JP2017192005A - Observing system by artificial intelligence agent - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an observing system capable of reducing human cost by facilitating grasp of a state of a person to be observed.SOLUTION: The observing system detects singularity from conversation data of a plurality of artificial intelligence agents and terminal users 23 of the plurality of artificial intelligence agents and a conversation data conversion part 12 and learning data 17 developed by a learning data development part 13 and grasps states of terminal users of artificial intelligence agents. The artificial intelligence agent 5 performs utterance and on-hook processing, thereby automatically performing state confirmation of the terminal users of the artificial intelligence agents.SELECTED DRAWING: Figure 5

Description

The present invention relates to a watching technique using a plurality of artificial intelligence.

2. Description of the Related Art Conventionally, there is a technique in which a person to be watched has an information communication terminal, and a caregiver calls the person to be watched through the information communication terminal.

JP2015-216565

JP2015-207077

However, although the above-mentioned watching system can check the read flag of the information communication terminal, there is a drawback that it is impossible to grasp the situation of the watching target person.

Moreover, it is necessary for the caregiver to periodically notify the person to be watched over, and there is a drawback that it takes too much human cost to notify individually.

In order to solve the conventional problems, the present invention includes an artificial intelligence agent terminal, a monitoring server, and a monitoring terminal connected via a communication line, the artificial intelligence agent terminal including a plurality of artificial intelligence agents, and conversation data. A transmission / reception unit that exchanges between the watch server and the watch terminal, a conversation display unit that exchanges conversation data, and a response unit that, when receiving the conversation data, refers to the learning data of the watch server and returns the conversation data; Conversation data conversion means for converting conversation data received from the artificial intelligence agent terminal into multidimensional data, comprising conversation start means for referring to conversation data and learning data of the watch server and starting conversation. Record conversation data and conversation data from conversation data converted to multidimensional data Learning data generating means for generating learning data weighted by numerical values; transmission / reception means for watching and interacting with conversation data and learning data; and recording means for recording conversation data and learning data; and It is provided with a notification means for extracting conversation data exceeding a threshold value from a numerical value and transmitting it to the watching terminal, wherein the watching terminal exchanges conversation data and learning data with the artificial intelligence agent terminal, the watching server and the watching terminal. A monitoring system characterized by comprising:

According to the present invention, the watching server extracts conversation data that exceeds the threshold from the numerical value of the conversation data of the learning data, and transmits it to the watching terminal. For this reason, the state of the artificial intelligence agent terminal user can be grasped.

In the present invention, the artificial intelligence agent refers to the conversation data and learning data of the watching server and actively starts conversation. For this reason, it is not necessary for the watching side to always have a conversation or monitor the contents of the conversation, and the human cost can be reduced.

It is a figure showing the composition of the watching system concerning one embodiment. It is a figure which shows the functional structure of the artificial intelligence agent terminal which concerns on one Embodiment. It is a figure which shows the functional structure of the watching server which concerns on one Embodiment. It is a figure which shows the functional structure of the monitoring terminal which concerns on one Embodiment. It is a flowchart which shows the flow of the response by an artificial intelligence agent. It is a flowchart which shows the flow of the speech by an artificial intelligence agent. It is a flowchart which shows the flow of the end story by an artificial intelligence agent. It is a figure which shows an example at the time of the artificial intelligence agent responding. It is a figure which shows an example at the time of the artificial intelligence agent uttering. It is a figure which shows an example when an artificial intelligence agent ends a talk. It is a flowchart which shows the flow which detects specificity from learning data and conversation data. It is a figure which shows an example of learning data. It is a flowchart which shows the flow of production | generation of learning data. It is a figure which shows the program implementation comparison of a conversation data conversion part. It is a figure which shows an example of the improvement of learning efficiency by multidimensionalization of conversation data.

Hereinafter, an embodiment of the present invention will be described with reference to the drawings.

FIG. 1 is a diagram illustrating a configuration of a watching system 1 according to an embodiment. The monitoring system 1 is a monitoring system for grasping the state of the artificial intelligence agent terminal user 23 by talking with the artificial intelligence agent terminal user 23, and includes the artificial intelligence agent terminal 2, the monitoring server 3, and the monitoring terminal 4 Is provided. The artificial intelligence agent terminal 2, the watching server 3, and the watching terminal 4 are in a state where data can be transmitted and received bidirectionally over the Internet via a communication line 24 such as a wired or wireless LAN.

The artificial intelligence agent terminal 2 responds to the conversation data input by the artificial intelligence agent terminal user 23, ends the conversation to end the conversation with the artificial intelligence agent terminal user 23, and uses the artificial intelligence agent terminal. It is a terminal which performs the utterance which talks to person 23. The artificial intelligence agent terminal 2 uses a smartphone application as an embodiment, but may be another terminal such as a robot, a dedicated terminal, or a tablet.

The watching server 3 is a server that transmits conversation data corresponding to responses, utterances, and endings from the learning data to the artificial intelligence agent terminal 2 based on the data received from the artificial intelligence agent terminal 2. When the learning data 17 and the conversation data 18 are referred to based on the data received from the artificial intelligence agent terminal 2, if the numerical value calculated by the learning data 17 is within the specified range, the monitoring terminal is notified.

