JPH05342258A - Natural language processing system - Google Patents

Abstract

PURPOSE:To provide a natural language processing system which can prevent the execution of unnecessary operation caused by responsing an erroneous meaning analytic result and can prevent an input which is unfair for the system. CONSTITUTION:The system selects/extracts corresponding message information from a table storing message information which should be displayed for a user based on word data of a part specified its failure by meaning analysis and information expressing the meaning analytic part at the time of natural language analysis, prepares a message by burying word data into message information and alarms it. Thus, the user can confirm the meaning analytic result before the execution of the operation so that the execution of the unnecessary operation can be prevented. Besides, even when the operation corresponding to the meaning analytic result is not allowed by the system, the user can know it before the operation so that the input which is unfair for the system can be prevented.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a natural language processing system that operates in accordance with the meaning of an input natural language.

[0002]

2. Description of the Related Art In a natural language processing system that operates in accordance with the meaning of an input natural language, the efficiency of the user's work is improved by notifying the user of the content and result of the semantic analysis of the natural language. Can be improved.

As a technique related to this, for example, as described in Japanese Patent Laid-Open No. 63-317879, when displaying the result of translating a sentence from a first language to a second language, a dependency is displayed. There is a technique that allows the user to understand the dependency analysis result by clearly indicating the analysis result using parentheses or the like.

Also, as described in Japanese Patent Laid-Open No. 63-103379, when converting the first language into the second language, not only the translation result but also the input first language or translation result. There is a technique of outputting information about the first language and the second language as additional information corresponding to the output of the second language.

Further, as described in Japanese Patent Application Laid-Open No. 2-120976, there is a technique for generating and displaying a confirmation sentence in order to confirm the meaning of the input natural language with respect to the input of the natural language.

[0006]

In the above-mentioned prior art, since the additional information such as the dependency analysis result is also output together when the operation result of the system is displayed, the input semantic analysis result of the natural language is erroneous. Even if they do, (unnecessary) actions (translations) are executed, and as a result, there is a problem that the work efficiency of the user is reduced.

Further, when the system does not allow the operation corresponding to the result of the semantic analysis of the input natural language, the system does not have a means for notifying the user of the result, so that the system is prevented from illegal input. There is a problem that you can not.

Further, in the technique disclosed in Japanese Patent Laid-Open No. 2-120976, it is premised that the semantic analysis, which is the basis of the confirmation sentence generation, obtains a correct result or forcibly obtains the analysis result. As a result, even if the semantic analysis is not performed sufficiently, it is not conscious of the insufficientness in the semantic analysis processing. As a result, the confirmation sentence is displayed, and the user who is requested to confirm does not know how to correct the input natural language sentence when updating it.

An object of the present invention is to provide a natural language processing system capable of preventing the execution of an unnecessary operation according to the result of incorrect semantic analysis.

Another object of the present invention is to provide a natural language processing system capable of preventing illegal input for the system.

[0011]

According to the present invention, there is provided output means for outputting a message including a natural language of a specified part regarding a part specified as an analysis failure by the semantic analysis of the input natural language. A natural language processing system that operates in response to a response input to the output message is provided.

Further, according to the present invention, means for detecting that the system does not allow the operation corresponding to the analysis result of the input natural language, and the meaning indicated by the input natural language in response to the detection result. A natural language processing system including means for notifying that the corresponding action is not permitted.

Further, according to the present invention, means for recognizing an analysis point where the semantic analysis of natural language may fail,
A means for outputting information about the cause of the failure of the semantic analysis regarding the analysis location, the analysis method of what kind of analysis the system performed for the cause, and the coping method that the user should take to prevent the failure of the semantic analysis. A natural language processing system is provided.

[0014]

The message including the natural language of the part specified by the semantic analysis of the input natural language may have the following contents, for example.

(1) A failure of the semantic analysis of natural language (2) A failure of the semantic analysis (3) What caused the failure of the semantic analysis (analysis (4) Representation of failure cause (4) Representation of what kind of analysis processing (analysis method) was performed by the system for the cause of failure of semantic analysis (5) User's action to prevent failure of semantic analysis (6) Semantic analysis was successful, but the system indicated that the action corresponding to the semantic analysis result was not acceptable (7) The action corresponding to the semantic analysis result If the system does not allow it, it indicates the action to be taken by the user. Before executing the operation corresponding to the result of the semantic analysis of the input natural language, output the messages such as (1) to (7) above. By doing so, the user Since the analysis failure can be confirmed and the semantic analysis result can be corrected if necessary, it is possible to prevent the unnecessary operation from being performed according to the incorrect semantic analysis result, and improve the work efficiency of the user. In addition, even if the system does not allow the operation corresponding to the semantic analysis result,
By referring to the output message, it is possible to know that before the operation. In addition, since a message regarding the corresponding treatment can be obtained, it is possible to prevent illegal input to the system.

[0016]

EXAMPLES Examples of the present invention will be described in detail below.

The present invention is applicable not only to Japanese but also to systems using natural languages such as English and German. Here, a database search system in Japanese (hereinafter simply referred to as a system) is used as an example. As described below. In the present embodiment, the user inputs to the computer mainly Japanese sentences instructing the search of the database, and generally (including exceptions) a search target and a search condition for the search target in Japan. It is a word sentence.

FIG. 1 shows an outline of processing of this embodiment.

First, the user is prompted to input Japanese (1). In this embodiment, Japanese is input from the keyboard. Japanese input 1 results in Japanese character string 1
It suffices that a can be input to the system, and therefore, as the Japanese input device, a handwritten character input device, a voice input device, etc. can be easily substituted in addition to the keyboard.

Next, the Japanese analysis processing unit 2 refers to the word dictionary 12 and the operation instruction target definition information storage table 13 defining the configuration information of the target database 16 to input the input Japanese character. Analyze column 1a. That is, the words (words, phrases, phrases, etc.) that make up the Japanese character string 1a are identified, the meaning (part of speech) of each word that makes up the Japanese character string 1a is identified, and the qualifying relations of the phrases are identified. An intermediate expression 2a representing the analysis result of the Japanese character string 1a is generated.

In this embodiment, as a message containing a natural language of a part specified by the semantic analysis of the input natural language,
Consider the following three types.

(1) Regarding the part where the semantic analysis fails,
What expresses the "cause of analysis failure", which is the cause.

(2) Cause of failure in semantic analysis ((1) above)
"Analysis method (analysis policy)" that indicates what kind of analysis the system performed.

(3) "Handling method" of what the user should do to prevent the failure of the semantic analysis
Which represents.

"Part where semantic analysis has failed" in (1) above
Regarding this, in the present embodiment, the following two points are considered as the analysis points where the analysis may fail.

(A) In the case where there are ambiguous meanings in the words constituting the input natural language and the meanings of the preceding and following words cannot be uniquely determined.

(B) A case where an unknown word which is not defined in the word dictionary 12 exists in the input natural language.

On the other hand, in the present embodiment, as an example of the case where the system does not allow the operation corresponding to the input semantic analysis result of the natural language, (c) the search target cannot be detected.
Consider the case where the analysis result cannot be converted into a search command.

In the present embodiment, the qualification of the analysis location where the analysis may possibly fail, or the certification that the system does not permit the operation corresponding to the result of the semantic analysis, is defined in the analysis processing program in a procedural manner. It will be realized.
That is, if a certain step is reached in the processing program, the analysis may have failed.
Alternatively, the system certifies that the analysis result is not acceptable by the system. Then, in that step, the word data and the flag defined corresponding to the analysis location (hereinafter,
It is called “analysis code (1401)”. ) And are stored in the analysis status data. In this method, since the certification of the analysis point is realized as a program, it is difficult to maintain, but it can be expected to increase the processing speed.

As another method of recognizing the analysis points where the analysis may fail, the analysis rules are declaratively defined independently of the processing program, and those analysis rules are applied in the actual analysis. There is a method of recognizing that the analysis may have failed if the application of a certain analysis rule is successful or if the application of any analysis rule is not successful. Although this method slightly slows down the processing speed, it is easy to perform data maintenance work such as addition, deletion, and correction of analysis rules.

As another method, conditions for failing the analysis are declaratively described in a table in advance,
In the analysis, there is a method in which those conditions are collated one by one, and when the conditions are satisfied, it is determined that the analysis corresponding to the conditions may have failed. This method also facilitates data maintenance work.

