JP4579400B2 - NATURAL LANGUAGE RESPONSE DEVICE AND METHOD, AND RECORDING MEDIUM CONTAINING NATURAL LANGUAGE RESPONSE PROGRAM - Google Patents

Description

[0001]
[Technical field to which the invention belongs]
The present invention relates to a natural language response device and method for asking a question in natural language for knowledge data obtained by extracting information stored in a character string expressed in a natural language and storing the information, and acquiring the stored information About.
[0002]
[Prior art]
In conventional morphological analysis, the meaning of a vocabulary is defined in advance using probabilities, frequency of use, set theory, and tree structure, and the meaning of a sentence is determined based on a logical value based on a combination of vocabularies and an area based on set theory. In addition, the computer operation by words stores a specific word, registers an operation corresponding to the word, a word, and a sentence, and executes the operation by a combination of logic circuits.
[0003]
Grammar analysis techniques are applied in context analysis of sentences. In this method, the grammar is classified into a fixed pattern, and the meaning is determined by the semantic connection condition of the vocabulary that matches the pattern.
[0004]
However, the meaning of the vocabulary that makes up the sentence is determined by the relationship between the vocabularies,
As a result, the grammar needs to be constructed, and fixedly defining the meaning and grammar of the vocabulary limits these relationships. The analysis of sentences using vocabulary or grammar increases the number of fixed patterns, and the analysis of sentences using limited patterns involves approximation by pattern recognition, making it difficult to grasp accurate information.
[0005]
Semantic analysis as described above
Japanese Patent Laid-Open No. 4-191338 “Question Answering Method and Device”,
Japanese Laid-Open Patent Publication No. 5-230701 “Information Extraction Device”,
Japanese Laid-Open Patent Publication No. 7-0778355 “Natural Language Semantic Analysis Processing Device”,
Japanese Laid-Open Patent Publication No. 10-304583 “Natural Language Processing Device and Method” is available.
[0006]
[Problems to be solved by the invention]
The meaning of sentences expressed in natural language is determined by the combination of the vocabularies that compose the sentences, and the meaning of each vocabulary that composes the sentences as a result of the influence of one vocabulary on the other vocabulary. And the meaning of the whole sentence needs to be determined. Further, since the sentence includes the meaning indicated by the whole sentence, it is necessary to use a method of determining the image expressed by the vocabulary constituting the sentence and determining the meaning of each vocabulary constituting the sentence.
[0007]
Therefore, in order to realize a morphological analysis that does not require a grammatical pattern for syntactic analysis, the present invention autonomously determines the influence of vocabularies constituting a sentence on the influence of other vocabularies, and determines a combination of vocabularies constituting the sentence. By changing the vocabulary (hereinafter, this is defined as the influence on other vocabularies), the vocabulary that constitutes the sentence is determined, the meaning of the combination of the vocabulary is estimated, and the meaning that gives an overview of the whole sentence It is an object to obtain information stored in a natural language by estimating and extracting information contained in a sentence and storing the extracted information in a natural language.
[0008]
[Means for Solving the Problems]
In order to solve the above-mentioned object, the natural language response device and method according to the present invention uses a minimum unit constituting a sentence as a definition word, and the definition word is a kanji, hiragana, katakana, alphabet, numeral that can specify the meaning. , Symbols and combinations of character strings. How to use a definition word, extraction method, operation method, answer method to input a sentence in natural language and describe the processing procedure that describes the effect of all definition words that can be constituent elements of the sentence on other definition words in the description language Define as necessary.
Definition words that make up a sentence by defining the procedure described in the description language as part of the definition word data, where human beings have determined differences in meaning and usage as general common sense. By executing the procedure defined for each definition word, the definition word itself judges the influence on itself or other definition words, and the combination structure of the definition words constituting the sentence is determined. Change, create the optimal connection of the definition words that compose the sentence,
[0009]
Even when extracting information contained in sentences or when answering questions, search the stored information by executing the procedure in the descriptive language defined in the operation method of the definition word and the answer method, which is caused by the combination of phrases By creating conditions, it is possible to accurately create answers according to individual situations.
[0010]
When the definition word information is data defining the definition word and the procedure describing the above method for the definition word and the linguistic attribute of the definition word, the sentence can be defined as the connection state of the definition words. The definition word information stores data that can operate as a small program, and the procedure of the morphological analysis process can be determined by the combination of the definition word information constituting the sentence. This means that a robot can be arranged on a production line such as an assembly factory to produce one product, but other products can be produced by changing the function and configuration of the robot. This is because each robot uses only a limited function, but the entire production line can produce any product by balancing and configuring the individual robots. The procedure described in the definition word information corresponds to this robot, the procedure has a limited motion definition, and the sentence formed by connecting the definition word information corresponds to the production line. As a product, the syntax of a sentence can be analyzed and information contained in the sentence can be acquired by the operation of definition word information that is a robot.
[0011]
This is characterized by the fact that it is not necessary to interpret the grammar of the entire sentence, and that it is possible to cope with any sentence expression by a combination of definition words by describing the influence of the definition word on other definition words as data. Yes.
[0012]
By performing the above operations and extracting the information contained in the sentence as individual sentences, it is not necessary to convert the sentence or vocabulary into a mathematical model or numerical value, so the amount of information required for analysis and the amount of data to be stored are reduced. It is possible to talk at a natural speed by shortening the analysis time. Also, by storing the information contained in the text as knowledge data using character strings of combinations of defined words, it is possible to accumulate information and meaning contained in the expression of words that are difficult to quantify with less data It is characterized by being able to do it.
[0013]
DETAILED DESCRIPTION OF THE INVENTION
In order to solve the above-described problems, embodiments of the present invention will be described in detail with examples. In this example, as a simple exchange of words,
Information 1 “I went to the park and ate a lunch.”
Question 1 "What did you eat?"
Question 2 “Where did you eat lunch?”
Question 3 “What did you eat in the park?”
Question 4 "Did you eat lunch?"
