JPH06266779A - Controller - Google Patents

Abstract

PURPOSE:To provide a controller performing the control of a control object equipment by inputting a commonly used natural language. CONSTITUTION:The voice input (1) of a natural language is speech recognized (2), words and phrase related to the control of a control object equipment are extracted by referring to a words and phrases information table 10 based on a recognition result (3), control information is determined by referring to a control information table 11 by words and phrases (4), a control signal is generated form control information to which limit information is not imparted (6), and the signal is outputted to the control object equipment or a data base (7). If a response signal is received from the data base (13), the signal is stored in a response information table 12 (14), control information to which limit information is imparted is supplemented by response information and a control signal is generated (6), and the signal is outputted to the control object equipment.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a control device provided with a natural language interface for controlling the interface between controlled equipment and humans.

[0002]

2. Description of the Related Art Conventionally, individual operating instructions have been issued to a single operating system.

A technique relating to the present invention is disclosed in Japanese Patent Laid-Open No. 63-85.
966 “Intelligent Workstation”. In this document, "When setting up an event such as a conference, the schedule of each individual attending the above-mentioned event is checked according to the individual schedule registered in the database, and a confirmation of a new reservation for the event is made to the individual by telephone and a schedule change. Then, when the confirmation or consent of the telephone voice from the individual is recognized by voice, and when the confirmation or consent of all the individuals attending the event is obtained, the event is set and registered, and at the same time, each individual is registered. The event will be registered in the schedule. ” Also, JP-A-63-22
In 3965 "Intelligent workstation", "In response to a telephone inquiry, graphic information such as a guide map is retrieved from the database, and when the telephone terminal has a communication function of image data, the graphic information is returned as image data. If the telephone terminal does not have the image data communication function, the graphic information is analyzed to create text data indicating the content of the graphic information, and the text data is voice-synthesized to output a response. Will be done. ” Also, Japanese Patent Laid-Open No. 64-2460 "Intelligent Workstation"
"When outputting the document information retrieved and extracted from the database according to the voice information input through the telephone line, the configuration format of the document information is determined by the type of the output terminal, and the output of the document information is performed according to the result. An inquiry is made by voice message to determine the form. "

[0004]

In the above-mentioned prior art, a database search is performed by simply inputting a word such as a scheduled date and time of a conference or names of attendees from a telephone terminal. Specifically, when “providing guide information to Mr. Taro Yamada's home is requested”, “guidance”, “home”, “Yamada”, “Taro”, etc. must be sequentially input. As described above, in the conventional technique, there are strong restrictions on the word order and input expressions, and it has been difficult to perform flexible interpretation of natural language sentences that mean complex queries and cooperative operations. Further, the above-mentioned conventional technology is said to present a correct command candidate for the system used for the first time by inputting the command of the conventional system without knowing the command of the system used for the first time.
However, system commands are accompanied by syntactic restrictions, and it was difficult for a user who was originally unfamiliar with the command format itself to input the commands. Further, it has not been possible to present the user with a procedure (or command sequence) for performing a cooperative operation of a plurality of systems. An object of the present invention is to provide a control device capable of giving an operation instruction command in a natural language from a wireless transmitter or a telephone terminal to remotely control a device to be controlled. Another object of the present invention is to allow a user, who is unfamiliar with the format of commands, to use a natural language when operating a plurality of control target devices so that the control target devices can be properly used or a plurality of control target devices can be used. Another object of the present invention is to provide a control device that can execute a desired operation instruction command by combining controlled devices and performing a cooperative operation.
Other objects of the present invention will be understood from the description of the present specification and the drawings.

[0005]

The above objects are achieved as follows.

The control device comprises a voice recognition means for recognizing the contents of voice input, a database, a word / phrase information table storing word / phrase information relating to control of the control target device, the word / phrase and control information for the control target device. The processing device includes a storage unit accommodating a control information table storing the correspondence relationship and a response information table storing the search result of the database, and a processing unit. Means for referring to and extracting a word related to control of the controlled device, means for extracting control information corresponding to the extracted word by referring to a control information table, and restriction information of the extracted control information Means for generating a control signal based on the control information that is not assigned and outputting it to the controlled device or database, Means for storing response information in the response information table, and complementing the control information to which restriction information is added by the response information stored in the response information table, and generating a control signal based on the complemented control information. A means for generating and outputting to the device to be controlled is provided. Further, the means for extracting a word or phrase relating to the control of the controlled device includes at least one of a morpheme analysis means, a syntactic analysis means, or a semantic analysis means to extract the word or phrase.

Further, the control device includes a voice recognition means for recognizing the contents of voice input, a database, a word / phrase information table storing word / phrase information relating to control of the controlled device, and a function chain structure for performing meaning analysis. And a response information table that stores a search result of the database in the form of a function chain structure, and a processing device, wherein the processing device includes a morpheme analysis means, a syntax analysis means, Alternatively, at least one of the semantic analysis means is provided, and means for extracting a word / phrase related to the control of the control target device by referring to the word / phrase information table based on the recognition result by the voice recognition means, and controlling based on the extracted word / phrase A semantic analysis is performed by referring to the information table to search for a semantic network expression, Means for extracting control information response,
A means for generating a control signal based on the extracted control information and outputting it to a controlled device or a database; a means for storing response information in a response information table in the form of a function chain structure when there is a response from the database; When response information is stored in the information table, means for complementing the meaning network expression with the response information and extracting control information corresponding to the extracted word or phrase based on the complemented meaning network expression; A means for generating a control signal based on the control information and outputting the control signal to the control target device is provided. Further, a storage means for storing a semantic network expression which is a semantic analysis result is provided, and the semantic analysis result stored in the storage means is re-analyzed based on the voice recognition result for a new voice input.

Further, the voice input means is a microphone,
The microphone is connected to a voice recognition means via a wireless communication device or a telephone line to enable remote control of the controlled device. Further, the control target device is a combination of at least one device such as a VTR, a television, or an audio output device. Further, instead of the voice recognition means, at least one or more means of character recognition means, keyboard input means, button operation means, or menu selection means is provided to input a natural language.

[0009]

With the above means, a natural language that is used on a daily basis is input, and words and phrases related to the controlled device are extracted from the input, and control information is extracted from the control information table based on the extracted words and phrases. A control signal that is obtained and corresponds to this control information is generated, and the control target device is controlled by this control signal. Also, by inputting natural language,
It is possible to remotely control the controlled device. Also,
It is possible to perform a cooperative operation by combining a plurality of control target devices by inputting a natural language.

[0010]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS This embodiment will be described in detail below with reference to the drawings. This embodiment is a case where the present invention is applied to a computer system connected to a telephone terminal, a voice recognition device, a database system, a VTR and a voice synthesizer. In this embodiment, the database system, VTR, and voice synthesizer are called control target devices. Also, a TV is connected to the VTR to display the output video. Here, data information of the video recorded on the video tape (the reproduction start address of the video tape and the outline of the content of the video) is stored in advance in the database system. If the present invention is not applied, in order to retrieve the information in the database and operate the VTR, it is necessary to operate each separately. That is, it is necessary to search the database and operate the user according to a predetermined VTR operation procedure while embedding the database search result. The prepared language and command are referred to as an operation instruction language in the description in this specification. Further, in the present embodiment, it is assumed that the operation procedure (pressing order of buttons and the like) can be mutually converted as an operation instruction language having a one-to-one correspondence. In this embodiment, the natural language sentence by the voice uttered by the user is a sentence mainly instructing the operation (sentence instructing the operation of the database, the VTR, etc.). For example, the VTR operation “I want to see Ki Hanada” from the user's point of view (“Ki Hanada” is the sumo wrestler's name) and the like are also included in this operation instruction.

FIG. 1 shows an outline of processing in the embodiment of the present invention. First, a natural language sentence is input from the input device (step 1). For example, the user utters "I want to see Kihanada" in the handset of the telephone terminal. Next, voice recognition is performed on the basis of the input voice and conversion into a corresponding symbol string (step 2). For example, the voice recognition device performs matching of phonemes, accents, and intonations of the voice, and outputs the corresponding natural language character string "Kibanada ga Mitai." The voice recognition device only needs to output a one-to-one symbol string for a voice, such as “Takahana Daga ...” or “takahanad.
aga. ． ． "Or" abcdef ... ", and the output may be" Kihanada ga ... "As a result of Kana-Kanji conversion processing used in word processors, but the following processing For the sake of explanation, it is assumed that kana-kanji mixed sentences are output. Further, since the voice recognition device can store information in the storage area (buffer), the above-mentioned symbol string can be output even when the user utters the voice by dividing it by each sound or word. Since the storage area (buffer) is a technology well known to those skilled in the art, detailed description thereof will not be required. Furthermore, the voice recognition device may output a symbol string corresponding to each word according to the input. Next, a word information table 1 prepared in advance based on the input natural language character string
0 (specifically, refer to FIG. 3) is used to extract a phrase related to the control target device (step 3). The extraction of words is simply done by inputting natural language character strings and word information table 1.
This is achieved by matching the character string with the expression (heading) of the word 0.

Next, control information corresponding to the words and phrases extracted in step 3 is prepared in advance in the control information table 11
(Specifically, refer to FIG. 4) is used to convert into control information (step 4). The control information is information for identifying the control target device that outputs the control signal, information about the control signal that is the operation instruction content for the control target device, and limit information regarding the output of other control information. The conversion into control information is simply achieved by acquiring the control information corresponding to the extracted phrase from the control information table 11. A part of the control information may be abstracted as a slot of the template, and the value of the slot may be appropriately embodied according to the execution timing. For example, from the operation instruction in the natural language, (1) an operation instruction language for searching a “VTR control start address” under a search condition “a wrestler name is Takahanada” with respect to a database system which is a control target device (for example, SQL (structured query) in relational database system
language)), (2) an operation instruction language (a specific command example will be described later) corresponding to “play from the start address value” for the VTR that is the control target device, and (3) control target It is converted into an operation instruction language (a specific example will be described later) corresponding to the natural language sentence “Please wait a moment” for the voice synthesizer, which is a device. In this way, in response to the input natural language sentence, the operation instruction language for controlling a plurality of control target devices (or systems) in cooperation is output. Here, in the control signal (or command) for controlling the VTR, there is a part depending on the search result of the database called "start address value", but this is set as the slot value, and when the signal is actually output to the VTR. , After the search from the database is completed, the slot value is filled in based on the search result and then output. Therefore, in the control information, “limit information regarding output” (see FIG.
35) in the control information table 11 is given that the control information is output after the control of the database is completed.

Next, based on the control information obtained in step 4 and the response information table 12 (specifically, see FIGS. 5 and 7) for storing the response information from the connected system, It is judged whether or not there is a control signal to be output by referring to the "limit information regarding output" 35 in the control information (step 5). If there is a control signal to be output, the process branches to step 6, otherwise to step 8.
In the specific example, the database search command and the voice output command to the voice synthesizer are control signals to be output, and the control signal of the VTR has not completed the database search.
It is determined that the control signal should not be output. If the output of the control information may be controlled in order (that is, not in parallel) in the “limit information regarding output” 35 or if it is independent of the order, the determination in step 5 is skipped and It is possible to branch to step 6 for each control information and process it. However, in general, control depends on the order, so control signals can be output in parallel.
That is, it is more desirable to determine whether or not the situation (meaning that the controlled device can be executed) is in a situation (step 5).

If there is a signal to be output in step 5, a control signal is generated from the control information obtained in step 4 (step 6). That is, the control signal information obtained in step 4 (34 in the control information table 11 in FIG. 4)
When the slot contains a slot, the response information table 12 is referred to, the slot value is substituted, and this is used as a control signal. If the control signal information 34 obtained in step 4 does not include a slot, this is used as it is as a control signal. Step 6
The control signals generated in step 1 are output to the corresponding controlled devices (step 7). The controlled device is controlled based on the control signal output in step 6. If there is no signal to be output in step 5, it is judged whether or not the series of processes should be ended (step 8). If not, the process branches to step 5.

If the voice input from the user is repeated, the above series of processing is repeated corresponding to those inputs.

The word / phrase information table 10 is a notation of words / phrases related to control (30 in the word / phrase information table 10 in FIG. 3).
(That is, a headline in the dictionary) and a semantic label (31 in the phrase information table 10 of FIG. 3) are stored. This is all that is required when extracting words by simple string matching. Morphological analysis, syntactic analysis,
The case of performing advanced natural language processing such as semantic analysis will be described later. The control information table 11 (see FIG. 4) is the meaning label 31 of the phrase information table 10 (see FIG. 3).
It is a table that stores control information corresponding to. The control information includes information for identifying a control target device that outputs a control signal (specifically, a control target device name or ID), information on a control signal that is the operation instruction content for the control target device, and other information. And limit information regarding output of control information. The case of performing advanced natural language processing such as disambiguation, semantic analysis, and discourse analysis will be described later. When the input natural language sentence includes a compound word (which has a meaning different from the meaning obtained from each word heading due to co-occurrence of a plurality of words), the plurality of headings are converted into one control information. To do this, add a word in the headline (“Kibana,
Look "), and when a sentence including all of those words is input, the corresponding control information is output (hereinafter, the case of this method will be described). Or, as an alternative, it is possible to prepare the idiom expression itself (“I want to see Kihanada”) as a headline, but when the system becomes complicated, the number of headings to be registered increases in combination, so a limited system Not valid except in the case of. As a further alternative, or exceptions may be made by linking to each heading related to the heading. Response information table 1
Reference numeral 2 is a table for storing control information and response information from a system connected thereto. The response information is information for identifying the control target device that has output the response signal (specifically, the control target device name or ID, etc.) and the information of the response signal from the control target device. These pieces of information are stored in the response information table 12 in step 14.