Next, the analysis result display unit 3 displays the analysis result on the display device based on the intermediate representation 2a. In the present embodiment, in order to make it easier for the user to understand the analysis result, an "intermediate expression expansion table" in which the intermediate expression 2a is expanded in a table format is displayed. Of course, other than this method, the display method of the analysis result to the user may be a display method in a tree structure format, a display method in a network format, a display method in a predicate logic format, or a display method in a natural language format. It can be easily replaced. Regarding the generation and display of the analysis result in the natural language format, JP-A-2-12097.
One method is to make it as described in Japanese Patent Publication No.

The user refers to the displayed analysis result,
Make sure the system parsed as you intended. If the analysis result is not intended by the user, that is, if the analysis of the natural language fails, the user may select "Cause of analysis failure" at the analysis point where analysis may fail and "Analysis method" at the analysis point. , And “handling method” for the analysis failure (collectively referred to as “natural language analysis status”) can be referred to as necessary (4). Of course, even if the analysis result is intended by the user, these natural language analysis situations may be referred to.

The message data table 15 is provided in advance with output data relating to the cause of the analysis failure, the analysis method (analysis policy) at the analysis location, and the coping method for the analysis failure in association with each analysis location where the analysis may fail. Stored in. As will be described later, the output message extracts the message information in the message data table 15 corresponding to the analysis code 1401 described in the analysis status data 14 and representing the analysis location where the analysis has failed, and extracts it in the variable portion of the message information. It is generated by substituting data.

Here, the "cause of analysis failure" is considered to be a direct element that performs an analysis different from the user's intention. In the present embodiment, (1) the words that make up the input sentence have polysemy, and (2) the unknown words that have not been registered in the dictionary appear in the input sentence.

"Analysis method (analysis policy)" represents a concrete method of what kind of analysis the system actually performs. This embodiment corresponds to an analysis method such as (1) interpreting as the meaning previously described in the word dictionary, and (2) ignoring and analyzing an unknown word not registered in the dictionary.

Further, the "handling method" means that, if it is assumed that the analysis fails due to the analysis location, the user performs the same analysis in the next and subsequent analyzes to prevent the same analysis from failing. It should be concrete treatment guidance. In the present embodiment, (1) the description order of the meanings in the word dictionary is changed, the meanings of the words are uniquely registered, or the word corresponding to the column name of the database and the word corresponding to the column value are described as a pair. Yes, and (2) registering an unknown word in the dictionary.

When the user gives an instruction to display the natural language analysis status, whether to display the information about the cause of the analysis failure (5) or the information about the analysis method at the analysis failure portion in that ((5) 6) Select whether to display information on how to deal with the analysis failure (7). If you have selected to display the natural language analysis status in any of steps 5, 6 and 7, generate an analysis status message.
In the display processing unit 8, the word data used in the analysis at the analysis location where the analysis may possibly fail and the analysis status data 14 storing the analysis code 1401 corresponding to the analysis location, and the natural language message character string to be displayed. With reference to the message data table 15 storing the information, a natural language character string representing the natural language analysis status is generated and displayed according to the display request content of the user. That is, the analysis status data 14
It is generated by extracting the message information in the message data table 15 corresponding to the analysis code 1401 described in 1. and substituting the data into the variable part of the message information. In the present embodiment, the display to the user is the screen display of the display device, but it can be easily replaced by voice output or the like.

In this embodiment, the natural language message data relating to the natural language analysis situation is stored in the message data table 15.
To be defined in advance. Therefore, it is possible to generate and display the natural language message regarding the natural language analysis status based on the word data stored in the analysis status data 14 and the message information defined in the message data table 15. Of course, as will be described later, there are bunsetsu dependency, search condition dependency, etc. in addition to these, as analysis points where analysis may fail. It is possible to store and display the information about the natural language analysis status based on the message information defined in the message data table 15.

Next, the user refers to the displayed analysis result and natural language message information and selects whether or not to correct the analysis result (9). In the "correction" here, an unknown word is newly registered in the word dictionary 12, or the word dictionary 12
It also includes a process of updating the description content regarding the word registered in. In the present embodiment, when the analysis result itself is modified, the analysis result modification processing unit 10 uses the intermediate expression expansion table displayed as the analysis result as it is as a modification editor. That is, the user specifies, updates,
Modify the corresponding data by deleting or adding. This can be easily realized by associating the displayed data with the data stored inside. The technique of modifying the analysis result is well known and is not the essence of the present invention, and thus it may not be necessary to further describe it in detail. Of course, as a modification method, in addition to this, a method of displaying a tree structure form, a network form, a predicate logical form and correcting it,
How to display the input Japanese sentence 1a and correct it,
It is also possible to easily substitute a method of displaying a sentence obtained by converting the analysis result into a Japanese sentence and correcting the sentence.

Next, in the command generation / execution processing unit 11, the intermediate representation 2a is converted into an operation instruction command language, and the converted operation instruction command is executed to the database 16,
Display the execution result.

As is apparent from FIG. 1, the Japanese analysis processing unit 2, the analysis result display processing unit 3, the analysis status message generation / display processing unit 8, the analysis result correction processing unit 10, and the command generation / execution processing unit 11 are The processing is shown, and the word dictionary 12, the operation instruction target definition information storage table 13, the analysis status data 1
4, message data table 15, database 16
Indicates a file (also called a table).

As described above, according to this embodiment, each function is divided into blocks (modularization), and the modularity is improved when it is realized by program (software) logic, and when it is realized by hardware logic. It is possible to improve the circuit modularity of each function and to make it into an LSI, which facilitates system design and maintenance.

The structure and elements of various files (tables) used in this embodiment will be described.

First, the structure of the word dictionary 12 will be described. As shown in FIG. 12, the word dictionary 12 includes “word stem (1
201) ”,“ regular word (1202) ”,“ item attribute (1
203) ”and“ semantic code (1204) ”.

The "word stem (1201)" is a word that expresses some concept by itself.

The "regular word (1202)" represents a natural language character string that has a one-to-one correspondence with a column name or column value in the database 16. By defining the regular word 1202, different language expressions corresponding to column names or column values at the same position in the database 16 (for example, “price” for a column name representing “sale price”,
With respect to (language expressions such as “price”), the language expressions can be accurately associated with the constituent elements of the database 16 by replacing the language expressions with regular words and analyzing the expressions.

The "item attribute (1203)" is a classification of the character in the target world of the search command, and is an "attribute item name" corresponding to the column name of the database 16 and a "column value" of the database 16. There are “attribute item data”, “logical instruction word” indicating a logical instruction relationship between the attribute item name and attribute item data, and “logical operation word” indicating a logical product or a logical sum. Further, in the Japanese analysis 2, when an unregistered word is detected in the word dictionary 12,
The item attribute of the word is “unknown word”.

The "semantic code (1204)" represents the position of the word in the target world of the search command. The columns constituting the database 16 and the meaning codes 1204 have a one-to-one correspondence. The correspondence between the position of each column forming the database 16 and the meaning code 1204 is defined in the operation instruction target definition information storage table 13 described below.

Next, the structure of the operation instruction target definition information storage table 13 shown in FIG. 13 will be described. The operation instruction target definition information storage table 13 is a table in which the configuration information of the database 16 is defined.
1) ”,“ semantic code (1302) ”,“ affiliation table (1303) ”, and“ default target designation item (1
304) ”.

The "attribute item (1301)" is an item defining a regular word character string corresponding to a column name forming the database 16.

"Semantic code (1302)" is the word dictionary 1
This is synonymous with “Semantic code (1204)” of 2. Here, a meaning code 1302 corresponding to each column of the database 16 is defined. It should be noted here that the first alphabet of the meaning code 1302 represents the table to which it belongs, and the meaning code 3
It means that the remaining number string of 02 represents the column position in the table to which it belongs. In this way, the meaning code 1302 represents the position information in the database 16, and in the ambiguous resolution processing 202 described later, the meaning of the word is uniquely determined based on this position information. In addition, these contents are “semantic code (12
04) ”.

A character string representing the name of the table to which each column of the database 16 belongs is defined in the "belonging table (1303)".