Question 5 “Did you eat lunch in the park?”
Question 6 “Did you eat lunch in the park?”
Question 7 “Did you eat an apple in the park?”
Assuming a conversation with questions 1 to 7 with respect to the above information 1, humans can easily answer, but information 1 is stored in a computer and the above questions are input by humans. In the prior art, it is not easy for the computer to answer the question. This is because the conventional method attempts to realize dialogue with the computer by using limited words and grammar. In the present invention, by accumulating information such as information 1, a method for accurately answering variously expressed questions such as the above questions 1 to 7 is described in detail, and the natural language response device is described. Do.
[0014]
The present invention is applied to, for example, a natural language response device 1 as shown in FIG. The natural language response device 1 includes an input unit 2 for operating the natural language response device 1 by an operator inputting information or a character string. The input unit 2 is composed of a pointing device such as a keyboard or a mouse, for example, and becomes an object of input processing when operated by an operator, and is sent to the character conversion unit 3 as a character string indicating a natural language converted by the input unit. Is output.
[0015]
The input unit 2 described above is not limited to a keyboard or the like. For example, the input unit 2 may convert a natural language into a character string by performing voice recognition on a voice signal input from a microphone. Moreover, it is possible to input external text data written in a data file, and any means capable of inputting a natural language can be widely applied.
[0016]
When the sentence of the information 1 is input by the input unit 2 in a natural language, if the character string output to the character conversion unit 3 is, for example, hiragana,
The character conversion unit 3 inputs “I have eaten the word” as the character string of the information 1 converted by the input unit 2, and the definition word information stored in the definition word storage unit 4 is input. Use it to convert to a standard character string normalized.
The definition word information 20 has areas R1 to R14 as shown in FIG. 2 and corresponds to what is called a dictionary in the conventional morphological analysis, but is not simply stored fixed information such as vocabulary and part of speech. A part of the morpheme analysis process is defined as data of the definition word information 20. This removes the analysis process from the conventional main program that determines the connection information and syntax of the vocabulary constituting the sentence, and sets only the standard analysis process in the main program. For sentence analysis processing that cannot be performed by the standard analysis processing of the main program, the analysis processing is defined as a procedure in a description language as data of definition word information. Therefore, the analysis processing of the natural language response device 1 is determined by a combination of the procedures described in the definition word information 20 constituting the sentence.
[0017]
FIG. 2 shows the data structure of the definition word information 20. The definition word information 20 records, as definition words, character strings whose meaning can be specified by kanji, hiragana, katakana, alphabetic characters, numbers, symbols, and combinations thereof. As long as the definition word is a character string whose meaning can be specified, a file name, image data, or the like may be expressed by a character string, and need not be a word. In the definition word R1, the identification number R2, the reading R3, the sending kana R4, the part of speech R5, the special use R6, the use R7, and the notation method R8 are stored as attributes, and in order to digitize the meaning of the definition word R1, Usage method R9 describing the procedure for changing the standard selection procedure, operation method R10 describing the operation procedure, sentence meaning extraction method R11 describing the procedure for changing the standard sentence meaning extraction procedure, standard answer generation procedure Generation method R12 that describes the procedure for changing the name, related definition word R13 that describes one or more other definition words that can be imagined from the definition word, and other definition words that have the same meaning as necessary One or more synonymous definition terms R14 are stored in accordance with.
[0018]
The identification number R2 is a unique number assigned to the definition word R1, and is used to specify the definition word R1 when the meaning differs depending on the homonym and the reading. In “Kin” and “Kane”, the notation “Kin” is the same, but the usage and meaning differ depending on how it is read. In order to distinguish such vocabulary, the definition word R1 is distinguished by the identification number R2.
[0019]
The reading R3 is defined using a hiragana that is the pronunciation of the definition word R1 divided into one character of the definition word R1. When "structure" is defined as the definition word R1, in the reading R3, it is defined by separating it with a white space character as "this □ elephant" (□ represents a space character). As a result, it is possible to search “structure”, “construction”, and “construction” as “structure”. The sending pseudonym R4 defines one or more sending pseudonyms when the sending pseudonym is required in the use of an idiomatic phrase or the like. Sending kana such as verbs and adjectives are defined by definition word R1, special use R6, and use R7. In addition, in order to perform a search for abbreviated sending kana, a notation such as “done” or “done” is registered in R3 as “done”, and abbreviated sending kana as part of the attribute of the notation method R8 of the definition word R1. Define the attribute to be searched. If the last character of the reading R3 is the same as the last character of the reading R3, the definition word R1 having the sending kana abbreviation search attribute "does" or "does" by omitting the last character of R3 as a search target. It is possible to search as “perform standard character string”.
[0020]
The part of speech R5 defines a grammatical part of speech used in linguistics. Further, the distinction and unit of the first name and the last name with respect to the person name are defined in R5 as an extended part of speech.
[0021]
Special utilization R6 defines special utilization such as utilization of sub-variables, utilization of adjective verbs, utilization of union forms, utilization of adverbs, utilization of cucumbers, utilization of siku, utilization of talis, utilization of nari.
When “definition” is defined as a definition word, it is defined that special variable R6 is used as a sub-variable. This defines usage changes including particles such as “define”, “defined”, “not defined”, final particles, case particles, etc. , Past forms, progressive forms, affirmation, denial, questions, etc.
[0022]
Utilization R7 is a verb utilization other than special utilization R6. It is a five-stage utilization, an upper-stage utilization, a lower-stage utilization, an upper-second-stage utilization, a lower-second-stage utilization, a Ka-line modification utilization, a Sa-line modification utilization, a Na-line modification utilization, Define grammatical verb usage, such as the use of la line modification. As a result, similar to the special use R6, it is possible to recognize a grammatical present tense, past tense, progressive tense, affirmation, denial, question, and other sentence forms and logical values.
[0023]
The notation method R8 defines a standardized notation format. Expressing “Kilometer” as “Km” and “Apple” as “Apple” are defined as standard character strings. In addition, it is defined as a standardized standard character string that “do” is used for the sending of a sending pseudonym such as “do” or “do”.