The above word / phrase information table 10, control information table 11, and response information table 12 can be managed in the same meaning network expression format by using a function chain structure (described later), and in that case they are unified. It is possible to manage detailed data. In this case, of course, it is not necessary to dare to use different names (word information table, control information table, response image method table). However, unlike the other control information table 11 and the response information table 12, the word / phrase information table 10 is large in scale, and the frequency of accessing the data stored therein is relatively low. It is desirable not to use the same meaning network representation format because it is possible to use computer resources more effectively than to use main memory. Although the semantic network expression format is used for the sake of convenience in the description below, it can be logically replaced by a table expression and may be treated as a table inside the computer.

In another process (or daemon in Unix), it responds to the response signal from the controlled device. That is, it is determined whether or not there is response information from the controlled device to be stored (step 13). If there is response information to be stored, the process branches to step 14, and if not, the process branches to step 13. If there is no processing to be stored, wait a while and then step 13
It is preferable to make the above judgment because the load on the computer is lightened. In this embodiment, in order to deal with the asynchronous response of each controlled device, step 13 and step 14
Is a separate process, but the response from the device to be controlled is stored in an appropriate buffer, and step 13 and step 14 are performed at an appropriate timing (for example, immediately before step 5) to perform multi-process (or The present invention can be implemented even in a computer environment in which multitasking cannot be performed. If there is response information to be stored in step 13, the response information is stored in the response information table 1
2 (step 14). In the specific example, when the search of the database is completed, the search result table is returned when the search is successful or successful. However, the table of the success or failure of the search and the inspection result does not necessarily have to be returned at once. In addition, the search result may be returned as (single) cell data instead of the table format. In this case, in the control information generated in step 4,
Instead of (1), it is described in the control information table 11 so that (4) an operation instruction language for a search condition and a search instruction and (5) an operation instruction language for sequentially inquiring search results one by one are generated. Just keep it.

FIG. 2 is a hardware configuration diagram in this embodiment. From the processor 15 and the storage device 16 connected to it, the voice input device (such as a telephone terminal handset or microphone) 17, the voice recognition device 18, the database system 19, the VTR 20, the television 21, the voice synthesizer 22 and the voice output device 23. Composed. Moreover, when connecting each device, it does not matter if they are connected via a communication modem or a communication line. A well-known technique may be used for a communication modem, a communication line, or a communication protocol (protocol) necessary for using them, and therefore detailed description thereof will be omitted in this specification. The voice input device 17 and the voice recognition device 18 may be combined on the device. Similarly, the voice synthesizer 22 and the voice output device 23 may be combined. Further, all or part of the database system 19, the voice recognition device 18, and the voice synthesis device 22 may be stored in the storage device 16 as software inside the computer. In addition, the control target devices to be connected are the database system 19 and the VTR
And the speech synthesizer 22, a decision support system,
It may be connected to an expert system, operating system, schedule management system or the like. The storage device 16 includes a natural language analysis program 24, a work area 25, a phrase information table 10, a control information table 11,
The response information table 12 is stored.

From FIG. 2 and the above brief description, those skilled in the art will recognize that the hardware for implementing the present invention may be a computer. The computer is not limited to a large computer, and may be a workstation or a personal computer. Natural language analysis program 2
4 can be executed, the phrase information table 10, the control information table 11, the response information table 12 and the work area 25 can be stored. It may be a computer configuration of a distributed system. The connection between the computers may be functionally connected via a storage medium, but this is not preferable due to the nature of the present invention.

Here, the natural language analysis program 24 is a program for executing a series of processes starting from step 1 shown in FIG. Also, steps 1, 2, 3, 4, 5,
The program that executes the processes of 6, 7, and 8 and the program that executes the processes of steps 13 and 14 are
A multi-process (multi-tasking) may be used and a software structure in which each process is executed in each process may be used. When all or part of the database system 19, the voice recognition device 18, and the voice synthesis device 22 are executed as a software program, the process may be executed corresponding to the output of the control signal in step 7, but the program configuration In addition (from the viewpoint of development efficiency and execution efficiency), a process (task or program) that actually executes a process (task or program) may be executed. The meaning of step 7 in the natural language analysis program 24 means the output of the control signal and the start of the execution accompanying it, and does not necessarily mean the end thereof.

FIG. 3 is a specific example of the word / phrase information table 10 in this embodiment. The word / phrase information table 10 is a table that stores a word / phrase notation 30 (that is, a heading in a dictionary) related to control and a meaning label 31. Specifically, the meaning label 31 "Kihanada" corresponds to the word 30 "Kihanada" related to database control, and
The meaning label 31 "VTR play" is stored in correspondence with the control word 30 "view" and "play".

As in this example, the word notation 30 and the meaning label 31 may be exactly the same character string. In this case, the meaning label 31 in the following description of processing
The phrase notation 30 can be used instead, and the amount of data registered in the phrase information table 10 can be reduced. In addition, the notation 30 of a plurality of phrases (“see” and “play” in the above example) is one meaning label (“VTR in the above example”.
Playback ”) may be supported. In this case, it is indicated that the notations 30 of a plurality of terms are treated in the same manner in the processing of step 4 and subsequent steps, whereby the data registered in the control information table 11 is used for the synonyms of different expressions (or different notations). The amount of can be reduced. Further, the meaning label 31 may be a simple identification symbol such as “A001”. For the sake of clarity, the explanation will be given using the semantic labels as described above.

FIG. 4 shows a concrete example of the control information table 11 in this embodiment. The control information table 11 is a table that stores a semantic label 32 that has a one-to-one correspondence with the semantic label 31, a control target device name 33, control signal information 34, and output restriction information 35. In particular,
Corresponding to the meaning label 32 “Kibanada”, the control target device name 33 “database” and the control signal information 34 “se”
lect start address value from sumo wrestling recording table where wrestler name = “Kibanada”, further, “speech synthesizer” of control target device 33, control signal information 34
It stores the “show show”.
As in this example, a plurality of combinations of the control target device name 33, the control signal information 34, and the output limit information 35 may correspond to one meaning label 32 (“Kibanada” in the above example). Absent. In this case, the description indicates that a plurality of control signals can be processed in parallel, corresponding to the meaning label 32 “Kibanada”. Of course, the plurality of control signals may be sequentially executed instead of in parallel. In addition, the meaning label 32 corresponds to “VTR playback”, the control target device name 33 “VTR”, and the control signal information 3
4 "JP (start address value)" and "PL (pla
y) ”and the output-related restriction information 35“ database search end ”are stored.

FIG. 5 is a specific example of the response information table 12 in this embodiment. The response information table 12 is a table that stores a response information slot name 36 and a response information slot value 37. Specifically, in response to the “database search end” of the response information slot name 36, “incomplete” as the initial value of the response information slot value 37 corresponds to the “start address value” of the response information slot value 36. The value of the information slot value 37 is not stored. FIG. 6 is a sumo wrestling recording table 38 which is a specific example of the table stored in the database system 19 in this embodiment. The sumo wrestling recording table 38 is a table that stores wrestler names 39 and start address values 40. Specifically, a start address value 40 of "305" is stored corresponding to the wrestler name 39 of "Kibanada".

In the following, the content of the process will be described in detail with reference to FIGS. 1 to 7 based on more specific data. A workstation is used as the processor 15 and the storage device 16 in this embodiment. As a concrete example, the creative workstation 2050 (manufactured by Hitachi, Ltd.), the database system 19 is a relational database system RDB2 (manufactured by Hitachi, Ltd.), VTR2 connected as software on the workstation.
0 will be described as a device that can be controlled using a control signal.
The telephone receiver of the voice input device 17, the television 21 connected to the VTR 20, and the speaker of the voice output device 23 may be known ones, and need not be particularly described. The voice recognition device 18 is a device that outputs a predetermined signal in response to a voice-based signal, and the voice synthesis device 22 is a device that outputs voice in response to a control signal. However, since the present invention is characterized by inputting voice and outputting a control signal, the above device may have a predetermined function and can be replaced with other products. For example,
Due to the performance and production cost, the present invention can be implemented even in a device in which the intonation analysis function is removed from the voice recognition device and the accent imparting function is removed from the voice synthesis device. Further, for the same reason, the present invention is applicable to an apparatus that is controlled based on a control signal, such as an air conditioner or a microwave oven. Such applicability will be apparent from the description below.

First, "I want to see Ki Hanada." Is input by voice from the telephone receiver (step 1). A microphone may be used for inputting voice without using the telephone handset. Next, the voice input in step 1 is converted into a corresponding symbol string by the voice recognition device. Here, it is assumed that the natural language sentence "I want to see Kihanada." Is output. The voice input device and the voice recognition device may be physically combined. When using a widely used telephone receiver, it is desirable to provide the voice recognition device as a separate device. In this case, a voice recognition device may be connected to each telephone handset, or may be connected to a computer that executes the following steps. In addition, the method of speech recognition may be input by the user by dividing into phonemes, words, phrases, or sentences. In this case, the speech recognition apparatus finally obtains an output equivalent to the natural language sentence "Kihanada ga Mitai." You can If the result of the voice recognition is a phonetic symbol or a Roman character or Kana, Kana-Kanji conversion may be appropriately performed. However, it is not essential in the present invention to convert a kana-kanji mixed natural language sentence, and any character string can be used as long as it is a symbol string that can be matched with the word notation 30 of the word information table 10. It doesn't matter.
The voice recognition device may use a well-known device (such as commercially available). However, it is clear that a device that is functionally homogeneous, that is, a device that outputs different symbol strings according to a plurality of different voice waveforms is sufficient for implementing the present invention. Ah As will be apparent from the description below, this step 2 does not necessarily need to output this character string. In other words, a fragment of a sentence that excludes "ga" and "i."
Alternatively, the following processing can be continued even if another expression such as "Um, replay Kihanada." Is output. This means that the present invention can be used even when the voice cannot be heard because the voice is weak, the input sentence includes an unknown word, and the sentence is grammatically non-sentence. is doing.

Next, in step 3, based on the natural language character string "Kibanada ga Mitai." Input in step 2,
A word information table 1 prepared in advance as shown in FIG.
0 is used to extract a word or phrase related to the control target device.
Specifically, whether or not each character string of the word notation 30 (that is, “Kibanada”, “mi”, “reproduction”) is included in the input natural language character string “Kibanada ga Mitai.” Explore. As a result of the search, the meaning label 31 (that is, "Kibanada" and "VTR playback") corresponding to the character string of the collated word notation 30 (that is, "Kibanada" and "mi") is output as the output of step 3, and the work is performed. Remember in the area.
When the number of word expressions 30 increases, the word expressions 30 that should not be matched may match. For example, if the input natural language sentence contains the string "discovery",
The “mi” of the word notation 30 matches. In such cases,
Instead of the algorithm of this step, morphological analysis or syntactic analysis may be performed on the input sentence to improve the natural language analysis accuracy.

Next, in step 4, the meaning label 31 extracted in step 3 (that is, "Kibanada",
The control information corresponding to “VTR reproduction”) is prepared by using the control information table 11 prepared in advance as shown in FIG.
Convert to control information. That is, the control information (1) and the control information (2) are obtained from the meaning label 31 “Kibanada”. That is, the control information (1) is “database” of the control target device name 33 and “sel” of the control signal information 34.
ect start address value from sumo wrestling recording table where wrestler name = “Kibanada”, and control information (2) is “speech synthesizer” of the control target device 33, and “show show”. In addition, from the meaning label 31 “VTR playback”, the control information (3), that is, the control target device name 33 “VT” is displayed.
R ”,“ JP (start address value) ”of the control signal information 34
And “PL (play)”, output limit information 35
"End of database search" is obtained. In addition, VTR
In the control signal information 34 for controlling, there is a part called "start address value" depending on the search result of the database, but this is a slot, and the process of substituting the actual value into this slot is performed in step 6 (described later). ). The control signal information 34 of the control information (3) consists of two signals (that is, “JP (start address value)” and “PL”), but these control signals are based on the control information (3). In this case, the data is output to the control target device (that is, VTR) in this order.

Then, in step 5, based on the response information table 12 and the limit information 35 regarding the output in the control information (1) to (3) obtained in step 4,
Determine whether there is a control signal to output. Specifically, it is assumed that the control information (1) and (2) which are not restricted are output, and the control information (3) which is restricted is referred to the response information table 12 of FIG. Since the value of the “database search end” of the response information slot name 36 of the response information table 12 is “unfinished”, it is judged that it should not be output. That is, the control information (1) and (2) branches to step 6, and the control information (3) branches to step 8. The control information (1) to (3) may be processed in parallel. As will be apparent from the description below, in the present embodiment, the processing for the control information (3) is eventually processed based on the result of the control information (1). It does not matter if the processing is performed sequentially. Further, when the processing is not a parallel processing system, whichever of the control information (1) and (2) is processed may be performed first. According to the present invention, it is possible to process the control information (2) after the control information (3), but this is not desirable from the point of control. This is because the control information (1) is a control of the database that generally takes a long time to control, and the control information (2) is accompanied by a voice message to the user that it will take a long time ("Please wait a moment."). Because it belongs to Therefore, it is desirable that the control information (2), (1), and (3) be carried out in this order in a processing system in which all processes (or multitasking) are impossible.