The watching terminal 4 is a terminal that receives the conversation data detected by the watching server 3. The watching terminal 4 can browse the conversation data and learning data of the watching server 3 as necessary. Further, the watching terminal 4 can transmit conversation data to the artificial intelligence agent terminal 2 without using the watching server 3.

FIG. 2 is a diagram showing a functional configuration of the artificial intelligence agent terminal 2 according to the embodiment. The artificial intelligence agent terminal 2 includes an artificial intelligence agent 5, a conversation display unit 8, an artificial intelligence agent terminal transmission unit 9, and an artificial intelligence agent terminal reception unit 10 in order to have a conversation with the artificial intelligence agent terminal user 23.

The artificial intelligence agent 5 includes a response unit 6 and a conversation start unit 7 to transmit conversation data to the conversation display unit 8 viewed by the artificial intelligence agent terminal user 23.

When responding to the conversation data input by the artificial intelligence agent terminal user 23, the conversation data input to the conversation display section 8 is received by the response section 6 of the artificial intelligence agent 5, and transmitted to the artificial intelligence agent terminal. Send conversation data to part 9. The conversation data returned from the watching server 3 is received by the artificial intelligence agent terminal receiving unit 10, transmitted to the response unit 6, and displayed on the conversation display unit 8.

When ending the conversation with the artificial intelligence agent terminal user 23, the conversation data input to the conversation display section 8 is received by the response section 6 of the artificial intelligence agent 5, and transmitted to the artificial intelligence agent terminal. Send conversation data to part 9. The conversation data returned from the watching server 3 is received by the artificial intelligence agent terminal receiving unit 10, transmitted to the response unit 6, and displayed on the conversation display unit 8. The response to the conversation made by the artificial intelligence agent terminal user 23 and the end talk for ending the conversation with the artificial intelligence agent terminal user 23 are performed based on the data received from the watching server 3.

When speaking to the artificial intelligence agent terminal user 23, the conversation start unit 7 transmits the conversation data to the artificial intelligence agent terminal transmission unit 9 based on the state of the conversation display unit 8 and the past conversation data. . The conversation data returned from the watching server 3 is received by the artificial intelligence agent terminal receiving unit 10, transmitted to the conversation display unit 8, and displayed on the conversation display unit 8.

FIG. 3 is a diagram illustrating a functional configuration of the watching server 3 according to the embodiment. The watching server 3 is a server that replies based on the data received from the artificial intelligence agent terminal 2 and the watching terminal 4, and includes a watching server receiving unit 11, a conversation data converting unit 12, a learning data generating unit 13, and a notifying unit 14. And a server transmission unit 15 and a recording unit 16.

The recording unit 16 of the watching server 3 includes learning data 17 and conversation data 18. The learning data 17 includes a process for calculating a numerical value from the received conversation data and a machine learning device called a learned neural network or a deep learning framework. The learning data 17 and the conversation data 18 are recorded in a database device called a relational database.

The watching server 3 watches the conversation data from the artificial intelligence agent terminal 2 and receives it by the server receiving unit 11, refers to the numerical value calculated by the learning data 17 of the recording unit 16 and the conversation data 18, and talks to the watching server transmission unit 15. Send data. The conversation data to be transmitted is the one having the highest numerical value from the numerical values calculated by the learning data 17 as shown in FIG. The numerical values shown in FIG. 12 are an example, and the data structure changes depending on the structure of the learning data generation unit 13 and the generated learning. The learning data will be described later.

The conversation data conversion unit 12 receives conversation data from the conversation data 18 of the recording unit 16 and performs data conversion for transmission to the learning data generation unit 13.

The learning data generation unit 13 receives the conversation data converted by the conversation data conversion unit 12, generates learning data, and records it in the recording unit 16. The learning data generation unit 13 is a machine learning device called a neural network or a deep learning framework. The learning data 17 recorded in the recording unit 16 includes a learned neural network and a deep learning framework and numerical values.

The notification unit 14 refers to the learning data 17 and the conversation data 18 in order to transmit data to the watching terminal 4, and when detecting the conversation data in which the numerical value calculated by the learning data 17 is within the specified range, The data is transmitted to the watching server transmission unit 15.

FIG. 4 is a diagram illustrating a functional configuration of the watching terminal 4 according to the embodiment. The watching terminal 4 is a terminal for grasping the situation of the artificial intelligence agent terminal user 23, and includes a watching terminal receiving unit 19, a watching terminal transmitting unit 20, a display unit 21, and a conversation data input unit 22. The watching terminal 4 watches the data transmitted from the watching server 3 and receives it by the terminal receiving unit 19 and displays it on the display unit 21.

The watching terminal 4 can transmit the conversation data to the artificial intelligence agent terminal 2 without going through the watching server 3 by inputting the conversation data to the conversation data input unit 22.