[0024]
The usage method R9, the operation method R10, the sentence meaning extraction method R11, and the answer generation method R12 are procedures written in a program description language for changing the natural language analysis processing procedure, and are related to the definition word R1 in the sentence. Define a procedure that changes the combination of definition words on the condition of.
[0025]
The related definition word R13 defines one or more other definition words R1 that can be imagined from the definition word R1 as necessary. When “apple” is defined as the definition word R1, definition words R1 such as “food”, “fruit”, “tree”, “red”, “delicious” and the like can be defined as the related definition word R13. Thereby, the unified meaning which arises by the combination of definition word R1 can be estimated using this related definition word R13. The synonymous definition word R14 defines one or more definition words having the same meaning and different expression methods as necessary. “Gold” and “gold” refer to the same “precious metal gold”, although they have different notations. When “Gold” is written in the text, it can be recognized as “Kin”.
[0026]
The character conversion unit 3 in FIG. 1 can acquire the definition word information 20 from the definition word storage unit 4 with respect to the character string “I have eaten the text” and can compose the character string. Create a combination. In this combination, based on the position information of the character string, the definition word R1 is defined as the definition word information 20 starting from the same position, and a plurality of lines are combined into one column and connected in the order of the position of each column. . The state of this combination is called a definition word conversion matrix and is shown in FIG. In the definition word conversion matrix of FIG. 3, vocabulary starting from “ko” that is the first character of the input character string is arranged in a first column, and the second column starts from the second character “u” of the input character string. Vocabulary is organized into columns. This definition word conversion matrix M1 shows the initial state for the input character string. An expression in which a definition word is surrounded by a rectangle in the definition word conversion matrix M1 indicates that the definition word information 20 includes R1 to R14 shown in FIG. 2, and the vocabulary in the rectangle indicates the definition word R1.
[0027]
In the column constituting the definition word conversion matrix M1, conversion candidates corresponding to the input character string are grouped according to the start position of the input character string, and therefore there are columns that are not used by the selection of the definition word R1.
FIG. 3 shows that “sponsorship” is selected as a conversion candidate for the input character string “Koen” in the first row of the first column. In this case, “Kouen”, which is a part of the input character string, is referred to as a first column conversion area. Assuming that the conversion area of the first column is “Koen”, the conversion areas of the second and third columns overlap with the conversion area of the first column and cannot be connected to the first column. The second column and the third column in FIG. 3 displayed one level lower indicate a state in which connection is not possible. The fourth column can be connected to the first column, and the conversion region of the first column is in contact with the conversion region. The conversion candidate “bento” can be connected to the conversion area “Bento” in the seventh column that can be connected to the fourth column. In this case, it is clear that the three conversion candidates are all nouns, but in order to determine a new noun by performing noun combination, it is a necessary condition that the meaning of the related definition word R13 in the definition word information 20 is uniform. It becomes.
[0028]
The related definition word R13 of the definition word information 20 is defined by another definition word that can be imagined from the definition word R1. This means that by defining the image that can be imagined from the character string and words of the definition word R1 with other words, the group to which the definition word R1 belongs can be conceptualized, and it is necessary to form a fixed group. Absent. By defining a plurality of words in the related definition word R13, a plurality of different concept groups can be defined, or by defining a similar concept group, a narrower concept group can be formed simultaneously. . For example, by defining the broad meaning of the noun “dog” using words such as “animal” and “living organism”, and at the same time, defining the breed of dogs such as “Akita Inu”, “Shepard”, and “Bulldog”. You can make up a group of dogs themselves and do not need to be aware of the interrelationships and dependencies of registered words.
[0029]
The conversion candidate can determine whether meaning can be created when the noun combination is performed by determining the uniformity of meaning using the related definition word R13. The related definition word R13 can be defined by any other definition word that can be created from the definition word R1, and assuming the related definition word R13 for the three nouns,
R13 such as “person”, “venue”, “meeting”, “lecture hall” for “support”
R13 such as “description”, “explanation”, “arrange”, etc.
R13 such as "Food", "Meal", "Love wife", "Home" for "Bento"
Define There is no common word for judging the unity of meaning from R13 of the top three nouns. Furthermore, even if the definition word information 20 of the words defined in R13 (for example, “person”, “description”, etc.) is obtained and the related definition words R13 are compared, the common definition word R13 is shared by the second related definition word R13. I can't find words. Therefore, it can be determined that there is no unity of meaning for the top three nouns, and the three nouns cannot be combined. It can also be determined that the combination of the top two nouns “sponsored” and “joint” cannot be combined. From the above, it can be determined that the “concurrent writing” that is the second conversion candidate is not a conversion candidate. For the fourth column of the transformation matrix M1 in FIG.
It can be determined that “weapons” and “peace” are not unified by noun combination. From this result, when it is assumed that the conversion area of the first column is “Kouen”, it can be determined that it is an error that the conversion area connected to the conversion area of the first column is “successful”. When the conversion candidate for the conversion region in the fourth column is moved to the lowest row in the fourth column, the conversion region becomes “to”.
[0030]
Assuming that the conversion region in the fourth column is “to”, the conversion candidate in the fourth column is “to” the case particle. For the usage method R9 of the definition word information 20 for the case particle “to”, the usage method for “to” is described in a description language. FIG. 4 shows an example of the usage method described in the description language registered in the usage method R9 of the case particle “to”. The content described in this usage method R9 is as a connection condition to the case particle "to"
Define a connection condition that has the priority of "noun + to + verb", "noun + to + adjective", "noun + to + adverb", or "noun + to", and connect to the case particle after "to" Specifies that the part of speech to be moved to the beginning of the column to which it belongs. However, in order to be able to move, the unification of the meaning of the noun that precedes “to” and the part of speech that moves is a necessary condition.
The fifth column connects to the fourth column conversion area, and the first column conversion area.