When the control information (3) is branched to step 8, first, in step 8, it is judged whether or not a series of processes should be ended. Specifically, since there is control information that has not been output yet, it is determined that the process should not be ended, and the process branches to step 5. Next, in step 5, it is determined whether there is a control signal to be further output. Specifically, for the control information (3) that has not been output yet, refer to the response information table 12 of FIG. 5, and refer to the response information slot name 3 of this response information table 12.
Since the response information slot value 37 corresponding to “database search end” in No. 6 is “unfinished”, it is determined that it should not be output. As a result, the process branches to step 8. As can be seen from the above description, regarding the control information (3), unless the response information slot value 37 is updated,
Steps 5 and 8 will be repeated. In terms of processing, this step may be repeated, or this step may be temporarily interrupted and restarted when the response information slot value 37 is updated.

Now, the description of the processing flow of the control information (1) and (2) branched to step 6 will be continued. In step 6, since there is no slot in the control signal information 34 of control information (1) and (2), nothing is done. next,
In step 7, the control signal information 34 "se" is sent to the database system 19 for the control information (1).
lect start address value from sumo wrestling recording table where wrestler name = “Kibanada” ”is output.
This control signal information is a kind of SQL (Structure).
d Query Language), and based on this, the database system 19 searches the sumo wrestling recording table 38 for the start address value 40 of the record data such that the wrestler name 39 is "Kiwanada", and outputs the database search end signal. "305" of the start address value 40 is returned. A well-known technique may be used to implement the database system 19. Depending on the type of database system, the return value may be the number of records, table format data, or simply storing the data in a file in the operating system, etc. Returning a response equivalent to that described above can be realized by a very easy program (or device), and thus the explanation here will not be required.

Further, in step 7, for the control information (2), the control signal information 34 "SHOW PLEASE START" is output to the voice synthesizer 22. Based on this, the voice synthesizer 22 outputs the corresponding voice to the voice output device. The voice synthesizer 22 may add accents, intonations, and appropriate pauses in correspondence with the control signal information 34. The speech synthesizer 22 is preferably a regular speech synthesizer, but may be a recording synthesizer or a tape recorder that outputs a voice that is pre-recorded in correspondence with the control signal information 34 "SHOW PERFORMANCE". .

The processing of another process (or daemon in Unix) corresponding to the response signal from the controlled device will be described. In step 13, it is determined whether or not there is response information from the controlled device to be stored. Specifically, until there is a response from the database system 19 corresponding to the above control information (1), it is determined in step 13 that there is no response information from the controlled device to be stored, and the process branches to step 13. . That is, the determination process of step 13 is repeated. Control information (1) above
When there is a response from the database system 19 corresponding to, it is determined that this is response information from the control target device to be stored, and the process branches to step 14. In step 14, the response information from the database system 19 is stored in the response information table 12. Specifically, with respect to the signal indicating the end of the database search, which is the response information, and “305” of the start address value 40, the response information slot corresponding to “database search end” of the response information slot name 36 in the response information table 12 “End” is stored in the value 37, and “305” is stored in the response information slot value corresponding to the “start address value” of the response information slot name 36.
To store. FIG. 7 shows the response information table 12 at this time.
Is.

Therefore, in step 5 in the control information (3) in which the steps 5 and 8 are repeated, the restriction information 35 regarding the output in the control information (3) is obtained.
Then, it is determined whether or not there is a control signal to be output, based on the response information table 12 updated as described above. Referring to the response information table 12 of FIG. 7, the control information (3) should be output because the value of the “database search end” of the response information slot name 36 of the response information table 12 is “end”. to decide.

Next, in step 6, since there is a slot in "JP (start address value)" of the control signal information 34 of the control information (3), the response information table 1
Substitute the slot value with reference to 2. Specifically, the response information slot name 3 to be compared with the slot name (that is, the start address value) included in the control signal information 34
The response information slot value 37 (that is, "305") corresponding to the "start address value" of 6 is replaced with the position of the slot of the control signal information. That is, “JP305” is obtained as the control signal information 34 of the control information (3). Next, in step 7, VTR2 is applied to the control information (3).
0 to control signal information 34 “JP305” and “P
L ”is output in this order. VTR20 is "JP3
05 "corresponding to the corresponding start address (ie,
The video tape is aligned with the position of the address 305 which is assigned to the video tape in advance, and the reproduction is started corresponding to "PL".

Next, at step 5, it is judged whether there is a control signal to be further output. Specifically, since all control information has been output, it is determined that there is no control signal to be output, and the process branches to step 8. Next, in step 8, since all the control information has been output, it is determined that there is no control signal to be output, and the series of processing ends.
Further, if the voice input from the user is repeated, the above series of processing is repeated corresponding to those inputs.

As described above, it is possible to provide a control device (eg, a remote controller) that analyzes a natural language sentence by voice and, based on the analysis, remotely operates a plurality of control target devices located apart from the user by interlocking them. Becomes In addition, as described in the present embodiment, a plurality of control targets (that is, the database and the VTR), such as "I want to see Takahanada."
It is possible to provide a control device that specifies a control target of a voice input in which is not explicitly specified and controls the control target in a predetermined sequence.

In the following, as a modification of the above-mentioned embodiment,
Further, regarding the case of performing advanced natural language processing, FIG.
~ It demonstrates in detail using FIG. In FIG. 8, FIG.
The same reference numerals as those in the drawings from FIG.
Here, what is largely different from the above-described embodiment is that instead of steps 3, 4, and 6,
It is being processed in steps 41-46. This makes it possible to realize advanced natural language analysis processing,
By reparsing the already parsed statement, more appropriate control can be performed, or for the operation already executed,
An operation to correct (or correct) it can be added. That is, in step 41, advanced natural language analysis such as morphological analysis and semantic analysis is performed. In step 42, based on the meaning of the sentence analyzed by the natural language,
Determine if the sentences entered so far need to be reparsed. If you need to re-analyze, step 4
Branch to 3. If there is no need to re-analyze, branch to step 44. For example, a negative response sentence such as “No”, an operation instruction sentence such as “analyze again” to instruct re-analysis, a sentence that adds a condition such as “However, it is sumo wrestling last year”, Day 15 means a sentence that adds knowledge, such as "use the synonyms."

When these sentences are interpreted, in the case of a negative response sentence or an operation instruction sentence for instructing re-analysis, the next candidate of the semantic analysis result of the previous sentence is searched. In other words, it is possible to find another analysis candidate by extending the distance of the searched route or relaxing the evaluation score criterion, which is a termination condition (described later). In the case of a statement that adds a condition or knowledge, the function chain structure obtained as a result of analysis is immediately used for the control information table 1.
1 and parse the previous sentence again, including this knowledge. A processing procedure for performing these processes (for example, a process of holding the previous sentence and its analysis result as a history and recalling it in the memory if necessary) and individual knowledge description methods (specific examples described later and Although a detailed description of substantially the same semantic structure conversion rules, etc.) will be omitted, those skilled in the art can easily infer from the following detailed examples.

In step 43, the sentence to be processed is re-analyzed with reference to the phrase information table 10 (or the control information table or the response information table 12). The processing procedure for natural language analysis in step 43 is step 4
It does not matter if it is processed in the same manner as the first method. In step 44, if the results analyzed in step 41 and step 43 need to be further updated by referring to the response information table 12, the process branches to step 45, and if not, the process branches to step 5. In step 45, the response information table 12 is referred to, and step 4
Update the semantic expression obtained in 1. In step 46, a more general-purpose knowledge description (function chain structure described later) in the control information table 11 and the response information table 12 is performed.
In order to cope with the above, a concrete function (a concrete example will be described later) is prepared and the function is evaluated. In the following processing steps, the function chain structure, which is a semantic expression format (or knowledge expression format) necessary to achieve high-level natural language processing, will be described first. The description of the functional chain structure is described in detail in the specification of Japanese Patent Application No. 4-152946.

In the basic structure of the function chain structure shown in FIG. 9, the child node 101 of the node 100 showing the function has its function 10
Nodes (inputs 1 to n) related to input data and conditions to 0, and a parent node 102 of the node 100 indicating a function
Is a node (outputs 1 to m) related to output data and conditions from the function. In the figure, nodes are represented by enclosing them with a rectangle (nodes related to input data and conditions) or hexagons (nodes showing functions), and expressing the relationship of nodes by connecting with directed links from child nodes to parent nodes. ing. In the following, a node indicating a function, particularly a node indicating a function of an operation instruction, may be referred to as an operation instruction node. When expressing that the output result of a certain function A becomes the input data or input condition of another function B, it is expressed by making the node of the output result of that function A the input data or input condition node of the function B. To do. A graph structure such as a tree structure or a semantic network may be used to represent the function chain structure. For convenience of processing, a root node (root node) or a node that modifies and limits a certain node may be provided in the actual function chain structure. Each node has the following information. It is a node ID (identifier) for identifying a node, (2) information on the link relationship of other nodes, and (3) an attribute (attribute) indicating the purpose of the node itself. Further, it may have the following information. (4) Information on natural language expression, (5) Information on adjunct words, (6) Information on functional labels, or information on semantic labels meant by these information (data), (7) Given natural language sentence (8) Information about the position on the notation, (8) Information about the range accompanying these information (data), (9) Of these information (data),
Information about the case relationship in the corresponding natural language sentence, etc. Further, (10) the ID of the function chain structure originally belonging to the node, (11) the type of the function chain structure of (10) (that is, whether the function chain structure is used as a rule for converting the semantic structure, or the system definition (12) the importance of the link relationship between nodes, whether it is performing knowledge, describing the knowledge of the application target world, expressing an actual operation target, expressing a virtual operation target, etc.) (In order to reduce the total number of rules that need to be defined by also using the rules for converting the semantic structure, it is necessary to define whether a link to a node is "arbitrary" or "required". Not necessary data), etc. These pieces of information will be described as needed in the following specific examples.

Nodes are classified into the following types according to the attribute indicating the purpose of (3). The first is a system node. This is a node used to process (manage, transform, etc.) the function chain structure. For example, in the case of expressing with a tree structure, there is a node (giving "root" as an attribute) for expressing a root. The second is a node indicating a function. This includes "features" as attributes
Is given. Further, the node indicating the function is further subdivided into an operation instruction node and a node other than the operation instruction node.
The operation instruction node is a node for expressing a function of instructing an operation of a control target device (or application). A method (command, etc.) for actually instructing operation of a control target device (or application) is described in the consequent part of the semantic structure conversion rule (described later) defined for this operation instruction node. Of the nodes indicating the functions, the nodes other than the operation instruction node are nodes for expressing a semantic relationship between the nodes regarding data and output described later. For example, a node for expressing the semantic relationship of the superordinate and subordinate concepts (giving "isa" as a function label) and a node for expressing the relationship between the target and its attribute (giving "hap" as a function label). ), And a node for expressing a semantic relationship of equivalence (giving “equ” as a functional label). The third is an operation target / condition node. This is a node for expressing input data or an output result that is an operation target or a condition for a node indicating a function. For example, a node for expressing the operation target (giving "operation target" as an attribute), a node for expressing the condition itself (giving "condition" as an attribute), and an attribute that is a component of the condition There are a node for expressing a name (giving an "attribute name" as an attribute), a node for expressing an attribute value that is a component of a condition (giving an "attribute value" as an attribute), and the like.

FIG. 10 is a diagram for explaining a function chain structure expressing an operation target and conditions. In order to express the condition that the natural language expression "S is V" by the function chain structure,
Let “isa” be the node 106 indicating the function, and the node 10
The "attribute name S" node 107 and the "attribute value V" node 108 are combined as child nodes of No. 6, and the "condition C" node 105 is connected as a parent node of the node 106. For example,
If the natural language expression is “the sumo wrestler's name is Takahanada”, the function chain structure is the node 106 “is” indicating the function.
“A”, a node 107 “wrestler name” indicating an attribute name, a node 108 “Kibanada” indicating an attribute value, and a node 105 indicating a condition are combined as described above.

When an expression in which the attribute name S is omitted such that the natural language expression is "V" is expressed by a function chain structure, the node 107 indicating the attribute name is connected as a dummy or the node 107 is used. May be omitted. Examples of specific natural language expressions in such a case include "is a spring place" and "Kibanada". Furthermore, when there are a plurality of corresponding attribute names S for one attribute value V, they may be expressed separately, or a plurality of nodes 107 indicating the attribute name S may be prepared to show them one function. It may be connected to the node “isa”. As a specific example of such a case, in the case of expressing the character string data “Kibanada” stored in the first field of the database,
Two nodes of "first field" and "character string data" are prepared as the node 107 indicating the attribute name.