"Default target designation item (1304)"
Is a flag that defines the central column in each table of the database 16, and each column has an integer value of "1" or "0".
In the analysis of the input natural language sentence 1a, when the search target cannot be recognized, the column in which the default target designation item 1304 is “1” is regarded as the search target, and the omitted Japanese sentence of the search target is also accepted. Is possible.

Next, the structure of the analysis status data 14 will be described. The analysis status data 14 is a table that stores word data related to an analysis point where analysis may fail and an analysis code 1401 corresponding to the analysis point. In the present embodiment, as described above, there are three meanings: existence of polysemous words that cannot be unambiguously determined, existence of unknown words, and absence of search target.
Display the natural language analysis status regarding types to the user. However, the data required for notifying the user of the natural language analysis status differs depending on the analysis content of the analysis location. In the present embodiment, a table that stores data required for each analysis location where analysis may fail is defined. That is, as shown in FIGS. 33, 34, and 35,
A table constituting the analysis status data 14 is defined for each analysis location. A detailed description of the elements of these tables is given below.

Next, the structure of the message data table 15 shown in FIG. 32 will be described. The message data table 15 is a table that stores natural language message information relating to the natural language analysis situation at an analysis location where analysis may fail. The message data table 15 includes “analysis code (1401)”, “message mode (1501)”, and “message template (15)”.
02) ”. In this embodiment, the natural language message information to be displayed to the user is generated using the message data table 15 shown in FIG.

In the message data table 15,
The "analysis code (1401)" is a code for identifying the analysis failure location. "Message mode (150
1) ”means that the content of the message is the cause of the analysis failure,
It is a code for identifying which of the three methods, the analysis method and the coping method, corresponds. The "message template (1502)" is a template of natural language message information displayed to the user as the natural language analysis status, and includes a fixed character string portion that does not depend on the analysis content and a variable portion that changes depending on the analysis content (FIG. 32). And the symbols "@", "$", "%"
Part sandwiched between each). When a message is generated, the natural language message information is generated by embedding the data stored in the analysis situation data 14 in the variable part of this template. In addition, in FIG.
The character string corresponding to the variable part is defined in natural language, but this is to make the description of this embodiment easier to understand.
In actuality, a character string (such as a variable name in the program) that is convenient for executing the program is described, and the data to be substituted is easily identified.

In the description of the processing in FIG. 2 and subsequent figures, specific Japanese sentences representing search instructions will be used as necessary. Specific Japanese sentences used in this embodiment are shown in FIGS. 14, 15 and 16.
14 is a Japanese sentence including a polysemous word, that is, a word “Hitachi” having a plurality of meaning codes 1204, and the Japanese sentence in FIG. 15 is an unknown word “Highly paid”. Is a Japanese sentence that exists, and the Japanese sentence in FIG. 16 is a Japanese sentence in which the search target is not specified.

FIG. 2 shows the hardware configuration of this embodiment, where 30 is an input / output device, 40 is a processor,
Reference numeral 50 is a storage device. The storage device 50 includes the following storage units. Reference numeral 51 is a Japanese analysis processing unit storage area, 52 is an analysis result display processing unit storage area, 53 is an analysis status message generation / display processing unit storage area, 54 is an analysis result correction processing unit storage area, and 55 is command generation / execution processing. A copy storage area, 56 is a word dictionary storage area, 57 is an operation instruction target definition information storage table storage area, 58 is an analysis status data storage area, 59 is a message data storage area, 60 is a database storage area, and 61 is for each processing execution. Represents a working area, which is the memory of.

The storage device 50 stores various programs and data such as the word dictionary 12. The processor 40 is
The program stored in the storage device 50 is executed, and data is processed through the input / output device 30 as necessary.

FIG. 3 is a schematic PAD of the Japanese analysis processing unit 2.
It is a figure (Problem Analysis Diagram), and shows the process of analyzing the input Japanese sentence 1a and generating the intermediate expression 2a. They are term recognition processing 201 and ambiguous resolution processing 20.
2, modified expression recognition processing 203, and intermediate expression generation processing 204.

FIG. 4 is a PAD diagram of the term recognition process 201 in the Japanese analysis processing unit 2. The word dictionary 1 from the Japanese character string 1a written in the phrase segment input from the input device
FIG. 2 shows the processing up to storing in the “bunsetsu table (stored in the working area 51)” as shown in FIG. The phrase table is a table in which grammatical information of words forming an input Japanese sentence and correspondence information between those words and elements of the database are described, and these pieces of information are acquired from the word dictionary 12. Less than,
This process will be described with reference to the PAD diagram.

First, the character string variable INP for storing the phrase segmentation input Japanese sentence 1a is initialized (step 20).
101). The variable INP merely represents a character string variable and does not have any particular meaning. Next, the phrase segmentation input Japanese sentence 1a is substituted for the variable INP (step 20102).

Next, while the variable INP contains a clause, the following processing is repeated for each clause. (Step 2
0103). By referring to the word dictionary 12, it is determined whether or not the stem 1201 of the word dictionary 12 is included in the phrase (step 20104). If it is included, the character string corresponding to the stem 1201 of the phrase is extracted. The part is registered as the "word stem (1201)" of the phrase table, and the remaining character string part is registered as the "adjunct word (1205)" of the phrase table (step 20105).

Next, referring to the word dictionary 12, "regular word (1202)", "item attribute (1203)", and "semantic code (120) of the stem 1201 in the phrase table.
4) ”are registered (step 20106).

If the word stem 1201 is not included in step 20104, it is determined whether or not a numeric string exists in the phrase (step 20107). If it exists, it is registered as a numeric string in the phrase table (step 2010).
8) Then, the remaining character string is registered in the clause table as an adjunct word 1205 (step 20109). Step 201
If it does not exist in 07, the phrase is regarded as an “unknown word”, and “unknown word” is stored as the “item attribute (1203)” of the phrase table. As described above, in the present embodiment, since it is determined that the analysis may fail due to the existence of the unknown word, the integer “1” is stored as the analysis code 1401 in the analysis situation data 14, and The phrase number 1402 to be executed and the unknown word character string 1403 are stored (step 20110). In this embodiment, the analysis code “1” indicates that an unknown word was present in the term recognizing process 201.

Next, after step 20104, the variable IN
The phrase is deleted from P (step 20111).

The term recognizing process 201 authenticates the phrase segmentation Japanese sentences shown in FIGS. 14, 15 and 16 by referring to the word dictionary 12 shown in FIG. 12, and FIGS.
It stores in each clause table shown in FIG. Here, in the Japanese sentence of FIG. 15, the phrase “high paid” is not included in the stem 1201 of the word dictionary 12, and
Since it is not a numerical string, it is recognized as an unknown word. On the other hand, in the unknown word data storage table that stores word data relating to unknown words in the analysis situation data 14 shown in FIG. 33, the integer “1” is set as the analysis code 1401 and the corresponding clause number 14
The phrase number “2” of the phrase “high paid” is stored as 02, and the character string “high paid” is stored as the unknown word character string 1403.

As described above, in this embodiment, it is determined whether or not an unknown word is present by determining whether or not each word is registered in the word dictionary 12.

FIG. 5 is a PAD diagram of the ambiguous resolution processing 202 in the Japanese analysis processing unit 2. Term recognition process 201
This shows a process of uniquely determining the meaning of a word having a plurality of meanings in the stem 1201 recognized in (1). "Meaning" here means the words that make up Japanese 1a,
The correspondence relationship between the elements forming the database 16 is shown. Hereinafter, this processing will be described with reference to the PAD diagram.

While data is present in the clause table, the following processing is performed (step 20201). It is determined whether the word stem 1201 of the phrase has a polysemous meaning, that is, whether the meaning code 1204 of the word stem is not uniquely determined (step 20202). In the case of polysemy, it is determined whether or not the clause is accompanied by an adjunct "ga" or an adjunct representing a logical operation relation (step 20203). Here, as described above, the “logical operation word” is a natural word (“or”, “katsu”, “to”, etc.) representing a logical product or a logical sum.