When the combination of the conversion candidate having the noun attribute for “Kouen” and the conversion candidate for the fifth column conversion area “Iki” is subjected to the above-described uniformity of the meaning, “noun + to + verb” in the connection conditions With respect to “,” “park” in the first column and second row and “go” in the first row in the fifth column, it can be determined that “going to park + to + go” holds. This is because the noun “park”, R13 such as “place”, “play”,
R13 such as “place” and “building” for the verb “go”
Can be defined, and it can be determined that the meaning is uniform by the related definition word “location”. Therefore, it can be determined that “Go to Koen” is “Go to the park”. The same process can be performed for the remaining “between meals”. As a connection condition to the usage method R9 of the conjunction ""
"Noun + a + verb", "Noun + a + adjective", "Noun + an adverb",
By defining “noun +”, etc., “eating lunch” can be established by considering the unity of the meanings of the nouns in front of “to” and the parts of speech behind “to” Can be judged. Therefore, the transformation matrix M1 shown in FIG. 3 finally has the structure of the transformation matrix M1 shown in FIG. By using the definition word in the first row of the connected columns in FIG. 5, the conversion result for the input character string can be obtained.
[0031]
The information extraction unit 5 in FIG. 1 extracts information contained in the input character string using the definition word conversion matrix M1 shown in FIG. Information is extracted by dividing it into the smallest units that make up a sentence. Usually, sentence division is performed in the form of “noun + verb” as one unit, but when the extraction method is described in the sentence meaning extraction method R11 of the definition word information 20, the described procedure is executed. By doing. If the conversion result “I went to the park and ate lunch” in FIG. 5 is divided into units of “noun + verb”, it can be divided into “go to the park” and “eat the lunch”. The above sentence divided into two is a character string corresponding to the input character string
“Kenen Iki” and “Bento Eta” are acquired, and the respective definition word conversion matrices M2 and M3 of the divided sentence shown in FIG. 6 are created using the above conversion method.
[0032]
The definition word conversion matrices M2 and M3 are the result of dividing the definition word, which is a component of M1, into the smallest units that can constitute a sentence, and the influence of the usage method R9 of the definition word information 20 and the related definition word R13. The result is minimized. For this reason, M2 and M3 are converted by a direct vocabulary relationship. If the results of M1, M2, and M3 are different, the conversion results of M2 and M3 have higher reliability. In such a case, the reliability can be further improved by reflecting the results of M2 and M3 on M1. In general, it is possible to determine whether or not the conversion result of M1 is correct by M2, M3,... Mn (n indicates the number of information included in the input character string), and if the conversion result of M1 is incorrect Can automatically correct M1 using M2, M3,... Mn.
[0033]
The definition word conversion matrix M1 for the input character string is a standard character string conversion result based on the definition word, and the definition word conversion matrices M2 and M3 are information included in the input character string.
The information storage unit 6 in FIG. 1 creates the definition word conversion matrices M1, M2, and M3 in sentence units with the structure shown in FIG. 7, and stores them in the knowledge data storage unit 7 in FIG. In the structure of FIG. 7, as for the above input character string, as “M2“ Fact to go to the park ”, as M3“ Fact to eat the lunch ”, M2 and M3 connection,“ After going to the park and ate the lunch ” "Time passage and conditions".
[0034]
The information stored in the knowledge data storage unit 7 in FIG. 1 is stored using the data unit shown in FIG. 7 using the hierarchical structure shown in FIG. The hierarchical structure of FIG. 8 is created as necessary. For example, by describing a sentence for storing information in a file, and adding a chapter number and a section number to the chapter or section title, Create a hierarchical structure. In the hierarchical structure, address pointers PV (P2) and NX (P3) for connecting other records are provided in M1 in which definition word conversion matrices M2, M3. For the title text, address pointers BP (P1), PV (P2), and NX (P3) are provided. For title sentences that form a hierarchical structure, hierarchical end data P4 is provided, and the leading address of the P4 record is stored in P1 at the hierarchical starting position, and the leading address of the P1 record is stored in P4. M1 and M3 are connected to the sentence of title 1 (M1), and P1, P2, and P3 are provided in M1 as a whole, and record 1 (R1) is shown, and the connection relationship from R1 to R6 is shown. R1 has a hierarchical structure, defines R6 as an end record of R1, stores the start address of R6 in P1 of R1, and stores the start address of R1 in P1 of R6. P2 stores the top address of the previous record, and P3 stores the top address of the next record. R1 and R6 are peculiar records, the head address of R6 is stored in P2 of R1, and the head address of R1 is stored in P3 of R6. By storing the top address of the record for P1, P2, and P3 of R1 to R6, a hierarchical structure of information stored in the sentence is created, and the positional relation of the entire sentence managed for each line is constructed. By constructing a bi-directional circular search chain with P1, P2, and P3 for each row of records, it is possible to estimate the range of necessary information and enable high-speed search.
[0035]
When the information included in the input character string is insufficient, the character conversion unit 3 and the information extraction unit 5 estimate the insufficient information by searching the knowledge data stored in the knowledge data storage unit 7. For example, when converting “can be used”, there is not enough information for converting “ro”, and it cannot be determined whether it can be converted to “furnace” or “櫓”. In such a case, if the knowledge data storage unit 7 has information converted to “ride on a boat”, define “boat” in the related definition word R13 shown in FIG. Therefore, it can be converted to “can be used” by judging that it is converted to “櫓” due to the influence of “boat”. If the knowledge data storage unit 7 has been converted to “made dish”, define “ceramic”, “ceramic”, “glass”, “seto” to the related definition word R13 of “dish”, and “furnace” By defining “baking”, “ceramics”, “melting”, etc. in R13 of “”, it can be converted to “furnace” and can be converted to “furnace can be used”. If the corresponding knowledge data does not exist in the knowledge data storage unit 7, the conversion candidate can be determined by asking what it is to be used for using the related definition word R13 of “furnace” or “櫓”. It can be confirmed.
[0036]
FIG. 9 shows an example in which the conversion result and the information extraction result by the above operation are displayed on the screen.
