basically, it has the same configuration as a normal personal computer, workstation, or general-purpose computer. 1 is a device for inputting a question such as a keyboard; 2 is a display device such as a display or printer for displaying search results; 3 is a central processing unit of the computer;
4 is a main memory for storing programs and work data, and 5 is an external storage such as a magnetic disk for storing knowledge databases and the like.

FIG. 1 shows an example of the structure and contents of each knowledge data 10 in the knowledge database stored in the external storage 5. As shown in FIG. Common L I S P function ch
This is an example of knowledge data of ar-code and code-char. As shown in the figure, the name column 11 of knowledge data
contains the function name, and the semantic structure data column 12 stores semantic structure data representing each of the nine functions. Here, we include the semantic structure that corresponds to the realization of the function. The content column 13 contains detailed explanations regarding each function.

If we write the semantic structure data of each knowledge data in Figure 1 in ordinary sentences, we can use the function char to convert a character into a number.
-code) Convert r number to character” (function co
de-char). These semantic structures consist of pairs of case markers that indicate the semantic relationships between each word and other syntactic elements in a sentence, and semantic markers that express the meaning of the story itself, and consist of a tree structure that expresses the dependence relationships between them. ing. In the case of the function char -code, "object" and "end point" are case indicators, and "convert,""character," and "number" are word meaning indicators. In this case, the case marker ``object'' indicates that the ``character'' is the object case of the verb ``convert,'' and the ``terminal''
represents the final case of ``exchange'', that is, the destination as a result of conversion.

FIG. 3 shows a schematic processing procedure of the first embodiment of the knowledge retrieval method according to the present invention. First, a question text input from the keyboard 1 is received, and it is analyzed to extract semantic structure data corresponding to the content of the question.

Next, the semantic structure data of the question sentence is compared one by one with the semantic structure data of each knowledge data in the knowledge database, and if there is a successfully matched one, it is extracted and the result is output.

Next, each step in FIG. 3 will be explained in more detail.

The first step of inputting the question text is normally performed using an input program attached to an operating system program installed as standard in a general computer device as shown in FIG.

FIG. 4 shows a more detailed processing procedure for analyzing the question text in the second step of FIG.

First, a morphological analysis is performed to divide the character string that makes up the human-written question into words or morphemes.

In the case of Japanese, morphological analysis is generally performed in units of clauses. In the analysis, a word dictionary that registers the words or morphemes used within the system, and clause white grammar rules that indicate the connections between words or morphemes within a clause are used.

From a system perspective, there are two possible ways to enter a question: one is to use a mixture of kanji and kana, and the other is to manually enter the question using kana. In the case of kana input, there are differences in the morphological analysis results, such as increased ambiguities such as homonyms, but overall there is no essential difference in the processing content of morphological analysis for both inputs. The processing power method of morphological analysis for kana input sentences is described, for example, in Information Processing Society of Japan's 28th National Conference (1981), 4M-6 to 4M-9 (1237 to I24).
(page 4) etc.

[What is the function that converts characters to numbers? ” is shown in FIG. 5.

The second step in question sentence analysis is syntax-semantic analysis. Using the morphologically analyzed question as input, the semantic structure of the question is generated. Syntax-semantic analysis generally uses so-called grammatical rules and a case frame dictionary that describes the semantic relationships between words such as nouns and verbs. Regarding the specific processing method of syntactic-semantic analysis, see, for example, Oshima, Abe, Yuura, and Fufu, "Resolving Ambiguity in Kana-Kanji Conversion Using Case Grammar," Transactions of the Information Processing Society of Japan, Vol. 27, No. 7 (July 1986). ) 679 pages ~
It is written on page 687.

FIG. 6 shows an example of the syntactic and semantic analysis results for the question "What is the function that converts characters into numbers?". The resulting semantic structure has the same form as the semantic structure of the knowledge data in FIG.

The third step in question text analysis is semantic standardization processing.

The redundancy in the semantic expression contained in the semantic structure of the question sentence obtained as a result of the syntax-semantic analysis is absorbed, and the semantic structure data is converted into W4S. Examples of cases where standardization of meaning is required are shown below.

(1) Diversity of terminology In the example above, “What is the function that converts letters to numbers?”
``Convert'' is the same as ``change''.

You can also say, ``I'll fix it and replace it with Jr.''. Also,"
``number'' is ``code'' or ``number code'' or rn
You can also enter umberJ rcodeJ.

(2) Diversity of Structural Expression ``Character string part'' ``List part'' is the same as ``r substring'' or ``partial list.'' Also, ``reverse (a list)'' (2nd clause) means ``(
It has the same meaning as "to reverse the list" (1 clause).