When an adjunct word "ga" or an adjunct word representing a logical operation relation is accompanied in step 20203, it is determined whether or not the immediately following clause has a meaning code 1204 common to the meaning code 1204 of the clause (step 20204). ), If there is a common meaning code 1204, the common meaning code 1204 for the clauses that are not uniquely determined.
As the semantic code 1204 of both clauses (step 202)
05). If there are a plurality of sets of common meaning codes 1204, all sets are set to the meaning code 120 of both clauses.
Set to 4. That is, in this case, the meaning code 1204 is not uniquely determined at this point. Step two
After 0205, the process once exits the loop of step 20201, and the process is repeated from step 20201. This is because the meaning code 1204 of a certain stem 1201 is uniquely determined, so that the meaning code 1204 of the stem 1201 that could not be uniquely determined up to that point may be uniquely determined. In addition, the ambiguous meaning code 1204 can be resolved.

In step 20203, if the adjunct word "ga" or the adjunct word representing the logical operation relation is not involved, it is judged whether or not the clause is a adnominal modifier clause (step 2).
0206), the meaning code 1204 in the same table as the clause meaning code 1204 is
It is determined whether or not the clause immediately after has (step 202
07), in the case of having a clause that is not uniquely determined, the meaning code 1204 in the same table is set as the meaning code 1204 of both clauses (step 2020).
8). When there are a plurality of sets of meaning code 1204 in the same table, all the sets are set as the meaning code 1204 of both clauses. For the same reason as the process after step 20205 after step 20208, the process returns to step 20201 and the process is performed again from the beginning.

In step 20206, if it is not the adnominal modifier phrase, the stem 1201 of the phrase immediately after the phrase.
Is a logical operation word or a logical instruction word for adnominal modification (step 20209). Here, as described above, the “logical instruction word” is a natural language (“or more”, which is a logical word) indicating a logical instruction relationship between the attribute item name and the attribute item data.
"Below", "other than", etc.). If YES in step 20209, it is determined whether or not the clause immediately after the logical operation word or the logical directive has the meaning code 1204 common to the meaning code 1204 of the clause (step 2021).
0), if there is, the common meaning code 1204 for the clauses whose meaning code 1204 is not uniquely determined is the meaning code 1204 of both clauses (step 2020).
5). If there are a plurality of sets of common meaning codes 1204, all sets are set to the meaning code 1204 of both clauses.
And After that, the process returns to step 20201 again to perform the process from the beginning. If not in step 20210, it is determined whether or not the clause immediately after the logical directive has the meaning code 1204 in the same table as the meaning code 1204 of the clause (step 20211). For the clauses that have not been set, the meaning code 1204 in the same table is set as the meaning code 1204 of both clauses (step 20208). When there are a plurality of sets of meaning code 1204 in the same table, all the sets are set as the meaning code 1204 of both clauses. afterwards,
The process returns to step 20201 and the process is performed from the beginning.

As can be seen from the above description, the processing executed in the loop of step 20201 is the ambiguous disambiguation processing with a strong restriction defined in the connection relation between adjacent clauses. The meaning code 1204 of the stem 1201 that cannot be unambiguously determined by the processing so far is unambiguously determined by the ambiguous resolution processing in which the constraint is slightly weakened in the loop processing of the next step 20212.

While data is present in the clause table, the following processing is performed (step 20212). It is determined whether or not the stem 1201 of the phrase has a plurality of meaning codes 1204 (step 20213). If there is, the meaning code 1204 of the phrase and the meaning code 12 in the same table are determined.
It is determined whether or not the adjacent clause has 04 (step 2
0214), if it has, the meaning code 1204 in the same table for the clauses that are not uniquely determined is set as the meaning code 1204 of both clauses (step 20208).
When there are a plurality of sets of meaning code 1204 in the same table, all the sets are set as the meaning code 1204 of both clauses. After that, the process returns to step 20201 again to perform the process from the beginning.

For words that cannot be uniquely determined by the processing up to this point, the meaning code 1204 is determined based on the description order in the word dictionary 12. In the meaning code 1204 of each stem 1201 described in the word dictionary 12 used in this embodiment, the description order matches the priority order of meaning. The priority order of the meaning code 1204 is determined in advance by the dictionary creator based on the usage frequency of the meaning. Of course, it is easy for the user to change this priority order.

While data is present in the clause table, the following processing is performed (step 20215). It is determined whether or not the stem 1201 of the phrase has a plurality of meaning codes 1204 (step 20216). If the stem 1201 of the phrase has the meaning code 1204 currently held by the stem 1201 of the phrase, the word is described earlier in the word dictionary 12. Meaning code 120
4 is determined as the meaning code 1204 of the clause (step 20217). Then, since the analysis may fail by forcibly determining the meaning code 1204 in step 20217, the polysemous word data storage for storing the word data relating to the polysemous words in the analysis situation data 14 shown in FIG. In the table, the integer “2” is set as the analysis code 1401, the corresponding clause number is set as the corresponding clause number 1402, and the decision meaning code 1405 is set in step 20.
The meaning code 1204 determined in 217 is stored as the other candidate meaning code 1406, and the meaning code 1204 rejected in step 20217 is stored. In the present embodiment, the analysis code “2” indicates that the meaning code 1204 of the stem 1201 cannot be uniquely determined from the meanings of the preceding and subsequent stems 1201, and as a result, the meaning code 1204 is forcibly determined halfway.

14 and 1 by the ambiguous resolution processing 202.
5, the meaning code 1 of a word having a plurality of meaning codes 1204 included in each input Japanese sentence 1a in FIG.
204 is uniquely determined. The meaning code 120 of each stem 1201 shown in FIG. 20 is obtained by performing the ambiguity elimination process 202 on the phrase table of FIG. 17, which is the result of the term recognition process 201 for the input Japanese sentence 1a of FIG.
4 generates a clause table that is uniquely determined.

In the clause table of FIG. 17, the word stem "Hitachi" has four kinds of meaning codes P5, C1, P4 and C2. The phrase "Hitachi's" is adjunctly modified to the immediately following stem "person",
Moreover, since the stem "Hitachi" has the meaning code 1204 in the same table as the meaning code 1204 of "person", in step 20208, the meaning code P5 and P4 belonging to the same table as the meaning code P1 of the stem "person" Is left as the meaning code 1204 of the stem "Hitachi", and the remaining two meaning codes C1 and C2 are rejected. Then, in step 20217, the meaning code P5 of the stem "Hitachi",
Of P4, the meaning code P5 described earlier in the word dictionary 12 (that is, the clause table of FIG. 17) is determined as the meaning code 1204 of the stem “Hitachi”. Further, at step 20218, the integer “2” is set as the analysis code 1401, the integer “1” is set as the corresponding clause number 1402, and the polysemous word character string 1404 is set as the character in the polysemous word data storage table of the analysis situation data 14 shown in FIG. The string “Hitachi” is stored as the determined meaning code 1405, the meaning code character string P5, and the other candidate meaning code 1406 is stored as the meaning code P4 rejected in step 20217.

Since there is no word stem 1201 forming the input Japanese sentence 1a shown in FIG. 15 having a plurality of meaning codes 1204 as shown in the clause table of FIG. 18, the ambiguous resolution processing 202 of FIG. As shown in the clause table immediately after, the clause table does not change before and after the ambiguous resolution processing 202.

In the phrase table of FIG. 19 which is the result of the term recognition processing 201 of the input Japanese sentence 1a of FIG. 16, the stem "Kawasaki" has two kinds of meaning codes P4 and C2. In this case, since the condition of step 20214 is satisfied, the meaning code P4 of the word "live" is determined in step 20208.
The meaning code P4 in the same table as is determined as the meaning code 1204 of the stem "Kawasaki", and the remaining meaning code C2.
Reject. FIG. 22 shows a clause table representing the result of the ambiguous resolution processing 202 for the clause table of FIG.

As described above, in this embodiment, the condition for eliminating the ambiguous condition is procedurally described in the program (steps 20203, 20204, 2021 in FIG. 5).
4, 20206, 20207, 20209, 2021
0, 20211) and apply them in order to resolve the ambiguity. These ambiguous resolution processing is performed by the above-mentioned condition part and steps 20205 and 202 of FIG.
It can also be realized by describing a consequent part (action part) such as 08 separately from the processing program.

FIG. 6 is a PAD diagram of the qualified expression recognition process 203 in the Japanese analysis processing unit 2. It shows the processing until the modification relation of the clauses forming the input Japanese sentence 1a is recognized using the clause table information, and this processing will be described below with reference to the PAD diagram.