Read "Kenen Ikibento Eta" (a1), search the definition word storage unit 4, create definition word conversion matrices M1, M2, and M3, define the definition word data in consideration of the meaning The analysis results of the word conversion matrices M1, M2, and M3 are displayed, and the contents of the extracted information stored in the knowledge data storage unit 7 are shown. The display of a2 shows the input character string. a3 is converted using the definition word conversion matrices M1, M2, and M3, and is displayed in accordance with the notation method R8 of the definition word information 20 in FIG. In the display of a4, the conversion result is displayed for each part of speech separated by a blank character. The display of a5 uses the conversion result of a3 and the reading R3 of the definition word R1, and displays the reading R3 by separating each definition word with a blank character. In the display of a6 and a7, information defined in the sentence a2 is extracted and displayed for each extracted information.
[0037]
A question is accepted in natural language about the knowledge data stored in the knowledge data storage unit 7 by the information storage method. Information accumulation and questions are switched by a pointing device such as a mouse. The operation is switched by determining from the definition word conversion matrix, which is the analysis result of the sentence, that the sentence is a question sentence from the special use R6 and the use R7 of the definition word information 20 of the verb.
The input character string of the information 1 is stored as information, and a response method when the question 1 is input will be described.
Question 1 input by the input unit 1 of FIG. 1 by the above conversion and information extraction method
"What did you eat?" Is output to the character conversion unit 3, and the result of conversion is input to the information extraction unit 5 to create the definition word conversion matrix shown in FIG. The information included in question 1 is only M2, the results of M1 and M2 are the same, and the conversion result is
"What did you eat?"
[0038]
The answer generation unit 8 in FIG. 1 receives the definition word conversion matrices M1 and M2 and generates an answer to the question 1. An answer can be generated by defining an answer method using a description language to an answer generation method R12 of the definition word information 20 for a definition word that is a key of the answer. In the case of the above question 1, the answer method is described in the answer generation method R12 of the definition word information 20 of the pronoun “what”. FIG. 11 shows a description example of the pronoun “what” in R12. The answer generation method shown in FIG.
Knowledge data stored in the knowledge data storage unit 7 by using the related definition word R13 of the verb “eat” for the question form of “what + conjunction + verb” and “what + case particle + verb” M2, M3,..., Mn, the noun connected to the matching verb, and the conjunction and verb retrieved from question 1 are used as the answer sentence. Therefore, the above procedure links the noun “bento” and the “question 1” from “M3” in FIG. 6 with the “eating” information 1 verb, and uses the R7 in FIG. By obtaining “I ate” as the response type of “I ate”, the response sentence “I ate lunch” can be generated.
[0039]
The answer display unit 9 in FIG. 1 outputs the above answer sentence to the output device, and outputs “I ate lunch” as an answer to the question 1 with respect to the information 1 by, for example, a display or voice output. Realize natural language response.
[0040]
If you entered the question 2 “Where did you eat your lunch?” To the sentence meaning extraction method R11 with the case particle “,” use “noun +” and “pronoun +” as the extraction unit, and the rest of the sentence Describe the procedure to extract by "noun + verb". In this case, as the extraction result of R11, “where” and “I ate lunch” can be defined as the definition word conversion matrices M2 and M3. In the question sentence, the extracted information that does not include pronouns is a condition for searching knowledge data. Since the information 1 satisfies this condition, it can be determined as an answer target sentence. In the answer generation method R12 of the pronoun “where”, a procedure is described in which a noun indicating “location” in the related definition word R13 is searched from the knowledge data as an answer condition. This makes it possible to create an answer condition of “in the park” and output “I ate in the park” as an answer.
On the other hand, the above question 3 can be output as “Lunch was eaten” by performing the same processing as the above processing, with “in the park” as the search condition.
In question 4 above, the pronoun “what” in question 1 is the noun “bento”
The question 4 itself is a search condition. In this case, “Yes” or “No” is the answer, and there is knowledge data that matches the conditions of Question 4.
In the form of “yes + verb”, “yes” can be output.
The above question 5 has two conditions of “in the park” and “eat lunch”, and the same result is obtained by the same process except that the number of conditions is different from that of question 4.
In the above question 6, the information extraction condition is question 2, and the search condition is the same as question 5,
The answer is the same as in question 5.
Question 7 is the same process as Question 5, but the sentence that satisfies the search condition “in the park” in the knowledge data does not contain a sentence that satisfies “eating an apple”, so “No” is output as an answer can do.
[0041]
As described above, sentences containing specific facts and information are stored as knowledge data as sentences, and sentences are used as search conditions to query the target information in natural language. Can be obtained. In addition, by describing the vocabulary-specific usage and answer methods that cannot be expressed by analysis patterns or grammatical patterns in natural language analysis in the key vocabulary, the meaning of the expression of many words can be estimated.
[0042]
【Example】
A specific example using the embodiment of the present invention will be described with reference to the drawings.
[0043]
FIG. 12 shows an embodiment of the standard character string automatic conversion apparatus 10 according to the present invention. The standard character string automatic conversion device 10 includes an input unit 2, a character conversion unit 3, a definition word storage unit 4, and an information extraction unit 5, which are the same as the natural language response device 1 of FIG. 1 described in the above embodiment. Further, a conversion result temporary storage unit 11 for temporarily storing the conversion result and a character output unit 12 for displaying the conversion result on the display device are provided.
The input unit 2 is not limited to a keyboard or the like. For example, the input unit 2 may convert a natural language into a character string by performing voice recognition on a voice signal input from a microphone.
Moreover, it is possible to input external text data written in a data file, and any means capable of inputting a natural language can be widely applied.
[0044]
The character string input by the input unit 2 uses the method of the above embodiment to obtain the definition word information 20 stored in the definition word storage unit 4 corresponding to the input character string, and the character conversion unit 3 A defined word conversion matrix M1 shown in FIG. 3 is created.