Representation of an operation target in the function chain structure (this operation target may be an operation target actually existing in the system or an operation target assumed during the inference) In this case, the operation target is expressed by a combination of conditions indicating the attribute of the operation target. That is, in order to represent the operation target with a function chain structure, "h
“Ap” is a node 104, the “condition C” node 105 is connected as a child node of the node 105, and the “operation target T” node 103 is connected as a parent node of the node 105. For example, consider an operation target whose natural language expression is “Rikishi is Kihanada”. In other words, a certain record in the database whose wrestler name is Takahanada is expressed as a function chain structure as an operation target. This is in addition to the function chain structure composed of the nodes 105 to 108 described in the above example, the node "hap" indicating the function and the node 1 indicating the operation target.
And 03 are combined as described above. In addition,
Generally, an operation target has a plurality of conditions (attributes), but when expressing this in a function chain structure, a node 105 indicating each condition is connected as a child node of the node 104 indicating a function. As an example of a specific natural language expression in such a case, there is “Rikishi is Kihanada and the search target table is a sumo wrestling recording table”. However, it is assumed that one operation target is indicated by this example sentence. In this way, by expressing the operation target and the condition in the same form as the function chain structure, as described above, the function chain structure of the operation target can be described including the function chain structure of the condition. It is possible in terms of expression that the expression of the condition in natural language can be interpreted as the anaphoric expression of the operation target including the condition. For example, it is possible in terms of expression to interpret that the condition “Spring place” is an anaphoric expression (one of) of an operation target including the condition (for example, Takahanada's approach in the spring place). As described above, the function chain structure based on the above description is a semantic expression effective for interpreting the anaphora of the conditional expression anaphorizing a certain operation target.

11 to 19 are specific examples of the control information table 11 and the response information table 12 in this embodiment. Here, for the sake of easy understanding of the description, a schematic expression based on the above-described function chain structure will be used instead of the table expression. It will be understood from the detailed description so far that the data representation form in the computer having the function chain structure can be realized by using the table. Now, in order to analyze the meaning of a natural language sentence and generate a command, the following knowledge is required. That is, (1) knowledge about vocabulary and grammar (for example, a dictionary or phrase information table 10), and (2) knowledge depending on fields (for example, knowledge about retailing, medical care, sumo, etc.). In the example, knowledge based on the database structure that expresses this knowledge),
(3) Knowledge of the system (knowledge of upper and lower relations in the system, etc.), (4) Knowledge of controlled device (or application) (knowledge of correspondence between operation instruction content of natural language sentence and command format (hereinafter In particular, it is sometimes called the semantic structure conversion rule)),
(5) Knowledge that depends on discourse (screen information, input / output history, operation history, response information, etc.). A specific example in which the knowledge of (2) to (5) is expressed by the above-mentioned function chain structure will be described.

The knowledge about vocabulary and grammar in (1) is also expressed by using a function chain structure, and (2)-
By using the expression format equivalent to the knowledge up to (5),
The series of analyzes from morphological analysis to semantic analysis in step 41 and step 43 may be integrated to further improve the accuracy of natural language analysis by compounding knowledge. However, in this case, the processing speed becomes slower than that of the conventional technique in which the dictionary (or the phrase information table 10) is searched as a simple table, which is not preferable in practice.
In the control information table 11, the headline and the control information may indirectly correspond to each other. That is, the headline is the intermediate expression 1
It is sufficient that the intermediate expression 1 corresponds to the intermediate expression 2, and this is repeated repeatedly, and the intermediate expression n corresponds to the control information. Such correspondences (intermediate representation n to intermediate representation n
(Correspondence to +1) can be expressed in record units of a table, or can be expressed as a partial network expression (that is, a function chain structure). Since the idiom can be expressed by combining these functional chain structures (that is, branching (or merging) in the search route), each headline may be registered for each word.

FIG. 11 is a specific example of the function chain structure to be stored in the control information table 11 in this embodiment. That is, this is a specific example of knowledge that depends on the field classified into (2) of the above knowledge, particularly knowledge about a database. The knowledge that the first field of the sumo wrestling recording table, which is a record file in the database, is the wrestler name, the second field is the date of the action, and the third field is the start address value It can be represented by a chain structure. That is,
For this operation target, the condition 112 that the attribute name “first field” 114 is the attribute value “wrestler name” 115, and the attribute name “second field” 118 is the attribute value “approach date” 119.
Condition 116 that is and attribute name "third field"
The condition 120 that 122 is the attribute value "starting address value" 123, and the condition 124 that the attribute name "table name" 126 is the attribute value "Grand Sumo recording table" 127
It is expressed as an operation target “record file” 110 that has both. Here, the attribute name "first field"
The relationship between 114, the attribute value “wrestler name” 115 and the condition 112 is expressed by the function “isa” 113, and the relationship between the attribute name “second field” 118, the attribute value “approach date” 119 and the condition 116 is the function “is”. isa ”117, and the attribute name“ 3rd
Field ”122 and attribute value“ start address value ”123
The relationship between the condition 120 and the condition 120 is expressed by the function “isa” 121, and the relationship between the attribute name “table name” 126, the attribute value “sumo wrestling recording table” 127 and the condition 124 is the function “is”.
a ”125, and the condition 112, the condition 116, and the condition 1
20 and condition 124 and operation target “record file” 11
The relationship with 0 is represented by the function “hap” 111. In this way, knowledge about the database structure can be expressed by a function chain structure.

12 and 13 are specific examples of the function chain structure to be stored in the control information table 11 in this embodiment. That is, this is a specific example of knowledge about a system classified into the above knowledge (3), particularly knowledge about upper and lower relationships of the system. Knowledge of the superordinate and subordinate relationships that the first field is a kind of field that is a component of the record file can be expressed as a kind of condition by a function chain structure. That is, the condition is that the attribute name “field” 129 is the attribute value “first field”.
It is expressed as 130. Attribute name "field" 12
The relationship between 9 and the attribute value “first field” 130 is expressed by the function “isa” 128. Similarly, the second
Knowledge that a field is a field and knowledge that a third field is a field can be expressed. In this way, knowledge about the upper and lower relations in the system can be expressed by the function chain structure.

14 and 15 are specific examples of the function chain structure to be stored in the control information table 11 in this embodiment. That is, it is a specific example of knowledge about the thesaurus of vocabulary classified as the above knowledge (1). The knowledge that Kikanada is a wrestler's name is expressed as that the attribute name “wrestler's name” 138 is the attribute value “Kibanada” 139. The relationship between the attribute name “wrestler name” 138 and the attribute value “Kibanada” 139 is expressed by the function “isa” 137. Similarly, the knowledge that Chiaki is the work day is attribute name "work day" 1
41 is expressed as the attribute value “Chiakiraku” 142, and the relationship between the attribute name “approach date” 141 and the attribute value “Chiakiraku” 142 is expressed by the function “isa” 140. These pieces of knowledge are not necessarily stored in the control information table 11 because they are expressed as a chained function structure. This is because the knowledge that "Kibanada is the sumo wrestler" and "Chiaki Raku is the working day" are stored as column names and cell data values (or column values) in the sumo recording table of the database 19, respectively. It is because there is.

In order to simplify the description of the present embodiment, the function chain structure is given to the control information table 11 as described above, but information equivalent to this is symbolized and the word knowledge is stored in the word information table 10. May be given in advance as. Also, in this case, since this thesaurus relationship has a very simple relationship of column name and cell data value, the table in the database given this relationship is searched in advance to find the cell data value and the cell Word information table 1 with the column to which the data belongs as a superordinate concept
It may be stored in 0. Furthermore, the process related to knowledge acquisition from this database may be (semi) automatically maintained by a process provided separately. According to this method, knowledge can be expressed as a function chain structure for each one upper-lower relationship, so even if the thesaurus of vocabulary is not systematized, partial upper-lower relationships of the thesaurus are incorporated as a function chain structure, and It can be used for word analysis 41 (and reanalysis 43 of natural language sentences and conversion 4 into control information).

16 and 17 are specific examples of the function chain structure to be stored in the control information table 11 in this embodiment. That is, this is a specific example of knowledge about applications classified into the above knowledge (4), particularly, meaning structure conversion rules. The knowledge unique to the relational database system that a table is generated as a search result when a specified field is searched is the "RD" in the first half of the knowledge.
The conditional part of the semantic structure conversion rule that is "B search" and the consequent part of the semantic structure conversion rule that is "generation" of the latter half of the knowledge are represented by a function chain structure. That is, the condition part (see FIG.
6) is expressed as the condition 144 that the search condition is the attribute value “field” 146, the condition 147, and the function (operation instruction) that inputs the condition 150 “RDB search” 143. Here, the attribute value “field” 146 and condition 1
The relationship with 44 is represented by the function “isa” 145. Other conditions are the same. Further, the attribute name “search condition” or the like may be provided in correspondence with the attribute value 146. In this embodiment, since this attribute name node is not used, it is omitted for convenience of explanation. In the consequent part (FIG. 17), the first column $ FIELD (node 146) is $ CELL.
At (node 146), the second column, $ FIELD
(Node 149) is $ CELL (node 149) and the third column, $ FIELD (node 152), is $ CE.
It is represented by a function chain structure that it is a search result table of the LL (node 152). A detailed description of this function chain structure is apparent from FIG. 16 and is the same as the description so far, and will be omitted. In addition, the semantic analysis process (step 4
A node for connecting the condition part and the consequent part may be prepared for the convenience of processing in the network search of 1). In addition to the above, control information (command, etc.) for the control target device (or application system) is described in the consequent part of the semantic structure conversion rule. Specifically, the control target device name 170 is “database” and the control signal information 171 is “select”.
$ FIELD (node 146, node 149, node 1
52) from $ TABLE (node 146, node 149, node 152) where $ WHERE
(Node 146, node 149, node 152) "is described. In this way, the semantic structure conversion rule can be expressed by a function chain structure. Here, "$ CEL
L ”,“ $ FIELD ”,“ $ TABLE ”and“ $
"WHERE" is a function that returns an appropriate character string based on each argument node (hereinafter, these functions may be referred to as an embodying function in particular).

As will be described in detail later, for example, the node 15
2 is the "start address name", and the control information table 11 or the response information table 12 connected to the node 146 on the network has "Rikishi name" and "Kibanada", and the attribute name node in the function chain structure including the node. "table name"
Corresponding to the attribute value node "Grand Sumo Recording Table", and the control information table 11 or the response information table 12 connected on the network to the node 149 has "Effort Date" and "Chiaki Raku". There is an attribute value node "Grand Sumo recording table" corresponding to the attribute name node "table name" in the function chain structure including, and "" is displayed on the control information table 11 or the response information table 12 connected to the node 152 on the network. When there is a "start address value" and a corresponding attribute value node "305", "$ CELL (node 146)" returns "Kibanada". Also, at this time, "$ CELL (node 149)"
Returns "Chiakiraku". Furthermore, at this time, "$ CE
"LL (node 152)" returns "305". "$ FI
ELD (node 146, node 149, node 15
2) ”returns“ Wrestler's name, work day, start address name ”.
Further, at this time, "$ TABLE (node 146, node 149, node 152)" returns the "sumo wrestling recording table". Furthermore, at this time, "$ WHERE
(Node 146, node 149, node 152) ”returns“ wrestler name = “Kibanada” and approach date = “Chiakiraku” ”. Therefore, the control signal information after performing these processes is “select wrestler name work day start address name from sumo wrestling recording table where wrestler name =“ Kibana ”and work day =“ Chikakuraku ””. Note that, in FIG. 16 and FIG. 17, for convenience of description, there are three condition nodes, but the number of condition nodes may be two or less, or four or more. It is also possible to prepare only one structure below the condition node and dynamically copy and use it.

FIG. 18 is a concrete example of a function chain structure to be stored in the control information table 11 in this embodiment. That is, this is a specific example of knowledge about applications classified into the above knowledge (4), particularly, meaning structure conversion rules. The knowledge unique to the VTR system that the VTR is controlled according to a table having the start address value as data expresses the conditional part of the semantic structure conversion rule, which is the "VTR control" in the first half of the knowledge, in a function chain structure. That is, a function (operation instruction) that inputs the operation target 173 having the condition 175 that the attribute name “start address value” 176 exists
It is expressed as “VTR control” 172. Here, the relationship between the attribute name “start address value” 177 and the condition 175 is expressed by the function “isa” 176, and the relationship between the condition 175 and the operation target 173 is expressed by “hap”. For simplicity,
Assuming that there is no change in the discourse world depending on the result of this operation instruction, it is not necessary to describe the functional chain structure in the consequent part. Incidentally, when the system manages the interaction model and the internal user model, the knowledge (that is, the function chain structure) that needs to be added (or deleted) to the consequent part of this semantic structure conversion rule may be described. Absent. In addition to the above, control information (commands, etc.) for the control target device (or application system) is described in the consequent part of the semantic structure conversion rule. It is desirable that these pieces of control information are described in the order in which they should be processed. Specifically, the control target device name 170 is "VTR" and the control signal information is "JP $ DATA (node 177)", and the control target device name 170 is "VTR".
And control information in which the control signal information is "PL" are described. Here, “$ DATA” is a function for reification. Although details will be described later, for example, the node 177.
The attribute name is “start address value” and the attribute value is “305” in the response information table 12 connected to the network
, And the function chain structure including that node includes a certain operation target node, “$ DATA (node 177)” returns “305”. Therefore, the first control signal information after the above processing is “JP30
5 ”. Since the second control signal information does not include the embodying function, it remains “PL”. In this control information, the control target device name 170 is “V
Control target devices other than "TR" may be defined. For example, as the control target device name 170, "speech synthesizer",
When the control signal information 171 is additionally provided with “Taihen Omata Seita Shimashita.”, When two control signals are output based on this control information (step 7), two control target devices (that is, VTR and voice synthesis) are output. It is possible to control the devices simultaneously.