So far, the functional configuration of the watching system 1 has been described. From here, an example of the flow of processing executed by each of the artificial intelligence agent terminal 2, the watching server 3, and the watching terminal 4 included in the watching system 1 will be described with reference to the drawings.

FIG. 5 is a flowchart showing a response flow by the artificial intelligence agent. The response processing shown in the figure is performed when the artificial intelligence agent terminal 2 and the watching server 3 are activated and connected by a communication line, and input by the conversation data of the artificial intelligence agent terminal user 23. It is executed (step S101).

The artificial intelligence agent 5 confirms the conversation data in the response unit 6, and transmits the conversation data to the artificial intelligence agent terminal transmission unit 9 (step S102).

The artificial intelligence agent terminal transmitting unit 9 transmits the conversation data to the watching server receiving unit 11 of the watching server 3 (step S103).

The watching server 3 refers to the learning data 17 (step S104).

The watching server 3 checks the history and time data of past conversation data and the numerical value calculated from the learning data 17 (step S105). At this time, if the numerical value is within the specified range, the peculiarity of the conversation content is detected, and the watch terminal 4 is notified. Specificity detection will be described later.

When detecting the specificity of the conversation content, the watching server 3 transmits the learning data 17 related to the conversation data received from the artificial intelligence agent terminal 2 to the notification unit 14 (step S106).

The notification unit 14 watches the received data and predetermined transmission destination information of the monitoring terminal 4 and transmits the data to the server transmission unit 15. The watching server transmitting unit 15 transmits data to the watching terminal 4 (step S107).

The watching terminal 4 receives data at the watching terminal receiving unit 19. The watcher who uses the watch terminal 4 confirms the content of the conversation data and inputs the status confirmation data to the conversation data input unit 22. The input data is transmitted to the artificial intelligence agent terminal receiving unit 10 via the watching terminal transmitting unit 20 (step S107).

The artificial intelligence agent terminal 2 transmits the received data from the response unit 6 to the conversation display unit 8 (step S115).

When the artificial intelligence agent terminal user 23 appropriately responds to the conversation, the situation has been confirmed, and the process ends (step S116). The conversation data input for response by the artificial intelligence agent terminal user 23 is transmitted to the watching server 3.

When the artificial intelligence agent terminal user 23 does not respond or does not respond appropriately, the user of the watching terminal 4 performs an emergency response (step S117).

Next, a case where the specificity of the conversation content is not detected in step S105 will be described.

The monitoring server 3 extracts the necessary conversation data from the conversation data 18 and transmits the conversation data to the monitoring server transmission unit 15 when the specificity of the conversation content is not detected (step S108).

The artificial intelligence agent terminal 2 receives the conversation data at the artificial intelligence agent terminal receiving unit 10 and transmits it to the response unit 6 (step S109).

The response unit 6 selects a conversation / artificial intelligence agent that responds from the plurality of artificial intelligence agents 5 (step S110). The reason why there are a plurality of artificial intelligence agents is that there are artificial intelligence agents that can be used for general purposes and artificial intelligence agents that correspond to the artificial intelligence agent terminal user 23. For example, there is an artificial intelligence agent corresponding to a dialect.

The response unit 6 transmits the selected conversation / artificial intelligence agent data to the conversation display unit 8 (step S111).

The artificial intelligence agent terminal user 23 browses the conversation via the conversation display unit 8 and ends the process (step S112).

The flow of response by the artificial intelligence agent is repeated each time the artificial intelligence agent terminal user 23 inputs conversation data. When the end story by the artificial intelligence agent shown in FIG. 7 is confirmed, the conversation ends. The end story by the artificial intelligence agent will be explained later.

FIG. 6 is a flowchart showing the flow of utterance by the artificial intelligence agent. The utterance process shown in the figure is in a state where the artificial intelligence agent terminal 2 and the watching server 3 are activated and connected by a communication line, and the conversation start unit 7 of the artificial intelligence agent 5 has not received conversation data. Thus, the process is executed (step S201).

The conversation start unit 7 automatically determines that conversation data has not been received based on past conversation data history and time data, and refers to the conversation server 3 via the artificial intelligence agent terminal transmission unit 9 I do. When the watching server receiving unit 11 receives the unreceived information of the conversation data from the artificial intelligence agent terminal 2, the watching server receiving unit 11 refers to the learning data 17 of the recording unit 16 (step S202).

Based on past conversation data history and time data, the numerical value calculated by the learning data 17 is confirmed, the conversation data associated with the highest numerical data is observed, and the artificial intelligence agent is passed through the server transmission unit 15. It transmits to the terminal receiver 10 (step S203).

The artificial intelligence agent terminal receiving unit 10 transmits the conversation data received from the watching server 3 to the conversation starting unit 7 (step S204).

As shown in FIG. 9, since a plurality of artificial intelligence agents 5 are prepared, the conversation start unit 7 selects an artificial intelligence agent that speaks from the conversation data, and sends the data to the conversation display unit 8 (step S205).

The conversation display unit 8 utters based on the received data (step S206). Figure 9 shows the image of the application when the second artificial intelligence agent actually speaks.

The artificial intelligence agent terminal user 23 browses the conversation of the artificial intelligence agent (step S207).