[0045]
The information extraction unit 5 uses the definition word conversion matrix M1 to extract information contained in the input character string by the information extraction method described in the above embodiment, and defines the definition word conversion matrix M1 shown in FIGS. , M2,... Mn are created, and this definition word conversion matrix is stored in the conversion result temporary storage unit 11.
The character output unit 12 outputs the converted character string to an output buffer for displaying.
[0046]
By the above-described processing, the vocabulary can be automatically selected by grasping the meaning of the sentence in the method that has been performed by the vocabulary use frequency or the menu selection of the homonyms until now. In addition, when automatic conversion cannot be performed due to lack of information, the operator does not need to learn an operation method necessary for conversion by a method of supplementing the shortage information by asking a question to the operator.
[0047]
In addition, the standard character string automatic conversion device 10 not only converts “Hiragana” to Kanji, but also converts character strings expressed by Kanji, Hiragana, Katakana, English, numbers, symbols, and combinations thereof in the above embodiment. As described, the definition word information 20 can be converted into a standardized standard character string using the definition word R1, the reading R3, the sending kana R4, and the notation method R8.
[0048]
The embodiment shown in FIG. 13 is a database search method in natural language,
The flowchart of the process which searches the data of a database by a natural language with respect to the external database connected with the communication line or the external database constructed | assembled on the same computer is shown.
By the method described in the above embodiment, an outline explanation of data registered in the external database is created as an external file, the external file is read, and knowledge data related to the external database 30 is created in the knowledge data storage unit 7. .
[0049]
For example, when the table A and the column B of the external database 30 are written to the external file “column B of the table A is the date of birth of the employee.”, The extraction information of the conversion matrices M1, M2,. Is
“Column B of table A is the date of birth of the employee”, and the result is stored in the knowledge data storage unit 7. “Birthday” is registered in the related definition word R13 of the definition word “birth date”, and “date” is registered in R13 of the definition word “birthday”.
[0050]
In step 1 (S1), a condition for searching the external database 30 from the input device is input in a natural language. The device to input the character string is a keyboard,
The present invention can be widely applied as long as it can input a natural language such as a character string input by a voice recognition system or a document stored in an external file from a voice input from a microphone.
[0051]
In step 2 (S2), when an operation to end the process is performed, the process ends. If the input character string exists, processing is performed using the method described in the embodiment of the present invention. Step 3 (S3) corresponds to the character conversion unit 3 in FIG. 1, and step 4 (S4) corresponds to the information extraction unit 5. The input character string creates a definition word conversion matrix M1 as shown in FIG. 3 in S3, and definition word conversion matrices M1, M2,... Mn shown in FIGS. In step 5 (S5), using the definition word conversion matrices M1, M2,... Mn, information is extracted with the structure shown in FIGS. 7 and 8 and stored in the temporary storage device.
[0052]
In step 6 (S6), the information stored in the knowledge data storage unit 7 related to the external database 30 is searched using the information stored in S5, and the data representation (table name, column name, If there is no stored information, the operator outputs insufficient information and prompts the operator to re-enter the conditions.
[0053]
In step 7 (S7), a condition for searching the external database 30 is generated using the condition extracted in S4. The search condition is generated using the part of speech R5 and the synonymous definition word R14 of the definition word information 20 constituting the accumulated information. Nouns are determined as table names, column names, and data values, and adjectives and adverbs are data range conditions. In addition, verbs, particles, case particles, and conjunctions are determined as calculation elements constituting conditions, calculation elements can be created using the synonymous definition word R14, and conditions with grammar are described in the answer generation method R12. The generation procedure is described using a language.
[0054]
For example, when “employees born in 1970” is entered as a search condition, the extraction information of the definition word conversion matrices M1, M2,.
“Employees born in 1970”, 1970 is a date, and employees can be judged as personal pronouns. In this case, the usage method R9 of the definition word “born” having the noun attribute in which “birthday” is registered in the related definition word R13 is
In the case of “date + born + personal noun”, a procedure for moving to the first row of the column of the definition word conversion matrix including this noun is described. “Born in 1970” becomes “noun + noun” and indicates a noun combination state. In this case, R13 of “1970” has “date” and R13 of “born” has “birthday”, but “date” and “birthday” do not match. In this case, referring to the definition words “date” and R13 of “birthday”, “date” exists in R13 of “birthday”. Therefore, R13 estimated for the combined noun “born 1970” is “date” or “birthday”. Therefore, in the above example, when the knowledge data stored in the knowledge data storage unit 7 is searched under the conditions of “date” and “employee”, “birthday” and “employee”,
“Column B of table A is the date of birth of the employee” can be acquired. This sentence contains nouns connected by case particles, but connection by case particles "
Since it can be unconditionally interpreted as a single noun, it can be determined that “A1 is composed of A2”, and it can be determined that A2 includes “birthday” and “employee”. Therefore, a procedure for extracting the table name and column name from the result of searching the knowledge data in the knowledge data storage unit 7 and the range condition from the input search condition is described in the answer generation method R12 of the definition word indicating that the conjunction “has”. Keep it.
[0055]
When the search condition in the natural language is applied to a database search described by an SQL sentence, when a table name and a column name are extracted from a result of searching the knowledge data storage unit 7, table name = “table A”, column name = ” Column B ", and when the range condition is extracted from the input search condition," 1970 "is obtained. Therefore, the SQL statement including the search condition is
"SELECT * FROM table A WHERE column B =
1970; ". The name of the external database to be accessed is known by the system, but the database can be opened by describing the database name in a document file and accumulating it as knowledge data. When a database exists, an external database can be selected from search conditions of the external database by describing table and column information for each external database in the document file.
The search conditions are stored as search conditions in step 7 (S7).
[0056]
In step 8 (S8), the external database 60 is searched using the above search condition or a question is asked to the operator. If the search is not performed, the following condition is input.
When “or an employee born in 1960” is input as the next search condition, a composite condition is created by adding the input condition to the search condition. In this case, “OR” is registered in the synonymous definition word R14 of the definition word indicating the conjunction “or”.