While the phrase table contains data that is not an unknown word, the following processing is performed (step 20301). It is determined whether or not the phrase is a logical directive word phrase (step 20302). If it is a logical directive word phrase, it is determined whether or not the preceding phrase is attribute item data (step 20303). In the case of attribute item data, the logical indicator corresponding to the logical instruction word is connected immediately before the attribute item data, and the two clauses are collectively determined as "D term" (step 20304). “D” in the D term is “Da
It represents the initials of "ta". Further, here, the “logical designator” is a symbol (<,>, ==,! =, Etc.) representing a logical designating relationship between the attribute item and the attribute item data, and is one-to-one with the logical designating word. Corresponding to. In this embodiment, a symbol representing a logical indicator is defined as the meaning code 1204 of the logical indicator and stored in the word dictionary 12. Accordingly, the correspondence between the logical designator and the logical designator can be identified by referring to the stem 1201 and its meaning code 1204. The correspondence between the logical indicator and the logical indicator may be defined in a table different from the word dictionary 12.

After step 20302, it is judged whether or not the relevant clause is a stem 1201 representing a logical product or a logical sum (step 20305). If yes, the item attributes of the stem 1201 before and after the relevant stem 1201. It is determined whether or not 1203 are the same (step 20306). If they are the same, a stem 1201 representing a logical product or a logical sum.
It is determined whether or not the item attributes 1203 of the clauses before and after are both attribute item names (step 20307). If the item attributes 1203 are attribute item names, the word stem 120 representing a logical product or a logical sum.
1 and the preceding and following clauses are collectively determined as “A term” (step 20308), and when both are not attribute item names,
It is determined as "D term" (step 20309). The "A" in the section A represents the acronym for Attribute.

After step 20301, the following processing is carried out while data that is not an unknown word exists in the clause table (step 20310). It is determined whether or not the clause is determined as the A term or the D term (step 203).
11) If not determined, it is determined whether or not the phrase is an attribute item name (step 20312). If the phrase is an attribute item name, the phrase is determined as "A term" (step 20313). ), If it is not an attribute item name, the clause is determined as “D term” (step 2031).
4).

FIG. 7 is a PAD diagram of the intermediate expression generation processing 204 in the Japanese analysis processing unit 2. The processing up to the generation of the intermediate representation 2a is shown, and will be described below with reference to this processing PAD diagram.

The following processing is performed for all the A terms and D terms determined in the qualified expression recognition processing 203 (step 20401). Same meaning code 1 in adjacent A and D terms
Those having 204 are determined as the minimum unit of the search condition (step 20402).

After step 20401, it is determined whether or not there is an undetermined D term as the minimum unit (step 2040).
3) If it exists, refer to the operation instruction target definition information storage table 13 and refer to the meaning code 1204 of the D term.
The item attribute character string (A term) corresponding to is supplemented and determined as the minimum unit of the search condition (step 20404). After step 20403, it is determined whether or not there is an undecided A term as the minimum unit (step 20405), and if there is, the A term is stored as a search target (step 2).
0406). If it does not exist, first, by referring to the operation instruction target definition information storage table 13, the A closest to the end of the sentence is displayed.
The attribute item name that is the "default target designation item (1304)" in the table to which the item belongs is recognized as the search target (step 20407). The default target designation item 1304 of the table to which the term A closest to the end of the sentence belongs is set as the search target for the following reason. In the Japanese sentence 1a instructing the database search, the search target is most likely to come to the end of the sentence. If the search target cannot be certified, it is considered that the search target is omitted. Further, the item that the user wants to search is not significantly different from item A at the end of the sentence. Therefore, it can be estimated that the search target exists in the table to which the A term at the end of the sentence belongs. Furthermore, when an item name other than the central item in the table is to be searched, it is usual not to omit the word corresponding to the item name. This is because, if omitted, it will often be a Japanese sentence that does not make sense. From the above, when the search target cannot be certified, the item (default target designation item 1304) that is the center of the table to which the A term closest to the end of the sentence belongs is certified as the search target.

Now, in step 20407, the search target is supplemented by the system. However, since the analysis may fail due to this processing, the analysis code 1 is stored in the search target data storage table of the analysis situation data shown in FIG.
As an 401, an integer "3", and a number obtained by adding 1 to the last bunsetsu number as the corresponding bunsetsu number 1402 (that means that the sentence is located at the end of the sentence of the Japanese sentence) is set as the end-of-sentence A term character string 1407 Insert sentence A term character string, insert search target 1
As 408, the default target designation item 1304 of the table to which the last sentence A belongs is stored. In the present embodiment, the analysis code “3” indicates that the search target cannot be recognized from the input Japanese sentence 1a.

After step 20405, the following processing is performed on the minimum unit of all search conditions (step 204).
09). It is determined whether or not the logical indicator is attached to the head of the item D (step 20410), and if it is attached, A
Immediately after the term, the D term is concatenated and determined as the “condition” in the search condition (step 20411). If not attached, the logical indicator “==” is connected immediately after the A term, and the D term is connected immediately after the A term to be the “condition” in the search condition (step 20412).

After step 20409, all "conditions"
Then, the following processing is performed (step 20413).
It is determined whether or not there is a stem representing a logical sum between the “condition” and the “condition” (step 20414), and if there is, a “or” is inserted between the “condition” and the “condition”. (Step 20415), if not present, "an
d "is inserted (step 20416).

After step 20413, all "conditions"
Then, the following processing is performed (step 20417). It is determined whether "or" exists in the search condition (step 2
0418), if it exists, it is determined whether or not the A terms of the “condition” before and after “or” have the same meaning code 1204 (step 20419).
Insert "{" before "condition" immediately before "r", and then "or"
"}" Is inserted immediately after "condition" immediately after (step 20420). Next, it is determined whether or not the logical indicators of the conditions before and after “or” are the same and other than “==” (step 20421). Connect the logical indicator of "condition" just before {},
The logical indicator of "condition" is replaced with "==" (step 20422). Next, it is judged whether or not the logical indicator in the step is negative "^ =" (step 2042).
3) If not, change "or" to "and" (step 20424).

The modified expression recognition processing 203 and the intermediate expression generation processing 204 generate the intermediate expression 2a of FIG. 23 from the clause table of FIG. 20 in which all the meaning codes 1204 are uniquely determined. Here, since the attribute item name corresponding to the stem "Hitachi" does not exist in the input Japanese sentence 1a, the attribute item corresponding to the meaning code P5 of "Hitachi" is referred to by referring to the operation instruction target definition information storage table 13. First name "workplace"
And is also certified as the minimum unit of the search condition 2001. Further, since the attribute item data corresponding to the stems “person” and “annual income” do not exist in the input Japanese sentence 1a, these are certified as the search target 2002.

Similarly, from the clause table in FIG.
To generate the intermediate representation 2a. The word stem “5 million” is supplemented with the attribute item name “annual income” as in the case of “Hitachi” described above, and is certified as the minimum unit of the search condition 2001. Further, the word stem “high paid” is an unknown word, so it is ignored in the analysis, and nothing is analyzed.

Similarly, from the clause table in FIG.
To generate the intermediate representation 2a. First, the stems "Kawasaki" and "live" (the regular word is "address"), the stems "hobby" and "golf"
Are collectively recognized as the minimum unit of the search condition 2001. In addition, the attribute item name “sex” corresponding to the word stem “male” is supplemented, and the minimum unit of the search condition 2001 is recognized. In addition, this Japanese sentence 1a has no undetermined A term as the minimum unit, that is, the search target 200
Since 2 cannot be certified, the operation instruction target definition information storage table 13 shown in FIG. 12 is referenced in step 20407, and the default target designation item “name” of the “personal table” to which the stem “hobby” to which the word end A belongs belongs. Search target 2
Certified as 002. Then, in step 20408, the integer “3” is set as the analysis code 1401 and the integer “6” is set as the corresponding clause number 1402 in the search target data storage table of the analysis situation data 14 shown in FIG.
End-of-sentence A character string "407" as end-of-sentence A character string 1407
The default target designation item “person name” is stored as the insertion search target 1408.

As described above, in this embodiment, when the natural language 1a is analyzed and the intermediate expression 2a is generated, the search target 200 is searched.
The fact that 2 cannot be recognized is confirmed by determining whether or not there is an A term that does not pair with the D term. Such conditions may be described in advance in a table or the like independent of the program as rules, and the same processing may be realized by collating them one by one in the analysis.