If we create standardization rules for the above-mentioned redundancies in expressions and use them to absorb the diversity of expressions contained in the semantic structure, we can create a wider range of input expressions for question sentences. This is effective in reducing the burden on the user when creating a question, and also in reducing search omissions. Furthermore, the burden of checking with the semantic structure of subsequent knowledge data can be reduced.

An example of the syntax of the semantic standardization rule is shown in FIG.

As shown in the figure, a rule consists of a matching pattern part, a matching condition part, and a result pattern part. The matching pattern part is a pattern for matching the whole or part of the semantic structure data that is input to the semantic standardization process, the matching condition part is for checking various conditions related to matching, and the result pattern part is a pattern of semantic structure data generated as a result when matching is successful.

FIG. 8 shows a concrete example of the semantic standardization rules. (1) (2)
are examples given above in explaining the variety of terms and the variety of structural expressions, respectively. In these examples, the conditions are not checked at the time of verification, so the verification condition part of the rule is "rt" (t indicates the truth of the logical condition), that is, the condition is always true. Furthermore, in the third rule in FIG. 8(2), the symbol "?X" is included, which means a variable. It is possible to match any variable that appears in the match pattern section.

If the matching is successful in the matching pattern section, data in the semantic structure that matches the variable is assigned to that variable. The value assigned to a variable can be referenced by a variable with the same name in the matching condition part and result pattern. Especially in terms of result patterns.

The final result pattern is what the variable is replaced with its value. For example, if the sentence ``reverse a list'' is expressed in the semantic structure ((result reverse (object
If you match it with the sentence "List)))", ? The value of X becomes [list j, and the generated result pattern is "(reverse (target list))".

FIG. 8(3) is an example of a rule using the matching condition part.

The matching condition part of this rule means "The data type of variable ?X is a subtype of format".
This is expressed in a program written in the mmon LISP'i language. Common L I S P'i
In Japanese, data types are hierarchical, with subtypes such as list type, vector type, and string type below the single column type.

Therefore, the meaning of this rule is "?x
To? When there is an input sentence "Add X", ? If the value of X is a size type or its subtype, change the sentence to ``?Concatenate X.'' Input sentence [Add list to list] [Add vector to vector]
[Add strings to strings] is converted to "concatenate lists,""concatenatevectors," and "concatenate strings."

meaning! The processing procedure of the M normalization process is shown in FIGS. 9 and 10. The gist of the process is to match all partial structures including the entire target semantic structure data with the semantic standardization rules, and if the matching is successful, replace the partial structure with the result pattern of the successfully matched rule. This means doing as much of this work as possible. The expression "as much as possible" includes the fact that patterns of conversion results that are successfully matched are also processed.

FIG. 9 shows the overall flow of processing, and a variable called a matching flag is provided to replace the above meaning of "as much as possible" with actual processing.

At the beginning of the loop, the verification match flag is set to njQ and the program convert is called as the primary call, but if verification is successful even once during processing within the program convert, the verification match flag becomes t and the loop is repeated.

FIG. 1o shows the processing procedure of the program cannvert which actually performs matching and conversion between semantic structures and rules. This program uses recursive calls to apply rules to all substructures of the semantic structure.

Processing steps 2301, 2302, 2303° 2304
．． The part consisting of 2305.2306 performs matching with all semantic standardization rules for a partial structure of a semantic structure, and if the matching with the matching pattern part of each rule and the condition check in the matching condition part are successful, , indicates the process of replacing the substructure with the result pattern of the rule.

Steps 2307, 2308, 2309, 2310゜2
The part consisting of 311 describes control for applying the above processing to all substructures of the semantic structure.

As mentioned above, control is performed here using recursive calls. In step 2307, it is checked whether the current partial structure is the terminal leaf of the tree. If the partial structure is a terminal leaf, it is not possible to decompose it further to obtain a partial structure, so the current partial structure is used as a return value and the process returns to the called program (step 2311). If the substructure is not the final leaf, further processing is applied to that substructure. In step 2308, the program con is placed at the beginning of the list of substructures.
Apply vert recursively. In step 2309, the program con
Apply vert recursively. Here, the beginning of the list or the rest of the list is defined as i in the LISP language terminology.
This corresponds to the Car part and cdr part of the list. In step 2310.

Based on the results obtained in steps 2308 and 2309, a partial structure at the same level as the original partial structure is generated, and this is used as a return value to return to the calling program.

Next, the processing of collating semantic structure data, searching for knowledge data, and displaying the results in the third and fourth steps of FIG. 3 will be explained.