When condition analysis is performed as in the above example, the condition items in the SQL statement are
“OR column B = 1960”. Therefore, the SQL statement that combines the two conditions is
“SELECT * FROM table A WHERE column B = 1970 OR column B = 1960;”.
[0057]
In the above example, the display conditions can be set similarly. When columns C and D exist in table A, if “display column C and column D” is input as a display condition, and if columns C and D are combined with the above search condition as a display column, the SQL statement is
"SELECT column C column D FROM table A WHERE
Column B = 1970 OR column B = 1960;
[0058]
If an instruction to search the external database 30 with the above search condition is given to the external database in S8, the external database 30 is accessed according to the search condition in step 9 (S9), the searched data is acquired, and the search result is sent to the display device. Can be displayed.
[0059]
【The invention's effect】
As described above, the natural language response apparatus and method of the present invention can facilitate the interaction between a human and a computer, and between the computer and the computer in a natural language. This enables dialogue by voice input or character string input.
[0060]
In addition, the natural language response device and method according to the present invention uses an analysis method that does not require a grammatical pattern, and handles a vocabulary processing method as vocabulary data, thereby simplifying many processing forms of words. Processing can be performed.
[0061]
By using the method of the present invention, the kanji conversion device uses the method of the present invention to convert a kanji, hiragana, alphabetic characters, numbers, symbols, and combinations thereof into a regular character string instead of the conventional kanji conversion. It can be easily converted into a standardized character string.
[0062]
According to the method of the present invention, it is possible to take out information stored in a computer while deriving an operation method by dialogue, and therefore, an operation that does not require knowledge of the computer or knowledge about stored data is possible.
[Brief description of the drawings]
FIG. 1 is a block diagram showing a configuration of a natural language response device to which the present invention is applied.
FIG. 2 is an explanatory diagram showing an area of definition word information to which the present invention is applied.
FIG. 3 is an explanatory diagram showing a data structure in an initial state of a definition word conversion matrix to which the present invention is applied.
FIG. 4 is a diagram showing an example of a procedure for registering in the definition word information usage method to which the present invention is applied.
FIG. 5 is an explanatory diagram showing a data structure of a defined state of a definition word conversion matrix to which the present invention is applied.
FIG. 6 is an explanatory diagram showing a definition word conversion matrix of a semantic division character string to which the present invention is applied.
FIG. 7 is an explanatory diagram showing an information extraction state of a sentence to which the present invention is applied.
FIG. 8 is an explanatory diagram showing a knowledge data structure having a hierarchical structure to which the present invention is applied.
FIG. 9 is an explanatory diagram showing a display example of a conversion result and an information extraction result of a definition word conversion matrix to which the present invention is applied.
FIG. 10 is an explanatory diagram showing a data structure of a defined state of a definition word conversion matrix for a question to which the present invention is applied.
FIG. 11 is a diagram showing an example of a procedure for registering in a definition word information answering method to which the present invention is applied.
FIG. 12 is a block diagram showing the configuration of a standard character string automatic conversion apparatus to which the present invention is applied. Example 1
FIG. 13 is an explanatory diagram showing a flowchart of a process for searching for data in an external database in a natural language to which the present invention is applied. (Example 2)
[Explanation of symbols]
1 Natural language response device
2 Input section
3 Character conversion part
4 definition word storage
5 Information extraction part
6 Information storage
7 Knowledge data storage
8 Answer generator
9 Answer display area
10 Standard character string automatic converter
11 Conversion result temporary storage
12 character output part
13 Display
14 Speaker
20 Definition word information
30 External database

Claims (15)

A character string composed of kanji, hiragana, katakana, alphabetic characters, numbers, symbols and combinations thereof that can specify the meaning is called a definition word.
Input means for receiving natural language input;
A conversion means for converting the input natural language into a normalized standard character string;
For each definition word, the description language describes how to use the definition word when it is used in the input natural language, how to extract the definition word, and how to answer the question about the definition word. Definition word storage means for storing a procedure and a plurality of attributes of the definition word as semantic determination information;
Among the meaning determination information, semantic estimation means for estimating the meaning of each definition word by determining a method of using each definition word and a combination of definition words constituting the standard character string using a plurality of attributes,
Information extracting means for extracting information contained in the natural language input by the method of extracting each definition word estimated by the meaning estimating means;
Knowledge data storage means for storing knowledge data information extracted by the information extraction means;
A natural language question is received by the input means, a combination of definition words constituting the question is estimated by the semantic estimation means, and a definition word answering method in the combination and knowledge data stored by the knowledge data storage means A natural language response device comprising: an answer generation unit configured to generate an answer corresponding to a question using information.   The meaning determination information stored in the definition word storage means stores a processing procedure in which the definition words constituting the input natural language influence each other due to the relationship in the definition word usage method described in the description language. The natural language response apparatus according to claim 1, wherein the natural language analysis is performed using the stored processing procedure. The semantic determination information stored in the definition word storage means stores a processing procedure in which the definition words constituting the input natural language influence each other due to the relationship described in the definition word extraction method in the description language. And using the memorized processing procedure,
The natural language response device according to claim 1, wherein information included in the input natural language is extracted with a character string expressed by a combination of definition words. The meaning determination information stored in the definition word storage means stores a processing procedure in which the definition words constituting the input natural language influence each other due to the relationship in the definition word answering method described in the description language. And using the memorized processing procedure,
The natural language response apparatus according to claim 1, wherein an answer generation process corresponding to an information acquisition request included in the input natural language is performed.   Character conversion, vocabulary meaning determination, sentence meaning determination, extraction of information contained in the sentence, vocabulary structure change and syntax structure change based on the vocabulary, and response sentence generation using a defined word conversion matrix The natural language response device according to claim 1. A natural language response method executed by a computer,
A character string composed of kanji, hiragana, katakana, alphabetic characters, numbers, symbols and combinations thereof that can specify the meaning is called a definition word.