FIG. 19 is a concrete example of the function chain structure to be stored in the response information table 12 in this embodiment. That is, knowledge that depends on discourse classified into the above knowledge (5), in particular, knowledge about an operation target that actually exists (or has existed) such as screen information, history information, response information, and the like (hereinafter referred to as “knowledge”). Then, it may be referred to as “materialization data”). In this example, in the search result table in which the first column “Rikishi name” is “Kibanada”, the second column “Effort date” is “Chiakiraku”, and the third column “Start address value” is “305” is there. Details of this functional chain structure are also omitted because they can be easily inferred from the above description. It should be noted that when this function chain structure is expressed by a table, it becomes as shown in FIG.

As described above, by expressing the above knowledge (2) to (5) by the function chain structure having the same expression format, it is possible to easily perform the semantic analysis in which the knowledge of two or more kinds is appropriately linked. Become. Further, as compared to the case where the expression form of the knowledge used and the intermediate meaning expression are different as in the prior art, the combined use of the function chain structure as the meaning expression of the natural language sentence as in the present embodiment is an algorithm of the semantic analysis processing. And is efficient in terms of storage usage. Further, the fact that the various knowledge expressions have the same expression format can reduce the total memory consumption because the knowledge editor can be integrated. Further, compared with the case where a plurality of different knowledge editors are provided, the operation method is unified, so that maintenance man-hours and development man-hours are reduced.

FIG. 21 is a detailed flowchart of the natural language analysis (step 41) processing in this embodiment.
Knowledge of the fields, systems, controlled devices (or application systems), and discourse already explained is
It is expressed by a function chain structure, which is a kind of semantic network. Natural language analysis in the present embodiment, after performing a morphological analysis, to search on various knowledge depending on this functional chain structure,
This is a processing method in which the semantic network that became the search route immediately becomes an intermediate semantic representation of a given natural language sentence. That is, the algorithm for this network search finds a function chain structure having an operation target or a condition as a starting point and a function as an end point by searching a function chain structure which is various kinds of knowledge, and combining those function chain structures. This is an algorithm that uses a chained structure of functions as an intermediate representation. First, morphological analysis is performed (step 410). This processing may use the simple word / phrase extraction processing described in the previous embodiment of this embodiment, or may use the morphological analysis processing or keyword extraction processing of the related art. However, this processing needs to be summarized as a node for each clause used below, so any processing that substitutes for this may be used. In the case of semantically analyzing the second and subsequent candidates, the morphological analysis process (step 410) is duplicated in most cases. Therefore, it is possible to save the waste of processing by storing the information of the result of the morphological analysis before and reusing it.

Next, in the semantic analysis using natural language,
It is selected whether to obtain a candidate or the second and subsequent candidates of meaning (step 411). In order to obtain the n-th candidate of meaning, the intermediate results of analysis processing up to the (n-1) -th candidate (such as a footprint list (described later)) are saved. When obtaining the first candidate of meaning, the intermediate result of the analysis process (footprint list etc.) is initialized and the process branches to step 412. Second
When obtaining the candidates and later, the process branches to step 414. According to this processing method, when the system performs an analysis that is not intended by the user who inputs the input sentence, or when the system wants to present a plurality of analysis candidates to the user all at once,
It is possible to efficiently realize the process of obtaining the second and subsequent candidates of meaning. Further, by storing the intermediate result of the analysis process historically, it is not necessary to perform redundant and wasteful analysis when it is necessary to re-analyze from the previous sentence as in step 43. There is also.

Next, among the nodes assigned to each natural language element (generally equivalent to a clause) in the processing of morphological analysis (step 410), the operation instruction node and this operation instruction node are syntactically or written. A pair of operation target / condition nodes (generally a plurality) which are related to the above position is found (step 412). If there is a set that has not been analyzed in the following steps 412 to 416 in this set, the process branches to step 413. Otherwise, the natural language analysis process (step 41) is finished. In the processing from step 413 onward, of the set obtained in step 412, the operation instruction node is set to be a “node for ending the search”, the operation target / condition node is set to be a “node for starting the search”, or a predetermined node is set. Default (or dummy)
The following analysis may be performed by supplementing the node of (1) as the “node that ends the search” or the “node that starts the search”.

Next, the footprint list is emptied (NULL),
The footprints prepared for each "search start node" are added to the footprint list (step 413). “Footprint” is a data structure in which a search route is stored, and each footprint corresponds to storing one step in network search. The “footstep list” is a set of footsteps. Each footprint has the following information. That is, (footstep information a) one step before the footstep (however, when two or more footsteps merge, a plurality of “steps before one footstep” are stored accordingly. (Footstep information b) the presently searched (Footprint information c) Distance of search path This distance is the distance from the search start node to the node of (footstep information b). The calculation method of the distance is that the passed (or joined) node ( Or a link), a value that is appropriately weighted according to the number of links is accumulated. , It is possible to find an appropriate search route with priority. In short, it is the total number of nodes that have passed through as a distance calculation method. (Footstep information d) Evaluation score, which is a value obtained by accumulating the degree of conformity of this footprint with respect to the node that starts the search to the evaluation point of the footprint one step before. Footstep information e) A set of function chain structures that can be searched in the next step, that is, a set of available knowledge. This set of function chain structures performs subject focusing by increasing or decreasing according to discourse. , The discourse tracking can be facilitated. The main purpose of the network search is to perform a memory-efficient search by using the footprint defined in this way. The knowledge that can be used is generated or disappears non-deterministically, so it is necessary to duplicate the knowledge for each search path. Since the path repeats branching and merging, the amount of memory used for duplicating each piece of knowledge becomes enormous, so if the network search is stored step by step as in the present invention, the search path can be saved. The common search path that has been generated at the time of branching is efficient because it does not need to use memory for replication.

Next, if there is a "footprint" to be processed in the footprint list, the process branches to step 415, and if not, the process branches to step 412 (step 4).
14). For example, from the footprint list, the smallest distance (footstep information c) of the searched route is sequentially taken out. Next, the node (footstep information b) during the search for the footstep obtained in step 414.
Starting from, the network is searched to find a node one step ahead, and the footprint of the node is added to the footprint list (step 415). Step 415 will be described in more detail with reference to FIG. Next, it is determined whether or not the search end condition is satisfied, and if satisfied, a natural language analysis process (step 4
Complete 1). If not satisfied, the process proceeds to step 414.
(Step 416). The search termination condition is
There are the following: (A) When the distance of the searched route (footstep information c) reaches a certain value. (B) When there is a footprint of a node that matches the operation instruction node obtained by the analysis of the natural language. (C) When there is a footprint with a sufficiently high evaluation score (footprint information d). The end of the search is determined individually or in combination. For example, search until the distance of the search route reaches a preset value, and if the footprint with the highest evaluation score is found among the footprints of the nodes that match the operation instruction node obtained from the natural language sentence, the search ends. To do. This footprint is the first candidate for meaning. In the case of the second meaning candidate or less, the footprint having a lower evaluation score is set as the meaning candidate. Further, when it is necessary to increase the number of footprint candidates in order to obtain the second or lower candidate of meaning, the distance of the search route of the search termination condition (a) may be dynamically extended.

FIG. 22 is a detailed flowchart of the processing procedure of the network search (step 415) in this embodiment. First, in the process of step 4150, if the (footstep information b) node of the footprint obtained in step 414 (hereinafter, this may be particularly referred to as “node currently being searched”) is the operation instruction node , The process branches to step 4151; otherwise, if the node is other than the operation instruction node, the process branches to step 4152. Next, in the processing of step 4151, the function chain structure of the consequent part of the meaning structure conversion rule is added or deleted to (footprint information e). By adding to this (footprint information e), the function chain structure of this consequent part becomes a function chain structure that can be searched as the next step of the node currently being searched. Further, the deletion from the (footprint information e) means that the function chain structure of the consequent part is a function chain structure that cannot be searched as the next step of the node currently being searched. As explained above, it is possible to define multiple function chain structures in the consequent part, and define whether to add or delete according to each function chain structure. You can keep it. A special node may be provided to indicate these “addition” or “deletion”. It may not be necessary to explain the processing in this case. When the function chain structure of the consequent part is added, only the node that is defined in advance as a node for continuing the search among the nodes forming the function chain structure is added as a new footprint to the footprint list. In this embodiment, in particular, only the attribute value node included in the function chain structure is added to the footprint list as a new footprint. In addition, a node for structurally clarifying that the search is continued may be inserted between the operation instruction node and the node where the search is continued. After this, the process proceeds to step 4152.

Next, in the processing of step 4152,
If the node currently being searched is a node corresponding to a pronoun (“this”, “that”, or “there”) in the input natural language sentence, the process branches to step 4153, and if not, the process is performed. To step 4154. In this embodiment, the clause analyzed as a pronoun as a result of the morphological analysis in step 410 is treated as a kind of operation target / condition node. Next, in the processing of step 4153, from the function chain structure generated as a result of analyzing the natural language sentence up to now, the function chain structure in the response information table 12, etc. (Described below) is added to the (footstep information e) of the new footstep. When a function chain structure is added, as a new footprint only for a node that is previously defined as a node that continues the search among the nodes that form the function chain structure,
Add to footprint list. In the present embodiment, similarly to the previous case, only the attribute value node included in the function chain structure is added to the footprint list as a new footprint. After this, the process proceeds to step 4154.

Here, the function chain structure that a pronoun may anaphorate simply refers to the operation target node of knowledge (the above-mentioned knowledge (5)) depending on the discourse according to the pronoun “it” or the like. Refers to the functional chain structure that includes. However, in order to improve the accuracy of interpretation of pronouns, the pronoun "there" and the like may be restricted to point to a function chain structure including "place" in the attribute name node (or attribute value node). . Also, of the knowledge that depends on discourse, in particular, history information (how long before the natural language sentence (or operation target)
Information, etc.), the likelihood of anaphora may be prioritized. Specifically, prioritization is performed so that there is a high possibility that the recently-referenced function chain structure will be adapted. In the present embodiment, the smaller the value of the (footprint information c), the more the meaning interpretation is output. Therefore, the (footprint information c) is decided based on the priority. That is, "0" is set as the initial value at the beginning.
Instead, a small value (short distance) is assigned to the footprint of the function chain structure that should be given high priority, and a large value (long distance) is assigned to low priority.

Next, in the processing of step 4154,
A node which is a component in the set of the functional chain structure of (footprint information e), or a node of (footstep information b) of a footprint that has already been processed in the footprint list, and a node currently being searched. Determine if the node matches. If so, the process branches to step 4155; otherwise, the process branches to step 4156. Whether or not they match is determined by matching, similarity, or disagreement of the information (described above) described in each node. Briefly, if the function label (or the semantic label) is the same in the node to be compared (that is, the node currently being searched) and the node to be compared, it is determined to be a matching node. Next, in the processing of step 4155, step 4154
Create a new footprint based on the node matched in and add it to the footprint list. At this time, in order to structurally clearly indicate that the matching source node and the matching destination node are connected, a node having a function label of "equ" may be placed in the middle.

Next, in the processing of step 4156,
A node in the direction that is one step further from the node currently being searched for toward the link connection is added to the footprint list as a new footprint. However, footprints having search paths that go back are not preferable in terms of processing efficiency, so do not add them to the footprint list. After this, the process proceeds to step 416. The connection direction of the link is the following direction. That is, the direction from the "hap" node to the operation target node. Direction from the "isa" node to the condition node or to the attribute name node. "Isa" from the attribute name (or attribute value) node
(Or "equ") direction to the node. The direction from the condition node to the "hap" node or the operation instruction node. The direction from the operation target node to the operation instruction node. When creating a new footprint in step 4151, step 4153, or step 4155, (footstep information a) is the pointer of the previous footstep, and (footstep information c) is an appropriate cumulative total of positive values. Is an addition of "1")
In (footstep information d), the degree of matching between this node and the start node is scored, and the total of the score and the evaluation score of the previous footprint is used to calculate each footprint information, and a new footprint is created. To do. However, the degree of conformity is calculated based on the coincidence, similarity, and disagreement of the information held by the two nodes. For example, in the information held by the two nodes, the natural language notation, the attribute, and the attached auxiliary word are respectively added with “1”, and when they do not match with each other, “0” is added. Is the degree of conformity.

As described above, when there are a plurality of conditions to be met, the search route, which is the result of the semantic analysis, is scored according to the importance of the conditions and the degree of agreement (or similarity). Can be used as an evaluation criterion for determining the priority order of. When there are a plurality of nodes that start the search, or when a branch occurs during the search, there is a footprint corresponding to the search route that joins during the network search. When the two tributaries meet the conditions required for merging, the merged search path is added as a footprint to the footprint list. This condition is simply that two or more footprints that have been network-searched so far are suitable as child nodes of the same node in the chained function structure that includes the footprints (in the above example, a positive value is If you have one), you must meet the confluence conditions. In the case of merging, the footprint information (footstep information d) included in each footprint is appropriately calculated (as a scalar or a vector, sum or merge, etc.).

The contents of the process will be described in detail below with reference to FIGS. 8 to 29 based on more specific data. Step 1 and step 2 of this embodiment
Is the same as in the previous embodiment, so detailed description will be omitted. That is, in step 1, a voice is inputted, "I want to see the efforts of Kihanada of Chiaki music." In step 2,
The voice is converted by the voice recognition device into a corresponding symbol string, that is, a natural language character string 200 "I want to see the approach of Kihanada of Chiaki music." Next, the natural language character string is subjected to natural language analysis (step 41). This process will be specifically described with reference to FIG. First, the natural language character string is subjected to morphological analysis (step 410). However, in step 410 of the present embodiment, essentially the same word extraction as in the previous embodiment may be performed.