Since the conversation data of the utterance has contents that the artificial intelligence agent terminal user 23 can easily respond to, the response of the artificial intelligence agent terminal user 23 is awaited (step S208).

When the artificial intelligence agent terminal user 23 responds, the conversation ends (step S209). The conversation data input for response by the artificial intelligence agent terminal user 23 is transmitted to the watching server 3.

When the artificial intelligence agent terminal user 23 does not respond, emergency response is performed (step S210).

FIG. 7 is a flowchart showing the flow of the end story by the artificial intelligence agent. The end-call processing shown in the figure is performed when the artificial intelligence agent terminal 2 and the watching server 3 are activated and connected by a communication line, and input by the conversation data of the artificial intelligence agent terminal user 23. Is executed (step S301).

The artificial intelligence agent 5 confirms the conversation data in the response unit 6, and transmits the conversation data to the artificial intelligence agent terminal transmission unit 9 (step S302).

The artificial intelligence agent terminal transmitting unit 9 transmits the conversation data to the watching server receiving unit 11 of the watching server 3 (step S303).

The watching server 3 refers to the learning data 17 (step S304).

The watching server 3 confirms the numerical value calculated by the learning data 17 based on the history of past conversation data and the time data, and watches the conversation data associated with the highest numerical data via the server transmitting unit 15. Is transmitted to the artificial intelligence agent terminal receiving unit 10 (step S305).

The artificial intelligence agent terminal receiving unit 10 transmits the conversation data received from the watching server 3 to the response unit 6 (step S306).

Since a plurality of artificial intelligence agents 5 are prepared as shown in FIG. 10, the response unit 6 selects an artificial intelligence agent that speaks from the conversation data, and sends the data to the conversation display unit 8 (step S307).

The conversation display unit 8 ends the conversation based on the received data (step S308). FIG. 10 is an image of the application when the second artificial intelligence agent actually ends the conversation.

The artificial intelligence agent terminal user 23 browses the conversation of the artificial intelligence agent (step S309).

Since the conversation data of the end talk is such that the artificial intelligence agent terminal user 23 can easily end the talk, the conversation ends (step S310). The conversation data when the artificial intelligence agent terminal user 23 responds is transmitted to the watching server 3. Concerning the conversation data of the end talk, the learning data 17 is generated via the conversation data converting unit 12 and the learning data generating unit 13, so that the conversation data having the highest numerical value from the calculated numerical values is used.

FIG. 8 is a diagram illustrating an example when the artificial intelligence agent responds. As the artificial intelligence agent terminal user 23 inputs the conversation data, one of the plurality of artificial intelligence agents responds.

FIG. 9 is a diagram illustrating an example when an artificial intelligence agent speaks. One of the plurality of artificial intelligence agents is speaking because the conversation start unit 7 of the artificial intelligence agent 5 is activated.

FIG. 10 is a diagram illustrating an example when the artificial intelligence agent ends the talk. When the artificial intelligence agent terminal user 23 inputs the conversation data, one of the plurality of artificial intelligence agents ends the conversation.

FIG. 11 is a flowchart showing a flow of detecting specificity from the numerical value calculated from the learning data 17 and the conversation data 18. The peculiarity detection shown in the figure is executed when conversation data is received from the artificial intelligence agent terminal 2 and the learning data 17 is referred to while the watching server 3 is activated (step S501). . Further, when the watching terminal 4 and the watching server 3 are activated and connected by a communication line, the watching terminal 4 can execute the process.

The watching server 3 checks the numerical value calculated by the learning data 17 based on the history of past conversation data and the time data (step S502).

When the conversation data associated with the high numerical data can be confirmed, the watching server 3 transmits the conversation data to the artificial intelligence agent terminal 2 (step S503). This step is the same processing as step S108 shown in FIG.

If the numerical value is within the specified range, specificity is detected (step S504).

The watching server 3 transmits conversation data to the notification unit 14 (step S505). This step is the same processing as step S106 shown in FIG.

By repeating such watching process, the learning data 17 and the conversation data 18 of the artificial intelligence agent terminal user 23 are recorded in the recording unit 16.

The reason why the utterance and the end story by the watching process described so far are used as an index for determining the state of the artificial intelligence agent terminal user 23 will be described in detail here.

For example, when an utterance is made by an artificial intelligence agent, the activity of the artificial intelligence agent terminal user 23 can be confirmed regardless of the activity time of the artificial intelligence agent terminal user 23. In the present invention, the artificial intelligence agent terminal user 23 is assumed to be an elderly person, but as the number of elderly people increases, one watcher corresponds to one elderly person. , Human resources are known to be great. However, the artificial intelligence agent 5 can automatically reduce human resources.

Further, since the speech process is performed if the communication line is connected, the state of the artificial intelligence agent terminal user 23 can be confirmed even when the user is not at home, for example, when moving away. For example, when traveling or traveling. Even in such a case, the state of the artificial intelligence agent terminal user 23 can be confirmed by the artificial intelligence agent terminal user 23 carrying the artificial intelligence agent terminal 2.