FIG. 11 shows the processing procedure for semantic structure matching and knowledge data retrieval.

First, a portion corresponding to the semantic structure of each knowledge data in the knowledge database is cut out from the output of the question analysis section, that is, the semantic structure of the question sentence that has been syntactically and semantically analyzed and standardized. Next, knowledge data is extracted one by one from the knowledge database and compared and verified against the semantic structure of the knowledge data.

If the verification is successful, the contents of the knowledge content (explanation) column are taken out and outputted through the display device 5. If the matching is not successful, the above process is repeated until there is no knowledge data in the knowledge database.

Here, it may be possible that the semantic structure of the question text and the semantic structure of a plurality of pieces of knowledge data match. If such a situation is possible, the process may be repeated until there is no more knowledge data in the knowledge database, without interrupting the process even if the matching is successful.

By the way, in the above embodiment, one question sentence is checked against all the knowledge data in the knowledge database (or until a successful check is found). When this happens, performance problems occur.

In such a case, in a typical file system, an index is created that focuses on some key in the search data, and this index is used to reduce the amount of search and collation processing. Furthermore, by creating this index into a tree structure and employing hashing technology, we are working to further speed up searches and matching within the index.

Common LI used in the explanation of this example
Regarding the S Pi language knowledge database, C
Many functions in ommonLISP require specific data types (
It is a function that performs processing on data belonging to a list, vector, character geodesic), and the functional expression of each function, that is, the semantic structure registered in knowledge data, usually includes words representing these data types. Therefore, if an index regarding this data type is provided 1 and the knowledge data of each function is placed under the index key of the data type, it is possible to speed up the verification search. Further, a verb in the functional expression of a function may be used as the index key instead of the data type, or both the data type and the verb may be used as the index key.

Furthermore, it is also possible to create a multi-stage index of semantic structure data using the discriminant net technique. Regarding the discriminative net technique, see H. Charniak.

C., Riesbeck, D.V., McDermoh
It is written in "Artificial Intelligence Ha',/Dobtsuku" by TT, Japan Computer Association, 1986.

Note that the technique for speeding up the matching search of knowledge databases described above can be applied in exactly the same way to speeding up the matching search of the above-mentioned semantic standardization rules.

By the way, in the first embodiment described above, it is not necessary to
There are cases where the CommonLISP function desired by the user cannot be obtained. Con+monL I S
There are many data types available in P, and they form a hierarchical structure, so if a generic function is provided in a higher-level data type, that data is Type-specific functions are often not provided. For example, there is no function that directly corresponds to the user's question, "What is the function that reproduces a vector?" However, copy -seq in the format which is a higher type of vector type.
There is a function called, which can satisfy the user's request, so in this case, you can use this function. Also, taking this situation into consideration, we originally created the [What is the function that compares strings?] '', but you may use a vector type, which is a higher-level data type for a single string, and ask, ``What is the function that compares vectors?'' However,
CommonLISP does not have a function that directly compares vectors, but there is a function that compares strings, which are subtypes of vectors, Stt"lng"+string<
+strjng/=, etc., is a function mismatch that compares a column that is a higher type of vector. In such cases, it would be helpful if you could provide both the functions corresponding to the upper and lower types.

A second embodiment for realizing such a request will be described below. The hardware configuration of this embodiment is shown in FIG. 2, similar to the first embodiment. Furthermore, the configuration of the knowledge database and the data format of the knowledge data are also the same as in the first embodiment (the format of the knowledge data is shown in FIG. 1). Furthermore, the general processing procedure of this embodiment is similar to that of the first embodiment, and is shown in FIG.

FIGS. 12 and 13 show differences between this embodiment and the first embodiment, that is, processing procedures for matching semantic structures and retrieving knowledge data.

The process shown in FIG. 13 is basically the same as the process shown in FIG. 11, and is a subroutine.

In step 601 of FIG. 12, first, a portion corresponding to the semantic structure of the function to be compared is extracted from the semantic structure of the question sentence. next,
As in the first embodiment, the semantic structure data is compared with the semantic structure data of the knowledge data (step 6o2). At this time, the collation program shown in FIG. 13 is called to execute the process. If the matching is successful, the process is over. If matching fails, check whether there is an upper data type or lower data type of the data reference expression included in the semantic structure data of the question, and if it exists, replace the data reference expression with the upper or lower data type for each. ,
Perform matching processing similar to the first processing. If there is a successfully verified item, the explanatory part of the knowledge content of the corresponding knowledge data is extracted and outputted through the output device 5 (steps 603 to 610).