That is, the phrase may be extracted by the character string collation between the input natural language character string and the phrase heading of the phrase information table 10 as described above. However, in the case of extracting a more appropriate word, a keyword extraction process, a morpheme analysis process, a syntactic analysis process, an ambiguous resolution process, a semantic analysis process, etc., which are conventional techniques, may be selectively used. For example, by using the parsing process together, even when multiple words are expressed in positions that are distant on the character string,
A compound word composed of those plural phrases can be recognized. Further, by the semantic analysis, it is possible to identify a word having the same expression in a plurality of systems but having a different meaning depending on the system. Although details of these processes are omitted, conventional techniques may be used. in this way,
When high-level natural language analysis processing (syntax analysis, semantic analysis, etc.) is performed in step 41, it is possible to output control information corresponding to a compound word or a syntactic structure. As described above, in step 41, the word / phrase information table 10 is preferably a table including the following information in order to perform the syntax analysis processing and the semantic analysis processing. That is, the word / phrase information table 10 includes information about morphological attributes (notation, utilization, etc.) and information about syntactically attributes (modifying function, case dominance function, etc.) for each morpheme or each constituent element of a natural language corresponding thereto , And information about semantic attributes (strength of operation instruction functional attribute, strength of relationship instruction functional attribute, strength of operation target attribute, strength of operation condition attribute, strength of attribute name attribute, attribute value type) Strength and weakness of attributes,
It is a table having functional labels, semantic labels, etc.).

In the present embodiment, first, the input natural language sentence is morphologically analyzed (divided into natural language elements that can be treated in the same manner as morphemes or morphemes in processing) using the phrase information table 10 having the above-mentioned information, and is ambiguous. Eliminate (characteristics such as master-slave relations and modified relations, features of co-occurrence, positions in writing, and features such as idiomatic usage to limit the scope of meaning). In the case of the current example, "Chiaki music",
Recognizing the clauses of "Kibanada", "Approach", and "I want to see."
3). By the way, according to the information of the word table 10 prepared in advance, "Chiakigaku no" and "Kibanada no" are operation target / condition nodes (semantic labels are "Chiakigaku" and "Kibanada"), and "Approach" is an unnecessary node, It is assumed that "I want to see" is an operation instruction node (the function label is "VTR control"). Here, the unnecessary node is previously registered in the word / phrase table 10 as an unknown word in the input sentence (a word that is not registered in the word / phrase table 10) or as a word that is not an unknown word but is unnecessary for the system. A node that is assigned to a word and is not used (ignored) in the following processing. Therefore, the node 203 will not be handled any more during the description of the present embodiment. Similar to the previous embodiment, the semantic label and the functional label are registered in the phrase information table 10 in advance for each morpheme unit.

As will be apparent from the following description, this step 41 does not necessarily need to analyze the sentence tail at a time. In other words, the following processing can be continued only by analyzing the fragment of the sentence excluding the portion "I want to see the approach of" from the input sentence 200 for "Kihanada of Chiaki Raku". This means that the present invention can be used even when an input sentence contains an unknown word or when the input sentence is grammatically non-sentence.

Next, the meaning of the function chain structure is analyzed (steps 411 to 416). The analysis of meaning by the function chain structure is to elucidate the meaning of a natural language sentence structurally, and is equivalent to constructing the meaning expression conceptually. Therefore, it does not need to be an explicit or descriptive meaning expression, and of course, this meaning expression need not be actually output. As will be described later in detail, in the case of the present example, this semantic representation is represented by a function chain structure for the sake of simplicity. In particular, it is a structure that combines the function chain structures that are expressions of the control information table 11 and the response information table 12. Representation of knowledge used, such as the control information table 11 and the response information table 12,
Since the formed semantic representation has the same format, it is not necessary to consider the representation system of the semantic representation, and therefore the processing from step 411 to step 416 can be executed at high speed.

Next, since the current analysis of the natural language sentence was the first analysis, it is judged that the re-analysis mode is not set (step 411). Next, in the process of step 412, the four nodes obtained in step 410 (that is, “Chiaki Raku's”, “Kibanada's”, “Approach”,
And "I want to see."), The combination of the operation instruction node and the operation target / condition node to be processed is determined.
In the current case, the operation target / condition node is "Chiaki Raku"
And "Kibanada no" as operation instruction nodes and "I want to see." It should be noted that this combination does not necessarily have to be a notational, syntactic, or semantic combination of natural language, as will be apparent from the description below.
This is because, in the case of such a combination, other operation target / condition nodes and operation instruction nodes are appropriately supplemented by network search so as to be semantically matched.

Next, in the processing of step 413, footprints are constructed for the nodes for starting the search for "Chiaki no Raku" and "Kibana no No". Since there is no footstep (footstep information a) one step before this footstep, it is set to be empty (NULL).
The node currently being searched (footstep information b) is a node “Chiaki Raku no” 201 and a node “Kibana no No” 202, respectively.
And In actual processing, the (footprint information b) may be stored with the node IDs of the node “Chiaki Raku” 201 and the node “Kibanada” 202, or the pointer of the storage area. “0” is given to the distance (footstep information c) and the evaluation score (footstep information d) of the search route. The set of function chain structures that can be searched in the next step is given a function chain structure corresponding to the knowledge currently available in the control information table 11 and the response information table 12. In the present description, at least the functional chain structure shown in FIGS. 11 to 18 is given. In this way, the obtained footprint is added to the footprint list.
As can be seen from the above description, the nodes, footprints, and footprint list may be handled in a table (array) format in processing. Correspondence between these data formats is a technique well known to those skilled in the art, and thus detailed description thereof will not be required.

Next, in the process of step 414, the process branches to step 415 to process the unprocessed footprints in the footprint list. Although the node 201 and the node 202 are treated equally, they will be described in order for convenience of explanation. That is, the node currently being searched will be described as 201. According to this procedure, even if the processor (and architecture) cannot perform parallel processing, this embodiment can be realized. The next step 415 is
It will be described in more detail. Since the node 201 is not the operation instruction node, the process branches to step 4152. Next, in the process of step 4152, since the node 201 is not the node corresponding to the pronoun, step 4154 is performed.
Branch to. Next, in the processing of step 4154,
Nodes included in (footprint information e) (that is, node 1
10-164 and nodes 172-177),
Find a node that matches node 201. Here, the nodes (that is, the nodes 153 to 164) belonging to the consequent part of the semantic structure conversion rule are not necessary until the semantic structure conversion rule is actually used. Should be ignored in the matching process. In this case, the independent word portion of the node 201 “Chiakigaku no” matches the node 142 “Chiakigaku” in the natural language notation. It is stored in the work area that "1" is added to the evaluation score of the new footprint which will be configured in step 4155 at this point.

Next, in the processing of step 4155, the node 1 determined to be suitable obtained in step 4154
42 based footsteps. In practice, the function "equ" node is connected to node 20 in order to combine the two nodes.
It is sandwiched between 1 and node 142. First, the footprint for the function "equ" node is constructed. For this footprint, (footprint information a) is 201, (footprint information b) is the pointer of this “equ” node, and (footprint information c) is “1” (this is “1” in the (footprint information c) of the node 201. "(Footprint information d) is" 0 ", and (footstep information e) is the same data as (footstep information e) of the node 201, and this is added to the footstep list. Next, a footprint for the node 142 "Chiakiraku" 142 is constructed. In this footprint, (footprint information a) is the pointer of the previous "equu" node, (footprint information b) is the node 142, (footprint information c) is "2" (this is the (footprint of the previous "equu" node. "1" for information c)
Is added), (footstep information d) is "1" (this is the value obtained in step 4154), (footstep information e)
Is the same data as the previous “equ” node, and this is added to the footprint list. Next, in the processing of step 4156, since no link is output from the function chain structure to which step 201 originally belongs, this processing is ended without being performed.

Next, in step 416, it is determined whether or not the current state of the footprint list satisfies the search termination condition. In the present example, for the sake of simplicity, the condition that the node (or the node that matches the node) for ending the search exists in the (footstep information b) of the footprint in the footprint list is set as the search end condition. Therefore, since the footprint list at the present time does not satisfy the search termination condition, the process branches to step 414. If the node that terminates the search cannot be obtained from the processing so far, the search may be terminated once each time the operation instruction node is passed on the search path, and the operation instruction node is included. The search may be terminated after finding some search routes. For example, if the input sentence is a fragment “Chiakigaku no Kihanada” and the idiom “I want to see.” Is missing, the first obtained search route is node 201 and node 201 in FIG. This is a search route having 202 as a starting point and a node 143 as an ending point. This search route is equivalent to interpreting the input sentence as "do an RDB search on the condition that the work day is Chiaki and the wrestler name is Takahanada." In this way, the input sentence in which the adjectives are omitted cannot always be properly interpreted (of course, if the command desired by the user is simple operation instruction content,
However, the analysis result and the execution result can be used in combination with the operation instruction by the following input sentence (for example, "control the VTR based on the data."). That is, once, the analysis result of the previous sentence (that is, the search path) has a function chain structure, and the execution result (search result in this example) also has a function chain structure stored in the response information table 12. So
By handling the knowledge in the same way as the knowledge expressed by other functional chain structure, it can be used for the analysis of the subsequent input sentence. In this way, it is possible to analyze a complicated operation instruction. As described above, according to the present invention, it is possible to analyze a sentence close to a conversational sentence such as a fragment (or an abbreviated sentence).

Next, in the processing of step 414, the footprint to be processed is determined to be the one with the smallest search distance among the unprocessed footprints in the footprint list. That is, the process is continued with the node currently being searched for as the node 202. Repeating the same process thereafter, the node 204
When the node 172 matching “I want to see” is added to the footprint list, the search termination condition is satisfied, and the process branches to step 42. The search route obtained at this point is shown in FIG. This search path includes a part that depends on the return value from the database. This portion is corrected in step 45 described later. In this way, by once determining the search path corresponding to all the input sentences (including the hypothetical operation target of the consequent part of the function chain structure),
It is possible to make a confirmation sentence based on this search path (sentence that presents to the user how the system is understood) to the user before executing the actual control.

In FIG. 24, node 205 and node 2
06 and 207 are the node 158 and the node 16 respectively.
3 is a node in which a value is assigned to the node 168. $ FIELD (node 146) of node 158
6. A node connected to the network 6 is searched and determined. Specifically, referring to the footprint list, the node returns to the direction opposite to the direction searched from the node 146, and the attribute value node (in this case, “Rikishi name”) inside knowledge of the database structure is set to $ FIELD (node 146). ) Value. Similarly, the value of $ FIELD (node 149) is the “approach date”, and the value of $ FIELD (node 152) is the “start address value”. As described above, the function for extracting the value related to the field is the materializing function $ FIELD. It is desirable that this value is actually calculated and assigned to each node (that is, the nodes 205 to 207) at the time when the operation instruction node “RDB search” 143 is added to the footprint list.

As shown in FIG. 24, the semantic networks may be merged (or branched). When a certain node in the function chain structure has multiple inputs, the search paths merge, and when it has multiple outputs, the search paths branch. In the present example, merging will be described in detail. That is, when the condition node 144 and the condition node 147 are in the footprint list, the operation instruction node 143 is added to the footprint list. At this time, AND and O are applied to these child nodes as constraints.
Another type such as R may be added. In the case of AND, the search to the parent node is not performed (or is held) unless all the child nodes are in the footprint list. In the case of OR, if any child node is in the footprint list, the parent node may be searched. In this case, depending on the combination of child nodes,
An operation instruction node 143 from the condition node 144, an operation instruction node from the condition node 147, or an operation instruction node 143 obtained by merging both of these condition nodes is obtained, and these are given appropriate evaluation scores. , You can add all to the footprint list. It is desirable to set the evaluation score so that the network that includes more nodes (that is, joins) corresponding to the phrase in the natural language is finally given priority. Also, AN
In addition to D and OR (or a combination thereof),
It is also possible to give the child node different restrictions such as essential and optional. That is, the essential child node must be provided when the search proceeds to the parent node, and may be used as a combination of merge sources or may not be provided when the child node is an arbitrary child node. These constraint conditions are stored in advance in the node (the node information (12) described above).

In the present example, the condition node 144, the condition node 147, and the condition node 150 are defined to be indispensable by AND. Therefore, in the process of step 4156, when the nodes 144 and 147 merge and the footprints of the node 143 are stored in the footprint list, at the same time,
It checks to see if node 150 is in the footprint list, and if so, will also join its search path. Further, for example, if there is no footprint corresponding to the node 150, the network search is performed in the reverse direction (this algorithm is the same as the above description except that the direction of the search so far is different, and therefore omitted). In this case, it is desirable that the evaluation of the node (field (node 122) in the database) that is not yet used is high. Of course, other nodes may be added to the footprint list in parallel as footprints as described above. This is because in this embodiment, a plurality of candidates are obtained as the meaning of the natural language sentence (that is, a plurality of search routes are output).
However, the search path (that is, the semantic network) that has a high evaluation finally becomes the first candidate for the meaning of the natural language sentence. In this way, when the footprint corresponding to the node 143 is actually added to the footprint list, it becomes a search route in which three condition nodes are connected.