However, the artificial intelligence agent terminal user 23 may not carry the artificial intelligence agent terminal 2 around. However, if there is no response to the utterance process, the peculiarity can be detected, so that the person who owns the terminal 4 knows that the artificial intelligence agent terminal user 23 is in an unusual state can do.

Also, there is a possibility that a third party other than the artificial intelligence agent terminal user 23 uses the artificial intelligence agent terminal 2 to respond. In that case, when the numerical value calculated by the learning data 17 of the watching server 3 is confirmed, if the correspondence is different from normal, the history and time of the past conversation data and the numerical value calculated by the learning data 17 Can be detected. For example, when the artificial intelligence agent terminal user 23 always responds to good morning speech processing, if a third party other than the artificial intelligence agent terminal user 23 responds with good morning, Since the reaction is different, the numerical value calculated by the learning data 17 shows a value different from the normal value, so that the specificity is detected and the monitoring terminal 4 is notified of the information.

Further, when only the same response is repeated with respect to the utterance by the artificial intelligence agent, it can be understood that some state abnormality has occurred in the artificial intelligence agent terminal user 23. For example, good morning, hello, the speech processing of evening, when the artificial intelligence agent terminal user 23 is always good morning and in response, becomes within a range of numerical values calculated by the learning data 17 is specified, specificity is detected. Specifically, when the designated range of the numerical value is 0.5 or less, as shown in FIG. 12, it becomes 0.1 at night and can be detected. For example, dementia can be determined from conversation data. For such conversation data, learning data 17 can be generated in the conversation data conversion unit 12 and the learning data generation unit 13 in advance. This can also be learned each time according to the case.

Further, in the present invention, the artificial intelligence agent 5 is provided with a plurality of artificial intelligence agents, so that it is possible to speak in accordance with the artificial intelligence agent terminal user 23. For example, dialect localization. As an embodiment of the present invention, a framework called neural network or deep learning is used for the learning data generation unit 13, but generally, these machine learning devices take learning time, so general-purpose learning has priority. Has been done. However, in the present invention, the processing of the conversation data conversion unit 12 can reduce the number of learnings by half or less, and can shorten the learning time by about twice. Therefore, a plurality of artificial intelligence agents can be learned separately, and even by using the same neural network or deep learning framework, it is possible to make a judgment using a plurality of indices. The processing of the conversation data conversion unit 12 will be described later.

Furthermore, since localization allows conversation by using words used by local elderly people on a daily basis, impersonation by a third party other than the artificial intelligence agent terminal user 23 is difficult.

Needless to say, the artificial intelligence agent terminal user 23 can also transmit his / her danger through the artificial intelligence agent terminal 2.

In the present invention, characters, pictures, and sounds are assumed as conversation data. Characters are character input by normal keyboard input. A picture is a combination of letters and pictures. For example, it is a combination of the word good morning and the design of the character greeting. The voice is input to the artificial intelligence agent terminal 2 by voice.

Therefore, if the artificial intelligence agent terminal user 23 responds to the utterance with speech, if a third party other than the artificial intelligence agent terminal user 23 responds with letters or pictures, the specificity is detected, It can be seen that something is abnormal in the artificial intelligence agent terminal user 23.

In addition, when the peculiarity is detected, it is possible to contact the artificial intelligence agent terminal 2 directly from the watching terminal 4, but it is also possible for the artificial intelligence agent 5 to check the situation instead. For example, when the AI agent terminal user 23 inputs a conversation when the head hurts, the user confirms what has happened. About these conversation data, accuracy is improved by learning each time and generating learning data.

For example, for speech processing, it is possible to confirm the schedule of the artificial intelligence agent terminal user 23 and convey the schedule for the day. In addition to simply managing the schedule, you can check whether you remember the schedule.

Even when an artificial intelligence agent is used to terminate the speech, the specificity can be detected as in the case of speech.

Further, the end-call processing can be used for health management of the artificial intelligence agent terminal user 23. For example, when a specific time has come, the artificial intelligence agent terminal user 23 can be urged to go to bed by performing the end-call processing as shown in FIG.

Furthermore, since the state of the artificial intelligence agent terminal user 23 is confirmed through conversation, the artificial intelligence agent terminal user 23 does not feel that the artificial intelligence agent terminal user 23 is being monitored. Can be confirmed. Since the image of the artificial intelligence agent 5 displayed on the conversation display unit 8 may use not only an icon but also a photograph or the like, for example, by using a family photograph, a natural act of contacting the family The state of the artificial intelligence agent terminal user 23 can be confirmed.

For example, it is possible to periodically take pictures of meals and upload pictures through the artificial intelligence agent terminal 2.

In this way, it can serve as an index for grasping the state of the artificial intelligence agent terminal user 23 through the utterance and the end of the watching process. For this reason, not only can the safety of the artificial intelligence agent terminal user 23 be confirmed by the utterance and the end of the talk, but also the health status can be grasped and managed.