For each definition word, the description language describes how to use the definition word when it is used in the input natural language, how to extract the definition word, and how to answer the question about the definition word. Semantic determination information consisting of a procedure and multiple attributes of its definition word is stored,
Accepts natural language input, converts the input natural language into a standardized standard string,
Among the above-mentioned semantic determination information, the meaning estimation process of each definition word is performed by determining the combination of definition words constituting the input natural language using the use method of each definition word and a plurality of attributes. The information included in the natural language input by each definition word extraction method is extracted, and the extracted information is stored as knowledge data information.
Accepts a natural language question input, estimates the combination of definition words constituting the question by the semantic estimation process, and responds to the question using the definition word answer method in the combination and the knowledge data information Generate an answer,
Natural language response method characterized by that. On the computer,
A character string composed of kanji, hiragana, katakana, alphabetic characters, numbers, symbols and combinations thereof that can specify the meaning is called a definition word.
For each definition word, the description language describes how to use the definition word when it is used in the input natural language, how to extract the definition word, and how to answer the question about the definition word. Semantic determination information consisting of a procedure and multiple attributes of its definition word is stored,
Accepts natural language input, converts the input natural language into a standardized standard string,
Among the above-mentioned semantic determination information, the meaning estimation process of each definition word is performed by determining the combination of definition words constituting the input natural language using the use method of each definition word and a plurality of attributes. The information included in the natural language input by each definition word extraction method is extracted, and the extracted information is stored as knowledge data information.
Accepts a natural language question input, estimates the combination of definition words constituting the question by the semantic estimation process, and responds to the question using the definition word answer method in the combination and the knowledge data information Generate an answer,
The computer-readable recording medium which recorded the program for performing the process with. A character string composed of kanji, hiragana, katakana, alphabetic characters, numbers, symbols and combinations thereof that can specify the meaning is called a definition word.
Input means for receiving natural language input;
A conversion means for converting the input natural language into a normalized standard character string;
For each definition word, the description language describes how to use the definition word when it is used in the input natural language, how to extract the definition word, and how to answer the question about the definition word. Definition word storage means for storing a procedure and a plurality of attributes of the definition word as semantic determination information;
Among the meaning determination information, semantic estimation means for estimating the meaning of each definition word by determining a method of using each definition word and a combination of definition words constituting the standard character string using a plurality of attributes,
Information extracting means for extracting information contained in the natural language input by the method of extracting each definition word estimated by the meaning estimating means;
Means for converting the input natural language into a more normalized standard character string by constructing a sentence by a combination of definition words using the analysis result of the semantic estimation means and the information extraction means. Standard standard character string automatic conversion device.   The semantic determination information stored in the definition word storage means defines a state in which the definition words constituting the input natural language influence each other depending on the relationship in the state used in the input natural language. 9. The input natural language is converted into a normalized standard character string using a processing procedure that is classified into a word usage method and an extraction method and described in a description language. Standard string automatic conversion device.   The meaning determination information stored in the definition word storage means stores a processing procedure in which the definition words constituting the input natural language influence each other due to the relationship in the definition word usage method described in the description language. 9. The standard character string automatic conversion device according to claim 8, wherein natural language analysis is performed using the stored processing procedure. The semantic determination information stored in the definition word storage means stores a processing procedure in which the definition words constituting the input natural language influence each other due to the relationship described in the definition word extraction method in the description language. And using the memorized processing procedure,
9. The standard character string automatic conversion device according to claim 8, wherein information included in the input natural language is extracted as a character string expressed by a combination of definition words.   9. The character conversion, lexical meaning determination, sentence semantic determination, extraction of information included in a sentence, vocabulary structure change and lexical structure change centered on a vocabulary are performed using a definition word conversion matrix. Standard string automatic conversion device described in 1. A standard character string automatic conversion method executed by a computer,
A character string composed of kanji, hiragana, katakana, alphabetic characters, numbers, symbols and combinations thereof that can specify the meaning is called a definition word.
For each definition word, the description language describes how to use the definition word when it is used in the input natural language, how to extract the definition word, and how to answer the question about the definition word. Semantic determination information consisting of a procedure and multiple attributes of its definition word is stored,
Accepts natural language input, converts the input natural language into a standardized standard string,
Among the above-mentioned meaning determination information, the meaning estimation process of each definition word is performed by determining a combination of definition words constituting the standard character string using a method of using each definition word and a plurality of attributes.
Perform information extraction processing to extract information included in the natural language input by the method of extracting each definition word estimated by the semantic estimation processing,
By using the analysis result of the semantic estimation process and the information extraction process, by constructing a sentence by a combination of definition words, the input natural language is converted into a more normalized standard character string.
The standard character string automatic conversion method characterized by this. On the computer,
A character string composed of kanji, hiragana, katakana, alphabetic characters, numbers, symbols and combinations thereof that can specify the meaning is called a definition word.
For each definition word, the description language describes how to use the definition word when it is used in the input natural language, how to extract the definition word, and how to answer the question about the definition word. Semantic determination information consisting of a procedure and multiple attributes of its definition word is stored,
Accepts natural language input, converts the input natural language into a standardized standard string,
Among the above-mentioned meaning determination information, the meaning estimation process of each definition word is performed by determining a combination of definition words constituting the standard character string using a method of using each definition word and a plurality of attributes.
Perform information extraction processing to extract information included in the natural language input by the method of extracting each definition word estimated by the semantic estimation processing,
By using the analysis result of the semantic estimation process and the information extraction process, by constructing a sentence by a combination of definition words, the input natural language is converted into a more normalized standard character string.
The computer-readable recording medium which recorded the program for performing the process with. The plurality of attributes of each definition word of the meaning determination information are:
One or more strings describing the reading of the definition word;
The definition part of speech,
Use of definition words by part of speech,
An attribute that describes the standardized standard string representation of the definition word,
For definition words, one or more other definition words that can be imagined as needed,
One or more other definition words representing the same meaning as the definition word,
A computer-readable recording medium on which the program according to claim 8 or 14 is recorded.

Images