Next, in the processing of step 42, the semantic network (FIG. 24) obtained in step 416 is
Since it is not a sentence for re-analyzing the input sentence so far (described above), the process branches to step 44. Next, in the processing of step 44, since there is no semantic expression to be updated,
Branch to step 5. Next, in the process of step 5, the semantic network obtained in step 416 is traced again in order, and the control information defined in the consequent part of the first semantic structure conversion rule (in this case, “RDB search”) is found. Since the control target device name 170 and the control signal information 171 shown in FIG. 17 are present, it is determined that there is a control signal to be output, this control information is stacked on the work area 25 stack, and step 46 is executed. Branch off.

Next, in the process of step 46, the embodying function in the control signal information 171 is evaluated. The instantiating function $ FIELD (node 146, node 149, node 152) uses the values obtained from the respective nodes in the same manner as $ FIELD (described above) when there is one argument,
Generates a string of characters connected by white space. In other words, it is “the wrestler's name approach date start address value”. Further, the reification function $ TABLE (node 146, node 149, node 152) is the original function chain structure to which the nodes related to the fields found in $ FIELD belong (in this case, the function chain structure of FIG. 11). Is a function that returns the value of the attribute value corresponding to the attribute name "table name". That is, it is a "sumo wrestling recording table". If there are multiple values, separate them with commas. For example, "sumo wrestling recording table, numbering table table". Furthermore, the materialization function $ WHE
RE (node 146, node 149, node 152)
If there is a node related to the field found in $ FIELD and a node corresponding to it, which represents the corresponding data further back in the network, the relation between these two nodes is expressed by the equation “=” (or “<”). , “>” And “! =” Inequalities), data is enclosed in double quotes, and when there are multiple equalities (or inequalities), the values are returned by delimiting with “and”. In other words, in the present example, “wrestler name =“ Kibanada ”and approach date =“ Chiakiraku ””. As described above, the position of the materialization function of the control signal information 171 is replaced with the return value,
Finally "select wrestler name work day start address value from sumo wrestling recording table where wrestler name =" Kibanada "and work day =" Chiakiraku "20
8 is obtained (FIG. 25).

The delimiters are space, comma, and "and
Etc. depends on the search language of the database that is specifically connected. A system using a search language different from that of the present embodiment appropriately gives a delimiter character accordingly. These slight differences in command formats can be dealt with by slightly changing each embodying function, so that it is not necessary to change the control information (170 and 171) of this embodiment. In this way, the definition of the function chain structure is
It is also possible to define it independently of each application system. If the reification function is nested, after evaluating the inner function,
Evaluate the outer function.

Next, in the process of step 7, a control signal is output to the actual controlled device based on the control information obtained in step 46. That is, the control target device name 1
The control signal information 208 is output to the connected database system 19 based on the "database" 70.
The control signal information 208 is an SQL (structured query la) for the database system 19.
signal), and in this embodiment, signals can be transmitted through the operating system. next,
In the process of step 44, in the semantic network (FIG. 24) obtained in step 416, a specific value is set on the route from the operation instruction node “RDB search” that has been processed to the operation instruction node “VTR control”. Undecided nodes (ie, node 159, node 16)
4 and the node 169) have a function chain structure (that is, a consequent part of the semantic structure conversion rule “RDB search”), and the data for embodying the values of these nodes is still in the response information table. Since it is not stored in 12, it is judged that the semantic network shown in FIG. 24 should not be updated, and the process branches to step 5. Note that in a scene different from this example, a node for which a specific value has not been determined may not be used in the subsequent processing. In this way, the processing of steps 13 and 14 can be omitted by generating the control signal while assuming that the operation target of the consequent part of the semantic structure conversion rule should exist. Particularly, this method is effective when the control target device is controlled in one direction and there is no response data, and all control signals such as batch processing can be converted and prepared in advance. Next, in the process of step 5, since there is no control information stacked on the stack, the process branches to step 8. Next, in the process of step 8, it is judged that the process has not been completed because there is an unprocessed semantic structure conversion rule (that is, VTR control semantic structure conversion rule) on the semantic network obtained in step 416. Then, the process branches to step 44. Next, at step 44, since the response information table 12 has not been updated, it is judged that it is not necessary to update the semantic network shown in FIG. In this way, the loop is executed between step 44, step 5, and step 8. Therefore, this process is. This is substantially equivalent to waiting until the response information table 12 is updated as a result of the search.

In the process of step 13, if there is no signal from the controlled device, the process branches to step 13. Therefore, this process is a process of waiting until a signal from the control target device arrives. In the present example, the control signal information 20
If the database system 19 responds with the search result corresponding to 8, the process branches to step 14. Next, in the process of step 14, the search result from the database is
It is expressed in a function chain structure and stored in the response information table 12. In the specific example of the present embodiment, as a result of the search, a search result is obtained in which the wrestler's name is Takahanada, the work day is Chiaki, and the start address value is 305. This is shown in FIG. 19 as a function chain structure. In this way, when the response information table 12 is updated, the processing looped in steps 44, 5 and 8 is changed to step 41 in step 44.
It is judged that the semantic network obtained in 6 should be updated, and the process branches to step 45.

In the process of step 45, the nodes (that is, nodes 153 to 169) of the consequent part of the semantic structure conversion rule (that is, nodes 153 to 169) are set to the nodes (that is, nodes 178 to 194) in the response information table 12. Replace. When the search result is one record as in the present example, it is substantially the same as the evaluation of the reification function defined by the nodes 159, 164, and 169. That is, the reification function $ CELL (node 146) defined in the node 159 is the value of the cell data corresponding to the value of the field name $ FIELD (node 146) of the search result (in this case, "wrestler name"). It is a function that returns ("Kibanada" in this case). With such a reification function $ CELL, $ CELL (node 14
9) is "Chiakiraku" and $ CELL (node 169) is "3"
05 ".

Here, if the search result is two or more records, the consequent part of the semantic structure conversion rule is copied, and a function chain structure is prepared for each record and stored in the response information table 12. For example, the result of a search
Two record data with the content that "the start address value of Kihanada's approach to Chiaki (in the spring place) is 305" and "the start address value of Kihanada's approach to the Chiaki music in (summer place is 2044)" 26) is obtained, in the process of step 45, for each record,
Nodes 153 to 169 are prepared, and values are appropriately substituted into $ FIELD and $ CELL as described above. By doing so, the present invention can be applied even when there are a plurality of database search results.

Furthermore, a functional chain structure after the node 207 to which the operation instruction node "RDB search" 143 in the semantic network of FIG. 24 is connected (that is, the nodes 153 to 164, "VT
Conditional part of the semantic structure conversion rule of "R control" (nodes 172 to 172)
Node 177), and nodes 207 and 177
"Equ" node that connects the
It is also possible to branch from 3 to the node 207 which is a duplicate of the semantic network. The purpose of updating the semantic representation in this way is to deal with the non-deterministic operation of the system by the semantic structure conversion rule (in this case, "VTR control") existing at two branch destinations.
That is, by copying the content of the operation instruction that has not been operated (in this case, “VTR control”), the control target device (VT
Even if R) performs different operations, it is possible to provide a control method in which those operations do not affect each other.

The description will be continued by returning to the original example.
In the processing of step 5, the semantic network obtained in step 45 is traced again in order, and the control information (ie , Control target device name 170 and control signal information 1 shown in FIG.
71), it is determined that there is a control signal to be output, and this control information is stacked on the stack in the work area 25 and branched to step 46. Next, in the processing of step 46, the embodying function in the control signal information 171 is evaluated. In this example, the materialization function $ DATA (node 177) is the materialization function $ CELL (node 1) described above.
This is a function with the same function as in 77). That is, $ DATA
The return value of (node 177) is the character string “305”. As described above, the control signal finally obtains “JP305” 209 in FIG. Next, in the process of step 7, this is output to the connected VTR 20. Next, in the process of step 44, it is determined that there is no need to update the semantic expression because there is no node to be embodied, and the process branches to step 5. After this, "VTR
The second control information for "control" is stacked on the stack (step 5). However, since this second control information does not include a materialization function, the process proceeds as it is (step 7), and In the same manner, the VTR 20 is connected to the “P
L ”210 is output. Next, in the process of step 44, it is judged in the same manner as described above, the process branches to step 5, and there is no control signal to be output (step 5).
It ends (step 8). FIG. 28 is a specific example of the finally-updated function chain structure corresponding to the semantic expression of the input sentence. Also, FIG. 29 is a specific example of the finally-updated function chain structure corresponding to the case where the search result of the operation instruction 208 to the database is 2 records (FIG. 26). However, it is an excerpt diagram of only the node 143 and the subsequent nodes. As described above, it is possible to semantically analyze a natural language sentence input by voice and operate a plurality of control target devices in cooperation.

In particular, according to this method, even if the end user does not know the operation instruction language depending on the device to be controlled (or the application system) at all, only the basic natural language words (sound) can be used. If is known, it becomes possible to operate the controlled device (or application system). In addition, the end user can accurately control the target device (or application system) through natural language without knowing in detail the complicated operation instruction language system, operation system, and data structure system that differ for each target device (or application system). It becomes possible to operate. Also, when the control target device (or application system) cooperates (in the present example, R
Also in the DB search and VTR control), a data transfer method between the control target devices (or application systems), a control procedure (order and combination) of the control target devices
The end user can operate the device to be controlled (or the application system) accurately without knowing the details. Further, by using a telephone line or the like, even when the end user and some control target devices (or application systems) are located apart from each other, they can be operated in cooperation with each other. In addition, the natural language sentence input by the user is an expression that directly instructs the operation (for example, “a wrestler's name is Takahanada, and the starting date is a search for a start address value in Chiaki music,
Play the VTR from the starting address value. )) Is not only parseable but also an expression that specifies the purpose (goal) (for example, “I want to see the efforts of Kihanada of Senshugaku.”), The operation directive is based on the result of the semantic analysis. Since it is generated, it is possible to appropriately operate a plurality of control target devices (or application systems) in cooperation even if the input is an expression close to a conversational sentence.

According to the present embodiment, knowledge depending on the field,
Various knowledge expressions, such as system knowledge, thesaurus knowledge, control target equipment (or application system) knowledge, and discourse-dependent knowledge, are described in the same expression format (function chain structure), which is convenient for processing procedures. Therefore, it is possible to appropriately use a plurality of pieces of knowledge, and a new meaning expression (FIG. 28) can be easily configured by combining the pieces of knowledge. Further, since the various knowledge has the same expression format, it is possible to commonly use the editors and functions for maintaining these knowledges, which reduces the development man-hours of those editors and functions. In addition, since the editors are used in a unified manner, the usability for knowledge maintenance of the user is improved. In addition, knowledge about the controlled device (or application system) can be defined independently, so even if the system configuration of the controlled device changes, the knowledge can be used additively (incrementally). , Knowledge can be changed easily. In this example, knowledge about RDB search (FIGS. 16 and 17) and knowledge about VTR control (FIG. 1)
8) and are defined as knowledge independent of each other. Similarly, when connecting a voice synthesizer, a second VTR, an expert system, or the like, all that is required is to add knowledge about each control target device (or application system) to the control information table 11. It will be possible. It is also possible to reduce knowledge in the same manner. For example, if the system to be connected is a ticket reservation system, by registering knowledge about the control of this system, issuing a reservation command for the ticket reservation system in response to "I want to see the efforts of Kihanada of Chiaki Raku." It becomes possible to do. When knowledge regarding the VTR and the ticket reservation system is stored in the control information table 11 at the same time, two or more semantic networks as a result of the analysis are found. In such a case, in addition to determining the priority order based on the evaluation score, a step of confirming (or selecting a candidate) the result of analysis of the meaning of the input sentence to the user may be provided between step 41 and step 42. . As described above, according to the present invention, even if a controlled object device that constitutes a certain system is dynamically added (or disconnected) later, the input / output device can be simply added with knowledge. It is possible to provide a natural language interface that solves the complex action or competitive selection of and presents an appropriate natural language interpretation candidate.

According to the present embodiment, since the semantic expression as the result of the semantic analysis is obtained by a simple combination of knowledge given in advance, the semantic analysis algorithm is different from the processing method of outputting another semantic expression format. Is efficient. Further, by using the same one (or a simple copy) on the memory of the computer, the memory consumption is reduced and the processing becomes faster. Further, since the semantic representation also has a function chain structure, it is possible to refer to it in the semantic analysis process by storing this semantic representation as history knowledge, as with the knowledge represented by other function chain structures. As a result, in particular, the function chain structure that is the result of semantic analysis such as a negative response statement or a statement that adds a condition afterwards, such as a proviso, and the function chain structure of a sentence that is input and analyzed before that are combined. It is also possible to do. Therefore, it becomes possible to analyze a negative response sentence or a sentence to which a condition is added, which is close to such a conversational sentence, and it is possible to easily re-analyze the previous input sentence.