FIG. 12 is a diagram illustrating an example of learning data. This learning data is a flowchart showing the flow of response by the artificial intelligence agent in FIG. 5, and when the artificial intelligence agent terminal user 23 inputs the conversation data for good morning, the number calculated by the learning data 17 or recorded It is an example of the data. As shown in the flowchart showing the flow of learning data generation in FIG. 13, the numerical value of the learning data 17 is constantly updated. The learning data is weighted with numerical values based on a plurality of parameters. In the embodiment, numerical data of morning and night parameters is shown, but parameters such as time zone, temperature, and weather can be handled. Numerical data takes a value including the decimal point from 0 to 1, with 1 being the highest numerical value.

In the present invention, implementation by a neural network or a deep learning framework is assumed. Actually, when learning data 17 is referred to, counting is performed and a numerical value is calculated. However, for general-purpose conversation data, for example, greetings and the like can be increased in response speed by holding numerical values in advance in the recording unit 16 or the artificial intelligence agent 5. This is similar to the spinal reflex in humans. Also in such processing, the conversation data is transmitted to the watching server 3.

As shown in FIG. 12, in the morning time zone, when good morning conversation data is input, the good conversation numerical value is high as the conversation data to be answered. On the other hand, the numerical value is low in the night time zone.

From this value, the peculiarity of the conversation shown in the flowchart showing the flow of response by the artificial intelligence agent in FIG. Specifically, when the artificial intelligence agent terminal user 23 inputs “Good night” in the morning time zone, since the numerical value is within the specified range, the specificity can be detected. The number specified here is set by the watcher. The numerical value ranges from 0 to 1. In the above case, the specificity can be detected by specifying a numerical value such as 0.5.

This artificial intelligence agent user of the terminal 23 and night morning, can detect even specificity if you type Hello.

FIG. 13 is a flowchart showing a flow of learning data generation. Learning data 17 is generated from past conversation data. The learning data generation process shown in the figure is performed periodically while the watching server 3 is activated. It is also possible to start from the watching terminal 4.

The watching server 3 receives the conversation data and transmits it to the conversation data conversion unit 12 (step S401). This data is data for generating learning data, and is different from the conversation data 18 of the recording unit 16. After the processing is performed, it is recorded in the recording unit 16 together with the learning data.

The conversation data conversion unit 12 multidimensionalizes the data in order to increase the learning speed of the learning data generation unit 13 (step S402). The multidimensionalization of conversation data will be explained later.

The learning data generation unit 13 generates learning data from the data received from the conversation data conversion unit 12 (step S403). The learning data generation unit 13 is software generally called an artificial intelligence engine. In the present invention, implementation by a neural network or deep learning framework is assumed, but other machine learning devices can also be used.

The generated learning data 17 is recorded in the recording unit 16. The original conversation data is recorded in the conversation data 18 of the recording unit 16, and the process ends (step S404).

FIG. 14 is a diagram showing a program implementation comparison of the conversation data conversion unit 12. As shown in FIG. This is an implementation when the learning data generating unit 13 performs multi-dimensional learning data when learning is performed. When entering data into the framework of deep learning or neural networks, the numerical value takes a decimal value between 0 and 1. For example, in black-and-white image data, learning data is often input with a white portion being 0 and a black portion being 1, which is real learning data. In the present invention, the learning data is input in a multidimensional manner. Of course, as the input data becomes multidimensional, the data input and output portions of the neural network and deep learning framework are also changed.

FIG. 14 shows program implementation in the case of two dimensions, multidimensions, multidimensions (complex numbers), and dual, but the dimensions can be further expanded. In the present invention, a Hello, can correspond to good night 0.

FIG. 15 is a diagram illustrating an example of improving learning efficiency by multidimensionalizing conversation data. The vertical axis represents the error rate, the horizontal axis represents the number of learnings, and the data converted by the conversation data conversion unit 12 is compared with the number of learnings of the data that has not been converted. The graph with the lower line in FIG. 15 is an example when the conversation data conversion by multi-dimension (complex number) shown in FIG. 14 is performed.

In the case of data converted by the conversation data conversion unit 12 in FIG. 15, the number of learning times when the error rate reaches 10% is less than half of the number of learning times of the data that has not been converted. In the case of the same number of learning times, the calculation time of converted data and non-converted data is 1.1 times. The converted data takes more time to calculate than the unconverted data, but since the number of learnings can be reduced to half or less, the calculation time to reach the 10% error rate is converted data. Is about twice as short as the unconverted data.

With this technology, it is possible to perform learning using two types of neural networks and deep learning frameworks using the same computer.

Further, the processing by the conversation data conversion unit 12 does not change the structure of the neural network or deep learning framework. Therefore, it can be applied to current neural networks and deep learning frameworks.

Furthermore, the learning time can be further shortened by combining with a learning speed-up process that changes the learning rule of an existing neural network or deep learning framework.

When the same neural network is used for the two-dimensional, multi-dimensional, multi-dimensional (complex number), and dual shown in FIG. 14, the multi-dimensional (complex number) can most shorten the learning time. Further, the learning time reduction rate varies depending on adjustment of other parameters of the neural network even with the same data.