Here, if there is no search target, the default target designation item is automatically stored as the search target. However, as will be described later, if the search target does not exist, it cannot be searched because there is no search target. There is also a method of generating and displaying a message indicating "," and letting the user define a search target using a correction editor or the like. In this case, the search target entered by the user and the input Japanese sentence are stored in association with each other, and when a semantically similar Japanese sentence is input, the stored information is matched with the stored information. By taking, it is possible to automatically qualify the search target.

FIG. 8 shows an outline PA of the analysis result display processing section 3.
It is a D figure and consists of the following processes. First, the intermediate representation 2a
Is formed into a table format to generate an "intermediate expression expansion table" 301, and then an intermediate expression expansion table display processing for displaying the intermediate expression expansion table by the output means.

FIG. 9 is a PAD diagram of the intermediate expression expansion table generation processing, which is the processing until the intermediate expression expansion table in which the intermediate expression 2a is shaped into a table format is generated.
This processing will be described with reference to the PAD diagram.

The following processing is performed on the condition of the search condition 2001 in the intermediate representation 2a (step 3010).
1). First, the regular word character string 1202 of the clause representing the A term is registered as the "attribute item name (2003)" that constitutes the intermediate expression development table (step 30102). Next, the "logical designator (2
004) ”, the logical indicator corresponding to the item A is registered (step 30103). Then, as the "attribute item data (2005)" that constitutes the intermediate expression expansion table, the regular word character string 12 of the clause representing the D term corresponding to the A term.
02 is registered (step 30104).

Next, the condition exists in {}, and
It is determined whether or not it is the last condition in {} (step 3
0105), if it is the last condition, the integer “0” is registered as the “condition modification range (2007)” of the intermediate expression expansion table (step 30106). If it is not the last condition, the integer "1" is registered (step 301).
07).

After step 30105, it is determined whether or not a logical operator exists immediately after the condition (step 30
108), if present, the logical operator is "an
It is determined whether or not it is "d" (step 30109), and if yes, the "logical operator (200
6) ”is registered as the character string“ & ”. If not, the character string “!” Is registered. If it does not exist in step 30108, it is judged whether or not the condition immediately after is "{" (step 30112). If not, the character string "as the" logical operator (2006) "in the intermediate representation expansion table is determined. "E" is registered (step 30113).

By the intermediate expression development table generation processing 301, the intermediate expression development table shown in FIG. 26 is generated from the intermediate expressions 2a in FIG. Similarly, the intermediate expression development tables shown in FIGS. 27 and 28 are generated from the intermediate expressions 2a shown in FIGS.

FIG. 10 shows the intermediate representation expansion table display processing 3
It is a PAD diagram of No. 02, and is a process until a Japanese analysis result is displayed to a user based on an intermediate expression expansion table. Hereinafter, this process will be described according to the PAD diagram.

First, with respect to the condition of the search condition 2001 (step 30201), the data of the intermediate expression expansion table corresponding to the condition is displayed by the output means (step 30202). Search target 2 after step 30201
The A term representing 002 is output from the output means (step 3
0203).

By the intermediate expression development table display processing 302, based on the intermediate expression development tables shown in FIGS. 26, 27, and 28, FIGS. 29, 30, and 31, respectively.
The analysis result is displayed to the user as shown in.

FIG. 11 is a PAD diagram of the analysis status message generation / display process 8, which is a process in which natural language message information is generated according to the display designation from the user and is displayed to the user by the output means. Yes, the process will be described below with reference to the PAD diagram.

First, the "message mode (1501)" is displayed indicating whether the cause of analysis failure is displayed at the analysis point where there is a possibility of analysis failure, whether the analysis method at the analysis point is displayed, or how to deal with the analysis failure. A variable MM representing "" is initialized (step 801). MM is
It is just a string that represents a variable and has no special meaning. Next, the content of the display instruction from the user is discriminated (step 802), and if it is the display instruction of the cause of the analysis failure, the integer "1" is substituted for the variable MM representing the message mode 1501 (step 803). If it is a display instruction of the analysis method, the integer “2” is substituted. If it is the display instruction of the coping method, the integer "3" is substituted (step 804).

After step 802, the analysis status data 14
For each table that stores the word data in
The following processing is performed (step 806). While data is stored for each table, the following processing is performed (step 807).

First, referring to the message data 15,
The variable MM representing the message mode 1501 and the message template 1502 corresponding to the analysis code 1401 stored in the table are extracted (step 808).

Next, the extracted message template 1
While the symbol "@", "$", or "%" representing the variable portion is present in 502, the following processing is performed (step 809).

While the message template 1502 has a portion surrounded by the symbol "@" and the symbols "@", "$" and "%" are not present in the portion (step 810), the analysis is performed. Referring to the situation data 14, the symbol "@"
It is replaced with the data corresponding to the item represented by the part surrounded by (step 811).

While the message template 1502 has a portion surrounded by the symbol "$" and the symbols "@", "$", and "%" do not exist within the portion (step 812), the operation is performed. By referring to the instruction target definition information storage table 13, the meaning code 1 represented by the part enclosed by the symbol "@"
The attribute item name 1301 corresponding to 204 is replaced (step 813).

While the message template 1502 has a portion surrounded by the symbol "%" and the symbols "@", "$" and "%" do not exist within the portion (step 814), the operation is performed. By referring to the instruction target definition information storage table 13, the meaning code 1204 represented by the portion surrounded by "@"
Is replaced with the belonging table name 1303 corresponding to (step 815). As can be seen from the above explanation, the symbol "@",
The difference between “$” and “%” is due to the difference in the table referred to for extracting the character string data to be replaced and the difference in the extracted data itself.

Here, as a result of analysis of the input Japanese sentence 1a shown in FIG. 14, when the intermediate expression expansion table in the table format shown in FIG. 29 is displayed to the user, if it is different from the user's intention, or Although it may be just the interest of the user, in any case, consider a case where the user receives an instruction to display the analysis status.

When the display content instructed by the user is "cause of analysis failure", the message mode variable MM is set to "1" in step 803, and all tables in the analysis status data 14 are set in step 807. Reference is made to check whether or not there is a table in which data exists. In the case of the input Japanese 1a shown in FIG. 14, it is understood that the data is stored in the polysemous word data storage table as shown in FIG. Next, referring to the message data 15 shown in FIG. 32, the message mode 1501 is the value “1” of the message mode variable MM, and the analysis code 1
401 is the analysis code “2” of the polysemous word data storage table
The message template ““ @ polymorphic character string @ ”of the field that is“, ”was interpreted as“ $ @ decision meaning code @ $ ”of“% @ decision meaning code @% ”based on the description order of meaning. "Is extracted.

Next, the above-mentioned extracted template 150
In FIG. 2, there are portions (3 places) surrounded by the symbol “@”, and the symbol “@”, “$”, or “%” does not exist inside them. By referring to the polysemous word data storage table of the data 14, the data "Hitachi" and "P5", which are surrounded by "@" and correspond to the "polysemous word character string" and the "determination meaning code", are extracted, respectively. Replace with the parts of "@ polymorphic character string @" and "@ decision meaning code @". At this point, the template 1502 has interpreted ““ Hitachi ”as“ $ P5 $ ”of“% P5% ”based on the order of description of meaning. Similarly, since the symbols “$” and “%” are present, the operation instruction target definition information storage table 1 shown in FIG. 13 is obtained in step 813 and step 815, respectively.
3, the attribute item name character string and the belonging table character string corresponding to P5 are extracted, and "$ P5 $" and "% P5" are extracted.
Replace with "%". As a result, the template 1502 interprets ““ Hitachi ”as“ workplace ”in the“ personal table ”according to the priority of meaning. ”
Since the complete sentence has been generated, it is displayed to the user as shown in FIG.

Even when the display designation from the user is the "analysis method", the same process is performed, and the message template 1502, "The meaning of polysemous words ... Get. Since the template "1502" does not have the symbols "@", "$", or "%", the message as it is,
This is displayed to the user as shown in FIG.

Even when the display designation from the user is the "handling method", similar processing is performed to display it to the user as shown in FIG.