According to this embodiment, all inputs are analyzed,
The operation target (in this example, the node 153 to the node 169) to be generated (or disappeared) is defined in the consequent part of the semantic structure conversion rule, and the semantic analysis (step 41 or step 43) is performed.
By assuming this operation target and correcting this assumption (step 45) based on the result of actually executing the operation ahead (in the present example, the search result of the RDB), the backward of the sentence is nondeterministic. It is possible to analyze even if it depends on the execution result in front of the statement. Or, simply, you can divide the input sentence appropriately based on breath breaks, pause breaks, punctuation (by keyboard input of characters, etc.), and analyze only the beginning of the sentence. , After the corresponding operation is performed, the backward of the sentence may be analyzed based on the execution result of the forward. Since an arbitrary function chain structure can be described in the consequent part of the meaning structure conversion rule, the operation target (node 153) is used as in this embodiment.
~ Node 169), the semantic structure conversion rule itself may be described in the consequent part of another semantic structure conversion rule. In this way, when the semantic structure conversion rules are described in the consequent part, the number of usable rules increases or decreases as a result of a certain operation instruction. This means that different knowledge can be used for semantic analysis depending on the situation of the discourse (or the progress of the dialogue). In this way, discourse-dependent analysis and semantic analysis that follows the focus of a topic become possible.

As a modification of the above-described embodiment, the case of interlocking with a voice synthesizer will be described with reference to FIGS.
~ It demonstrates using FIG. In the following description, the description of the parts that are substantially the same as those in the previous embodiment is omitted. FIG. 30 is a semantic structure conversion rule regarding voice output in this embodiment. That is, the knowledge that the character string 251 is converted into katakana (the materialization function $ KATAKANA) and is output to the voice synthesizer is described. FIG. 31 is a specific example of knowledge of upper-lower relations in the system in the present embodiment. That is, the knowledge that “the wrestler name is a character string” is described. FIG. 32 is a specific example of thesaurus-related knowledge in the target field in this embodiment. That is, the knowledge that "Sekiwaki is numbered" is described. FIG. 33 is a specific example of knowledge about the database structure in this embodiment. That is, a record file in which the first field is a wrestler name, the second field is a number, and the table name is a numbering table is described.
The above knowledge is the same as the above description, so further detailed description will not be required. In this embodiment, the knowledge of FIGS. 12 to 17 and FIGS. 30 to 33 is stored in the control information table 11.

It is now assumed that the user inputs "Tell me who is Sekiwaki?" 272 by voice (FIG. 34). According to the information in the word / phrase information table 10 prepared in advance, “Sekiwaki wa” 273 is an operation target / condition node (semantic label is “Sekiwaki”), “someone” 274 is an unnecessary node, and “teach me.” 2
75 is an operation instruction node (the function level is "voice output"). FIG. 35 is a function chain structure obtained as a result of the semantic analysis. A search instruction 276 to the RDB is generated based on this result (FIG. 36). In response to this search instruction 276, the database is searched and the search result (see FIG.
7) is stored in the response information table 12 and the semantic expression is updated (FIG. 39 (a diagram showing a search route after the RDB search)). Based on this result, the operation instruction “Takahana dades.” 277 and the operation instruction “Akebonodes.” 278 to be output to the speech synthesizer are generated (FIG. 38). In this example, "Kibanada" and "Akebon" are the names of wrestlers. As described above, it is possible to control a plurality of systems in which the voice synthesizers are combined.

According to this embodiment, knowledge for RDB search (FIGS. 16 and 17), knowledge for VTR control (FIG. 18), and knowledge for voice output (FIG. 30) are mutually related. being independent. Moreover, by additionally storing these pieces of knowledge in the control information table 11, it is possible to derive an interpretation that combines these pieces of knowledge (for example, FIG. 24). Therefore, by adding knowledge (semantic structure conversion rules, etc.) prepared for each control target device (for example, second-generation VTR, air conditioner, etc.), a system in which the control target device and an already existing system are integrated is created. On the other hand, the user can give an operation instruction in natural language without paying attention to the control of the individual system. In the present example, consider a case where the operation instruction node is defined as “voice output” or “VTR control” with respect to “tell me”. At this time, if the search end condition is ignored, at least the semantic networks include “voice output” via “RDB search” and “VTR control” via “RDB search”. These meanings are “teaching” by voice output of search results (ie, “Kibanada” and “Akebono” in the present example), or VT.
It is equivalent to "teaching" by outputting to the TV screen by R. As described above, according to the present invention, it is possible to interpret the meaning for the intended purpose of the input sentence even when the controlled device to be operated is different.

According to the above description, the response information table 12 in this embodiment can express all the data in the table format by the function chain structure. That is, generally, if the data set can be stored in a table format on a computer, it can be used in the natural language analysis process (step 41 or step 43), the semantic expression update process (step 45), and the like. Indicates that there is. For example, the last-handled data is stored as history data (FIG. 40 and FIG. 41 are corresponding function chain structures). At this time, the word "that" is defined as an operation target / condition node (the semantic label is "the last data handled") by the phrase information table 10, so that "I want to see the approach of the person on the first day. It is possible to analyze the input sentence such as ". According to this embodiment, it is possible to interpret a natural language sentence including a pronoun in this way.

As is clear from the above detailed description, by operating a plurality of systems that are distant from each other in cooperation, the user does not have to be aware of the physical layout of each system and the network connection form. It is possible to provide a natural language control device that enables a necessary database search and remote control of another system based on the search result. Further, it is possible to perform the cooperative operation without knowing the details of the configuration of the control target device and the control procedure. Furthermore, it is possible to control by voice input close to a conversational sentence including pronouns. Further, as an application example of the present invention, it is also possible to always accept a broadcast including audio such as television and radio as an input, and perform control, present related information contents, etc. in response to this input. For example, when the announcer broadcasts that the sumo program will start, the VTR recording control is started. Alternatively, when the value of the stock price is broadcast, it can be added to the database, aggregated, compared with the value of the previous day, and the result thereof can be displayed. Details regarding how to handle these systems will be readily inferred by those skilled in the art from the foregoing description.

[0101]

According to the present invention, it is possible to control the device to be controlled and to perform remote control by giving an input in a natural language such as normal Japanese or English. it can. Also, by giving an input in normal language such as normal Japanese or English, it is possible to combine a plurality of control target devices to perform a cooperative operation, and further, it is possible to remotely control the cooperative operation. Obtainable.

[Brief description of drawings]

FIG. 1 is a diagram showing an outline of processing in an embodiment of the present invention.

FIG. 2 is a block diagram showing a hardware configuration according to the present embodiment.

FIG. 3 is a diagram showing a specific example of a word / phrase information table in the present embodiment.

FIG. 4 is a diagram showing a specific example of a control information table in the present embodiment.

FIG. 5 is a diagram showing a specific example before updating a response information table in the present embodiment.

FIG. 6 is a diagram showing a specific example of a sumo wrestling recording table stored in the database system according to the present embodiment.

FIG. 7 is a diagram showing a specific example after updating a response information table in the present embodiment.

FIG. 8 is a diagram showing an outline of processing in a modified example of the present embodiment.

FIG. 9 is a diagram illustrating a basic structure of a function chain structure.

FIG. 10 is a diagram illustrating a function chain structure expressing an operation target and a condition.

FIG. 11 is a diagram showing a specific example of knowledge about the database structure in the embodiment based on FIG. 8;

FIG. 12 is a diagram showing a specific example of knowledge of upper-lower relations in the system in the embodiment based on FIG.

FIG. 13 is a diagram showing a specific example of knowledge of upper-lower relations in the system in the embodiment based on FIG.

FIG. 14 is a diagram showing a specific example of thesaurus-related knowledge in the target field in the embodiment based on FIG. 8;

FIG. 15 is a diagram showing a specific example of thesaurus-related knowledge in the target field in the embodiment based on FIG. 8;

FIG. 16 is an “RDB search” in the embodiment based on FIG. 8;
It is a figure which shows the specific example of the conditional part of the meaning structure conversion rule which defined.

17] "RDB search" in the embodiment based on FIG.
It is a figure which shows the specific example of the consequent part of the meaning structure conversion rule which defined.

FIG. 18: “VTR control” in the embodiment based on FIG.
It is a figure which shows the specific example of the meaning structure conversion rule which defined.

FIG. 19 is a diagram showing a specific example (function chain structure) of knowledge about an operation target and its attribute in the embodiment based on FIG. 8;

FIG. 20 is a diagram showing a specific example (table) of knowledge regarding an operation target and its attribute in the embodiment based on FIG. 8;

FIG. 21 is a diagram showing a detailed flowchart of a natural language analysis processing procedure in the embodiment based on FIG. 8;

22 is a diagram showing a detailed flowchart of a processing procedure for network search in the embodiment based on FIG. 8. FIG.

FIG. 23 is a diagram showing a specific example in which a natural language sentence is divided into phrases.

FIG. 24 is a diagram showing a specific example of a function chain structure obtained as a result of semantic analysis of a natural language sentence.

FIG. 25 is a diagram showing a specific example of control signal information for RDB search.

FIG. 26 is a diagram showing a specific example of knowledge about an operation target and its attribute.

FIG. 27 is a diagram showing a specific example of control signal information for VTR control.

FIG. 28 is a diagram showing a specific example of a final function chain structure for a natural language sentence.

FIG. 29 is a diagram showing a specific example of the final chained function structure for natural language sentences (when the search result is two records).

FIG. 30 is a diagram showing a specific example of the semantic structure conversion rule defining “voice output” in the embodiment based on FIG. 8;

FIG. 31 is a diagram showing a specific example of knowledge of upper-lower relations in the system in the embodiment based on FIG. 8;

FIG. 32 is a diagram showing a specific example of knowledge of thesaurus relations in the target field in the embodiment based on FIG. 8;

33 is a diagram showing a specific example of knowledge about the database structure in the embodiment based on FIG. 8. FIG.

FIG. 34 is a diagram showing a specific example in which a natural language sentence is divided into phrases.

FIG. 35 is a diagram showing a specific example of a function chain structure obtained as a result of semantic analysis of a natural language sentence.

FIG. 36 is a diagram showing a specific example of control signal information for RDB search.

FIG. 37 is a diagram showing a specific example of knowledge regarding an operation target and its attribute.

FIG. 38 is a diagram showing a specific example of control signal information for a speech synthesizer.

FIG. 39 is a diagram showing a specific example of the final chained function structure for a natural language sentence.

FIG. 40 is a diagram showing a specific example (table) of knowledge regarding operation history in the embodiment based on FIG. 8;

FIG. 41 is a diagram showing a specific example (functional chain structure) of knowledge about an operation history in the embodiment based on FIG. 8;

[Explanation of Codes] 10 word / phrase information table 11 control information table 12 response information table 15 processor 16 storage device 17 voice input device 18 voice recognition device 19 database system 20 VTR 21 TV 22 voice synthesis device 23 voice output device 24 natural language analysis program 25 work areas

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

Claims (7)

[Claims] 1. A control device for controlling a controlled device, comprising: a voice recognition unit for recognizing the contents of the voice in response to an input of a voice from a voice input unit; a database; and a control of the controlled device. A storage unit that stores a phrase information table that stores the phrase information, a control information table that stores the correspondence between the phrase and control information for the control target device, and a response information table that stores search results of the database, A processing unit, wherein the processing unit refers to the word / phrase information table based on the recognition result by the voice recognition unit, extracts a word / phrase related to control of the control target device, and corresponds to the extracted word / phrase. Means for extracting the control information by referring to the control information table, and a control based on the control information to which restriction information is not added among the extracted control information. Means for generating a signal and outputting it to the controlled device or the database; means for storing response information in the response information table when there is a response from the database; and limitation by the response information stored in the response information table. A control device comprising means for complementing control information to which information has been added, generating a control signal based on the complemented control information, and outputting the control signal to the control target device. 2. The control device according to claim 1, wherein the means for extracting a word related to the control of the controlled device is one of a morpheme analysis means, a syntax analysis means, or a semantic analysis means,
A control device comprising at least one and extracting a phrase. 3. A control device for controlling controlled equipment, comprising: a voice recognition means for recognizing the contents of the voice in response to input of voice from the voice input means, a database, and control of the controlled equipment. Storage means accommodating a word / phrase information table storing such word / phrase information, a control information table storing a function chain structure for performing meaning analysis, and a response information table storing a search result of the database in a function chain structure expression And a processing device, wherein the processing device includes at least one of a morpheme analysis means, a syntax analysis means, or a semantic analysis means, and refers to the word / phrase information table based on a recognition result by the voice recognition means. Means for extracting words related to control of the controlled device, and meaning for performing semantic analysis by referring to the control information table based on the extracted words. To explore the network representation,
Means for extracting control information corresponding to the extracted words and phrases based on the semantic network expression; means for generating a control signal based on the extracted control information and outputting the control signal to the controlled device or the database; Means for storing response information in the response information table in the form of a functional chain structure when there is a response from the database; and when the response information is stored in the response information table, complementing the semantic network representation with the response information, A means for extracting control information corresponding to the extracted phrase based on the complemented semantic network expression; and means for generating a control signal based on the extracted control information and outputting the control signal to the control target device. Characteristic control device. 4. The control device according to claim 3, further comprising a storage unit for storing the semantic network expression that is the result of the semantic analysis, and the semantic analysis stored in the storage unit based on a voice recognition result for a new voice input. A control device characterized by reanalyzing the result. 5. The control device according to any one of claims 1 to 4, wherein the voice input means is a microphone, and the microphone is coupled to a voice recognition means via a wireless communication device or a telephone line. A control device, wherein the control target device can be remotely controlled. 6. The control device according to any one of claims 1 to 4, wherein the control target device is a VT.
A control device which is a combination of at least one device of an R, a television, and an audio output device. 7. The control device according to claim 1, further comprising a character recognition means, a keyboard input means, a button operation means, or a menu selection means instead of the voice recognition means. At least one of the above means is provided so that a natural language can be input.