The learning data 17 and conversation data 18 of the artificial intelligence agent terminal user 23 are recorded in the recording unit 16 by repeating such learning data generation processing.

The reason why the generation of the learning data described so far is a technique for improving the accuracy of grasping the situation of the artificial intelligence agent terminal user 23 will be described in detail here.

For example, by shortening the learning time, it is possible to use a multi-layer neural network or a deep learning framework. Neural networks and deep learning frameworks are known to improve accuracy with more layers. By using the present invention, it is possible to improve the accuracy of results because the learning time does not change even if more layers are used than before.

In addition, shortening the learning time also makes it possible to use multiple neural networks and deep learning frameworks. Therefore, it is possible to make numerical judgments using a plurality of artificial intelligence, which has been difficult in the past. Specifically, there is generation of learning data by a neural network that has learned conversation data and a neural network that has learned speech data. When it is difficult to determine the situation of the artificial intelligence agent terminal user 23 only by learning data of conversation data, it is possible to improve the accuracy of detecting specificity by referring to the learning data of speech data.

Multiple neural networks and deep learning frameworks make it easy to create general-purpose learning data and learning data tailored to the artificial intelligence agent terminal user 23, so it supports local dialects and uses artificial intelligence agent terminals. It is possible to generate learning data according to the age, sex, etc. of the person 23.

Furthermore, the learning time can be reduced by shortening the learning time. For example, what took 24 hours to learn once is completed in 12 hours, so it is possible to cope with the situation of the artificial intelligence agent terminal user 23 by learning new words each time.

Specifically, there are localization of dialects and new words. For example, when a new word is used, the specificity is detected, but by learning the conversation data, processing is performed so that the specificity is not detected when a new word is used again. Is possible.

As a secondary effect of newly learning, there is no effect of modifying the program so that it does not react to the words, so there is an effect that the maintainability of the software is improved.

Since the learning data 17 can be generated for each artificial intelligence agent, when the artificial intelligence agent terminal user 23 uses it for the first time, the artificial intelligence agent that matches the situation of the artificial intelligence agent terminal user 23 is used. Thus, it is possible to improve the accuracy of utterances, end conversations, and responses from the beginning of use.

In addition, by making the numerical data used for learning multidimensional, it is possible to learn multiple learnings using a single neural network or deep learning framework. For example, both conversational data and voice data can be learned using a single neural network or deep learning framework. Specifically, image discrimination and dialogue can be realized simultaneously with one program.

A side effect that can shorten the learning time is that it reduces the use of computer resources. Therefore, since it is not necessary to use a high performance computer, the server cost to be used is reduced and the power consumption is suppressed.

In the present invention, the watching server 3 and the artificial intelligence agent terminal 2 are separated as an example of implementation, but since the use of computer resources can be reduced, learning data may be generated and recorded in a smartphone or tablet terminal. Is possible. Specifically, it can be incorporated into the response unit 6 and the conversation start unit 7 of the artificial intelligence agent 5.

As a result, the artificial intelligence agent terminal 2 can combine a plurality of artificial intelligence agents and a plurality of learning data 17. Even if there are multiple artificial intelligence agents, if the same learning data is used, it will be the same conversation data, so it is easy to notice that it is due to a machine.

In addition, if the learning data 17 is not updated, the same mistake will always be repeated, but in the present invention, since learning can be performed each time, it is possible to utter, end, and respond to the artificial intelligence agent terminal user 23, The accuracy of grasping the situation of the artificial intelligence agent terminal user 23 can be improved.

Thus, the process of generating learning data can be a technique for improving the accuracy of grasping the state of the artificial intelligence agent terminal user 23.

As mentioned above, although embodiment of this invention was described, this invention is not limited by embodiment. The present invention includes combinations of the embodiments and the following modifications as appropriate, and includes equivalents thereof.

For example, the artificial intelligence agent terminal 2 is not limited to smartphones, tablets, and robots, but is a terminal that can communicate with the artificial intelligence agent terminal user 23, and is a wearable device or mobile phone that can transmit and receive data to and from the watching server 3 via a communication line. But you can. Moreover, since it is sufficient that the numerical value calculated by the learning data 17 can be confirmed by transmission / reception of conversation data or voice data, the conversation data or voice data may be transmitted / received by e-mail or the like. Furthermore, by combining the present invention with an existing chat application, it is possible to grasp the state of the artificial intelligence agent terminal user 23 by using conversation data learning and peculiarity detection without using the artificial intelligence agent 5. Is possible.

For example, the watching server 3 may use a cloud service. In addition, a resource that is external to the watching server 3 may be used instead of the watching server 3 watching the neural network and deep learning framework of the learning data generation unit 13.

Further, for example, the watch terminal 4 is not limited to a smartphone, tablet, PC, or robot, and is a terminal capable of talking with the artificial intelligence agent terminal user 23, and transmits and receives data via the communication line with the artificial intelligence agent terminal 2 and the watch server 3. It may be a wearable device or mobile phone. Moreover, since it is sufficient that the numerical value calculated by the learning data 17 can be confirmed by transmission / reception of conversation data or voice data, the conversation data or voice data may be transmitted / received by e-mail or the like. Further, by combining the present invention with an existing chat application, it may be used as an emergency contact to the artificial intelligence agent terminal 2. The watching terminal 4 is not limited to a family or a caregiver, but may be a terminal used by a local government or community administrator. In addition, the artificial intelligence agent 5 may have a conversation from the watching terminal 4 via the watching server 3.