Similarly for the input Japanese sentence 1a shown in FIGS. 15 and 16, a message corresponding to the display designation from the user is generated and displayed. The input Japanese 1a in FIG. 15 is
There is an unknown word, "high paid". Regarding the natural language analysis situation of the Japanese sentence 1a, a message is generated by referring to the analysis situation data 14 and the message data 15, and according to the display instruction from the user, as shown in FIGS. 39, 40 and 41. Display the natural language analysis status to the user. When the input Japanese 1a shown in FIG. 16 cannot be identified as a search target, the display designation from the user regarding the analysis status of this Japanese sentence 1a is “analysis failure cause”, “analysis method”, and “handling method” In addition, analysis status data 14 and message data 15
Message is generated by referring to FIG.
2, displayed to the user as shown in FIGS. 43 and 44.

When the user inputs the Japanese sentence "Hitachi's person's annual income is" as shown in FIG. 14 in the sense of "the employee's annual income is that of Hitachi," as shown in FIG. If the analysis result is displayed, it will be regarded as the correct analysis result. However, if you enter "meaning that the address is Hitachi's annual income", you will recognize that the analysis result is different from your intention. Therefore,
By giving an instruction to display the cause of analysis failure,
The information shown in 6 is obtained. That is, "Hitachi" is a polysemous word, and "Hitachi" means "work" as "address".
Recognize that the analysis fails due to misinterpretation of polysemous words.

When the user wants to know how to determine the meaning of a polysemous word, the information shown in FIG. 37 is displayed by giving an instruction to display the analysis method. obtain. That is, it is understood that the meaning of "Hitachi" is determined from the meaning of the adjacent word "person", and if the meaning of "person" cannot be determined, the meaning is determined according to the description order of the dictionary.

Further, when the user wants to treat “Hitachi” as the meaning of “Hitachi city” and wonders what should be done, by giving an instruction to display a coping method, the information shown in FIG. Can be earned. The user selects one of these coping methods according to the situation and performs a processing operation, so that the system can be made to analyze the system more accurately from the next time onward.

The same can be said for the Japanese sentences shown in FIGS. 15 and 16 as in the case of FIG.

It should be understood that, in this embodiment, the system does not understand the user's intention in the analysis, so the analysis result displayed by the system is the user's intention. It is not possible to determine whether or not it matches Therefore, the cause of the analysis failure displayed by the system to the user is merely a candidate, and thus redundant information may be included. However, it is possible to narrow down the part where the analysis fails to some extent, and it is possible to notify the user of information about the analyzed part that causes the user's intention to be deviated for a certain percentage of Japanese sentences.

As described above, according to this embodiment, the input Japanese sentence 1a mainly instructing the search of the database 16 is input.
When the analysis result is displayed and the analysis result does not meet the user's intention, in the Japanese parsing 2 by the system, the user is Since information about the analysis method and how to deal with the analysis failure can be obtained, the user can understand what kind of analysis was performed at what analysis location, and as a result, the analysis did not match the user's intention. As a result, the analysis result can be corrected efficiently. From the next time onward, the user can input the natural language that the system can analyze exactly as intended by utilizing the natural language analysis information acquired in the previous analysis. Further, since the message data 15 is independent of the program, it is possible to easily change the message as needed and according to the proficiency level of the user.

The expansion and modification of the above embodiment will be described below.

(1) When any one of the cause of the analysis failure, the analysis method at the analysis failure location, and the method of coping with the analysis failure is displayed, and the display content includes a plurality of types of messages. , If the user specifies one of multiple messages and further instructs to display a message other than the currently displayed display content,
The user is notified of only the message information related to the analysis failure location corresponding to the message specified by the user.

It is assumed that the Japanese sentence 1a shown in FIG. 45 is input. In this input Japanese sentence 1a, the polysemous word “Hitachi” exists, and the search target 2002 is omitted. The analysis result shown in FIG. 46 is displayed by the Japanese analysis process 2 and the analysis result display process 3 described above. Here, when the user gives an instruction to display the cause of the analysis failure, the analysis status message generation / display process 8 displays the message shown in FIG. Here, of the two types of messages displayed in FIG. 47, if you want to know the analysis method for the first message, first select the first message, and then instruct to display the analysis method. By,
As shown in FIG. 48, only the analysis method message corresponding to the first content of the analysis failure cause message is displayed to the user. This can be easily realized by referring to the analysis code 1401 corresponding to the message selected by the user and the word data in the analysis situation data 14. As a result, only the information necessary for the user can be displayed, and the user can save the trouble of finding the relevant information.

Alternatively, when the cause of analysis failure, the analysis method, and the coping method are displayed at the same time, and when a certain message is selected for one display content, the message of the other display content corresponding to that is displayed in a different mode from the other. The same effect can be obtained by this method.

(2) In the displayed analysis result, when the user specifies a part different from his intention and then instructs to display an analysis status message, only the analysis status message concerning the part specified by the user is displayed. Display to the user. Further, when displaying the analysis result, the analysis result component corresponding to the analysis status message that can be displayed to the user is displayed in a manner different from the others. Also, instead of the analysis result, the input Japanese sentence itself may be displayed and the same operation may be performed.

In the displayed analysis result, the portion corresponding to the analysis failure portion is the clause number of each element constituting the analysis result and the "corresponding clause number (14
02) ”. However,
In this embodiment, the unknown word is not reflected in the analysis result, so that the Japanese sentence is used as the number of the analysis result component corresponding to the “corresponding phrase number (1402)” of the unknown word phrase in the analysis situation data 14. In, the phrase numbers of the phrases located before and after the unknown word are adopted. For the Japanese sentence 1a shown in FIG. 45, the analysis result is displayed as shown in FIG. In the Japanese sentence 1a of FIG. 45, the polysemous word “Hitachi” exists and the search target is omitted. Figure 5
In the polysemous word data storage table in the analysis situation data 14 shown in 0, since the "corresponding phrase number" of "Hitachi" is "1", it is the first phrase in the input Japanese sentence, as shown in FIG. "Hitachi," which is an analysis result component corresponding to "Hitachi's", is displayed in a manner different from the others. In the search target data storage table in the analysis situation data 14 shown in FIG. 51, the corresponding phrase number of the search target added at the end of the sentence is “3”, so as shown in FIG. "Person name" which is an analysis result constituent element corresponding to "Person name" which is the phrase (phrase inserted during analysis) is displayed in a manner different from the others.

By displaying FIG. 49, the user can understand that there is a possibility that some sort of analysis failure has occurred in the "Hitachi Ltd." and "person name" parts. Then, if the location matches an analysis location that does not match the user's intention, by selecting the location, the analysis status message regarding the location can be referenced. As a result, only the analysis status message relating to the analysis result element that the user wants to refer to can be displayed, and the processing efficiency of the user can be improved.

(3) When the result of the semantic analysis is not acceptable to the system, a message is output, but the action corresponding to the action performed corresponding to the message and the input natural language are stored and newly input. If the meaning of the stored natural language is equivalent to the previously stored meaning of the natural language, the stored operation is executed corresponding to the previously stored natural language.

For example, when the Japanese sentence of FIG. 16 is input, there is no search target, so the system is not allowed as it is. Therefore, the message shown in FIG. 52 is generated and displayed to inform that the system does not allow it. The user uses the modification editor to define the search target. At that time, the defined search target and the natural language shown in FIG. 16 are stored in the table as shown in FIG.

Next, when FIG. 16 is input again as a new input, it is found that the search target does not exist (step 20407), and then the natural language input by referring to the table of FIG. The matching with the natural language stored in the table is performed, and if there is a match, the search target “name” corresponding to the natural language is supplemented as the search target in the new input.

Here, the matching between the natural language and the natural language is used, but when storing the analysis result of the natural language together with the natural language when storing in the table, more flexible matching can be realized. For example, the analysis result of the Japanese sentence in FIG. 16 is stored as shown in FIG. Next, it is assumed that the natural language shown in FIG. 55 is input as a new input. FIG. 55
Since there is no search target for the sentence of, the system cannot accept it as it is. Therefore, after step 20407, the table of FIG. 54 is referred to. Since the database table referred to in Japanese in FIG. 16 is the same as the database table referred to in Japanese in FIG. 55 (personal table), the omitted search target is the one previously searched by the user. It can be estimated that it is the same as the defined one (“name” in FIG. 54), and the search target “name” is supplemented as the search target in a new input.

As described above, by storing the contents of the treatment once performed by the user and using it after the next time, the analysis reflecting the treatment of the user can be performed and the usability of the system is improved.