For example, when the question ``What is the function that compares vectors?'' is input, there is no function that directly corresponds to the request of this question, so ``the part of vector J should be converted into one column of the upper data type'' or the lower data type Replace it with a "string" and perform a match, using the functions mismatch and qtring=,
Obtain explanations such as 5triB<, string/=, etc.

The inference method described in this example does not necessarily apply to CommonL
The knowledge database is not limited to the ISP language. It is also applicable when the content of the knowledge database is other knowledge data. For example, if the knowledge database is about various properties of living things, and the question is ``Can sparrows fly?'', knowledge of ``bird,'' which is a superordinate semantic category of ``sparrow,'' is used. teYE
It can also be used for things like answering S.

〔Effect of the invention〕

As explained above, according to the present invention, it is possible to easily construct a knowledge database search system based on the semantic content of knowledge. Furthermore, it is possible to respond to a variety of question expressions. Furthermore, after the Dan system is completed,
Since it is easy to add knowledge data, terminology, and meaning data to the system, the system can be easily expanded.

The present invention is effective in all systems having a function of searching not only knowledge of programming i-words but also knowledge in general based on its semantic content. Searches based on the content of literature and newspaper articles, searches based on the meaning of words listed in Japanese and foreign language dictionaries, searches related to computer hardware and knowledge of operating procedures for application programs, computer-based teaching devices, so-called CAI It is also effective in devices that explain knowledge in response to questions. This invention is extremely effective in systems such as consultation systems, where searching and presenting knowledge from a knowledge base is an important point.

[Brief explanation of the drawing]

FIG. 1 is a diagram showing the structure and content of knowledge data in the knowledge database in the first and second embodiments according to the present invention, FIG. 2 is a diagram also showing the hardware configuration, and FIG. 3 is also a diagram according to the present invention. It is a figure showing the outline processing procedure of the knowledge search method in . Figure 4 is a diagram showing a more detailed processing procedure for question text analysis.
FIG. 5 shows the morphological analysis results for the question "What is the function that converts characters into numbers?" and FIG. 6 shows the semantic structure data representing the syntactic-semantic analysis results. FIG. 7 shows the format of the semantic standardization rules, and FIG. 8 shows a specific example thereof. FIGS. 9 and 10 are flowcharts showing the processing procedure of the semantic standardization process. FIG. 11 is a flowchart showing the processing procedure for collating semantic structure data and searching for knowledge data in the first embodiment. FIGS. 12 and 13 are flowcharts showing processing procedures for collating semantic structure data and retrieving knowledge data in the second embodiment. DESCRIPTION OF SYMBOLS 1... Keyboard, 2... Display device, 3... Central processing unit, 4... Main memory, 5... External memory, 1o.
...Example of knowledge data, 11...Name of knowledge data, 12
...Semantic structure data representing the content of knowledge data, 13.
...Contents of knowledge data.

Claims (1)

[Scope of Claims] 1. In a knowledge retrieval device that extracts desired knowledge data from a knowledge database consisting of a large number of knowledge data such as documents, there is a device for expressing meaning representative of the content of the knowledge data in the knowledge database. One or more semantic structure data consisting of the above semantic constituent elements are assigned in advance when knowledge data is registered in the knowledge database, and when searching for knowledge, the input question text is analyzed and the question text is Extract the semantic structure data corresponding to the semantic content,
A knowledge retrieval method characterized in that the semantic structure data assigned to each piece of knowledge data in the knowledge database is compared one by one with the semantic structure data of the question sentence, and those that are successfully matched are selected and output. 2. The semantic structure data includes the words that make up the meaning, that is, the semantics, the case indicators that express the relationship between the semantics to express a complex meaning by linking the semantics, and the information between the semantics and the case indicators. 2. The knowledge retrieval method according to claim 1, wherein the knowledge retrieval method is comprised of a tree structure composed of pairs as constituent elements. 3. The question text analysis process includes three morphological analysis that divides the question text into words or morphemes, syntactic-semantic analysis that generates an expression using semantic structure data of the question text based on the morphological analysis results, and syntactic-semantic analysis results. 2. The knowledge retrieval method according to claim 1, further comprising a semantic standardization process for generating standard semantic structure data by absorbing redundancy in semantic expressions present in the semantic structure data. 4. The semantic standardization process consists of a matching pattern part for matching the whole or part of semantic structure data including redundancy in expression, a condition part for checking various conditions during matching, and a result when matching is successful. 4. The knowledge retrieval method according to claim 3, characterized in that a semantic standardization rule consisting of a result pattern part generated as a result pattern part is used. 5. When the search of the knowledge database using the semantic structure data fails, a part of the semantic structure data of the question text is replaced with other related data and the verification is attempted again. Knowledge retrieval method.