Further, for example, the artificial intelligence agent of the artificial intelligence agent 5 may use artificial inability. In that case, the conversation data is watched and transmitted to the server 3, and the numerical value calculated by the learning data 17 and the conversation data are watched and transmitted to the terminal 4 or the like. Also, artificial intelligence agent and deep learning and neural network framework, learning data, conversation data are not one-to-one correspondence, multiple deep learning and neural network framework, learning data, behind one artificial intelligence agent Conversation data may be used. Of course, since there are a plurality of artificial intelligence agents, an N-to-N structure may be used.

Further, for example, the initial data of the learning data 17 may be a general-purpose conversation data corpus. In the present invention, learning is performed each time, but it is possible to perform only detection of specificity without performing learning each time.

Further, for example, the display unit 21 of the watching terminal 4 may display the conversation display units 8 of the plurality of artificial intelligence agent terminals 2. The data to be displayed may be combined with the numerical data weighted on the learning data 17, and the importance of the conversation may be displayed by color or the like.

By displaying the conversation display units 8 of the plurality of artificial intelligence agent terminals 2, it is possible to easily check the status of the plurality of artificial intelligence agent terminal users 23.

1 Watch system
2 Artificial intelligence agent terminal
3 Monitoring server
4 Watch terminal
5 Artificial intelligence agent
6 Response section
7 Conversation start
8 Conversation display
9 Artificial intelligence agent terminal transmitter
10 Artificial intelligence agent terminal receiver
11 Monitoring server receiver
12 Conversation data converter
13 Learning data generator
14 Notification section
15 Monitoring server transmitter
16 Recording section
17 Learning data
18 Conversation data
19 Watching terminal receiver
20 Monitor terminal transmitter
21 Display
22 Conversation data input section
23 Artificial intelligence agent terminal users
24 Communication line

Claims (6)

It is equipped with an artificial intelligence agent terminal, a monitoring server and a monitoring terminal connected via a communication line,
The artificial intelligence agent terminal comprises a plurality of artificial intelligence agents,
Sending and receiving means for watching conversation data and exchanging with the server and watching terminal,
A conversation display means for exchanging conversation data;
When the conversation data is received, a response means for referring to the learning data of the watching server and returning the conversation data;
Referring to conversation data and learning data of the watching server, and having a conversation start means for starting a conversation,
The monitoring server includes conversation data conversion means for converting conversation data received from the artificial intelligence agent terminal into multidimensional data;
Learning data generating means for generating learning data that is recorded with new conversation data from the conversation data converted into multidimensional data and weighted by numerical values of the conversation data;
A transmission / reception means for watching conversation data and learning data and interacting with the terminal;
A recording means for recording conversation data and learning data;
Extracting conversation data exceeding a threshold value from the numerical value of the learning data, comprising a notification means for transmitting to the watch terminal,
The watching terminal is a transmission / reception means for exchanging conversation data and learning data with the artificial intelligence agent terminal and the watching server and the watching terminal.
Display means for displaying data received from the watching server;
A watching system comprising means for inputting conversation data. The artificial intelligence agent comprises transmission / reception means for interacting with a plurality of artificial intelligence agents,
Means for referring to the learning data of the watching server, and having a conversation between the artificial intelligence agents by the conversation display means;
2. The watching system according to claim 1, further comprising means for performing conversation between the owner of the artificial intelligence agent terminal and a plurality of artificial intelligence agents. The learning means comprises a conversation data conversion means for converting conversation data into multidimensional data, a new conversation data recording means, and means for generating learning data weighted by the numerical value of the conversation data,
The conversation start means refers to the generated learning data, acquires the conversation start conversation data from a weighted numerical value, and transmits the conversation data to the artificial intelligence agent terminal. The monitoring system according to any one of 2 above. 4. The watching system according to claim 1, wherein the conversation display means displays the conversation data with characters, pictures, and sounds. Conversation data conversion means for converting the two-dimensional conversation data received from the artificial intelligence agent terminal into multidimensional data;
Learning data generating means for generating learning data that is recorded with new conversation data from the conversation data converted into multidimensional data and weighted by numerical values of the conversation data;
A transmission / reception means for watching conversation data and learning data and interacting with the terminal;
A recording means for recording conversation data and learning data;
A monitoring server comprising means for extracting conversation data exceeding a threshold value from the numerical value of the learning data and transmitting it to the monitoring terminal. Learning data obtained by acquiring conversation data from the owner of the artificial intelligence agent terminal and multiple artificial intelligence agents, converting the acquired conversation data into multidimensional data, and recording new conversation data and weighting the conversation data numerically And a means for extracting conversation data exceeding a threshold value from the numerical value of the learning data and transmitting the extracted conversation data to the watching terminal.