(4) In the above embodiment, the case where the invention is mainly applied to the information retrieval system has been described, but the invention is not limited to this and can be applied to other natural language analysis systems and natural language interfaces.

[0142]

According to the present invention, by outputting a message containing a natural language of a part specified by the input semantic analysis result of the natural language, the user can execute the operation corresponding to the semantic analysis result. In addition, since the accuracy of the semantic analysis result can be confirmed and the unnecessary operation according to the incorrect semantic analysis result can be prevented, the work efficiency of the user is improved.

Further, according to the present invention, when the system does not allow the action corresponding to the result of the semantic analysis, it is possible to know that fact before the action, and the message regarding the corresponding action can be obtained. It is possible to prevent illegal input for the system.

Further, according to the present invention, the analysis location where the semantic analysis of the natural language may fail is identified, and the cause of the analysis failure at the analysis location and the analysis method at the analysis location (according to the user's request) (Analysis policy) and information on how to deal with the analysis failure can be notified, so the user can
When the analysis result and the user's intention are inconsistent, it is easy to understand why the system performs such analysis. Also, by using these information, the user can efficiently correct the analysis result, and
From the next time onward, you can create Japanese sentences that the system can analyze accurately.

[Brief description of drawings]

FIG. 1 is a diagram illustrating a process outline of an embodiment.

FIG. 2 is a block diagram showing a hardware configuration of an embodiment.

FIG. 3 is a schematic PAD diagram of Japanese analysis processing.

FIG. 4 is a PAD diagram of term recognition processing.

FIG. 5 is a PAD diagram of ambiguous resolution processing.

FIG. 6 is a PAD diagram of qualified expression recognition processing.

FIG. 7 is a PAD diagram of an intermediate expression generation process.

FIG. 8 is a PAD diagram of analysis result display processing.

FIG. 9 is a PAD diagram of intermediate expression development table generation processing.

FIG. 10 is a PAD diagram of an intermediate expression development table display process.

FIG. 11: PAD of analysis status message generation / display processing
It is a figure.

FIG. 12 is a configuration diagram of an example of a word dictionary.

FIG. 13 is a configuration diagram of an example of an operation instruction target definition information storage table.

FIG. 14 is an example of an input Japanese sentence representing an operation instruction for information retrieval.

FIG. 15 is another example of an input Japanese sentence representing an operation instruction for information retrieval.

FIG. 16 is still another example of an input Japanese sentence representing an operation instruction for information retrieval.

FIG. 17 is a block diagram of an example of a clause table showing the recognition result of an input natural language sentence.

FIG. 18 is a configuration diagram of another example of a clause table showing the recognition result of an input natural language sentence.

FIG. 19 is a configuration diagram of still another example of the clause table showing the recognition result of the input natural language sentence.

FIG. 20 is a configuration diagram of an example of a clause table after the ambiguous resolution processing.

FIG. 21 is a configuration diagram of another example of the clause table after the ambiguous resolution processing.

[Fig. 22] Fig. 22 is a configuration diagram of still another example of the clause table after the ambiguous resolution processing.

FIG. 23 is an example of an intermediate representation.

FIG. 24 is another example of the intermediate representation.

FIG. 25 is still another example of the intermediate representation.

FIG. 26 is a configuration diagram of an example of an intermediate expression development table.

FIG. 27 is a configuration diagram of another example of the intermediate expression development table.

FIG. 28 is a configuration diagram of still another example of the intermediate expression development table.

FIG. 29 is an example of an analysis result display to the user.

FIG. 30 is another example of displaying the analysis result to the user.

FIG. 31 is still another example of displaying the analysis result to the user.

FIG. 32 is a configuration diagram of an example of a message data table.

FIG. 33 is a configuration diagram of an example of an unknown word data storage table that constitutes analysis situation data.

FIG. 34 is a configuration diagram of an example of a polysemous word data storage table that constitutes analysis situation data.

FIG. 35 is a configuration diagram of an example of a search target data storage table that constitutes analysis status data.

FIG. 36 is an example of a message display showing the cause of analysis failure of the input Japanese sentence 101.

FIG. 37 is a message display example showing an analysis method at an analysis failure portion of the input Japanese sentence 101.

FIG. 38 is a message display example showing a coping method for an analysis failure part of the input Japanese sentence 101.

FIG. 39 is an example of a message display showing the cause of analysis failure of the input Japanese sentence 101.

FIG. 40 is a message display example showing an analysis method at an analysis failure portion of the input Japanese sentence 101.

FIG. 41 is a message display example showing a method of coping with a failed analysis portion of the input Japanese sentence 101.

FIG. 42 is a message display example showing the cause of failure in analysis of the input Japanese sentence 101.

FIG. 43 is a message display example showing an analysis method at an analysis failure portion of the input Japanese sentence 101.

FIG. 44 is a message display example showing a coping method for an analysis failure portion of the input Japanese sentence 101.

FIG. 45 is still another example of an input Japanese sentence representing an operation instruction for information retrieval.

FIG. 46 is another example of displaying the analysis result to the user.

FIG. 47 is a message display example showing the cause of failure in analysis of the input Japanese sentence 104.

FIG. 48 is a message display example showing an analysis method at an analysis failure portion of the input Japanese sentence 104.

FIG. 49 is another example of displaying the analysis result to the user.

[Fig. 50] Fig. 50 is a configuration diagram of still another example of the polysemous word data storage table forming the analysis situation data.

[Fig. 51] Fig. 51 is a configuration diagram of still another example of the search target data storage table forming the analysis situation data.

FIG. 52 is a configuration diagram of another example of the message data table.

[Fig. 53] Fig. 53 is a configuration diagram of an example of a table that stores Japanese and search targets in association with each other.

[Fig. 54] Fig. 54 is a configuration diagram of an example of a table that stores Japanese analysis results and search targets in association with each other.

[Fig. 55] Fig. 55 is still another example of the input Japanese sentence representing an operation instruction for information retrieval.

[Explanation of symbols]

2: Japanese analysis processing unit, 3: Analysis result display processing unit, 8:
Analysis status message generation / display processing unit, 10: analysis result correction processing unit, 11: command generation / execution processing unit

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Satoshi Wakayama 5030 Totsuka-cho, Totsuka-ku, Yokohama-shi, Kanagawa Prefecture Hitachi Ltd. Software Development Division

Claims (10)

[Claims] 1. In a natural language processing system that operates in accordance with the meaning of an input natural language, regarding the part identified as poorly analyzed by the semantic analysis of the input natural language, the nature of the identified part A natural language processing system, comprising: output means for outputting a message including a word, and operating in response to a response input to the output message. 2. The natural language processing system according to claim 1, wherein said output means outputs said message in response to an output request from an input device. 3. The natural language processing system according to claim 1, wherein the output means outputs the input natural language and the message in association with each other. 4. The natural language processing system according to claim 3, wherein the output means outputs the natural language of the specified part included in each of the input natural language and the message in a mode different from the others. A natural language processing system characterized by: 5. The natural language processing system according to claim 1, wherein the specified portion is an unknown word not registered in the dictionary of the natural language processing system. 6. The natural language processing system according to claim 5, further comprising means for registering the unknown word in the dictionary in response to a request from an input device. 7. A natural language processing system that operates in accordance with the meaning of an input natural language, and detects that the operation corresponding to the analysis result of the input natural language is not allowed in the natural language processing system. A natural language processing system, comprising: means and, in response to the detection result, means for notifying that an operation corresponding to the meaning of the input natural language is not permitted. 8. The natural language processing system according to claim 7, wherein the informing means includes means for outputting information on a treatment corresponding to the detection result. 9. The natural language processing system according to claim 7, wherein the operation of the natural language processing system corresponding to the treatment performed in response to the notification and the input natural language are stored in association with each other. And a newly input natural language meaning is equivalent to the stored natural language meaning, the natural language processing system executes the stored operation corresponding to the stored natural language. . 10. A means for inputting a natural language consisting of at least one word representing an operation instruction, a means for analyzing the natural language, a means for converting the natural language analysis result into an operation instruction command language, and the operation. In a natural language processing system including a means for executing an instruction command and an output means, a means for recognizing an analysis location where the natural language analysis may fail, an analysis failure cause at the analysis location, and the analysis location A natural language processing system, comprising: an output method for notifying information about at least one of the analysis method and the method for coping with the failure in analysis.