JP2020177367A - Computer system for edge-driven collaborative ai, and program and method therefor - Google Patents

Abstract

To provide a computer system, a program and a method capable of requesting statistics on questionnaire results by others in the past as well as in the future and statistics on phenomena related to a specified viewpoint.SOLUTION: Questionnaire implementation result information derived from input by a user is transferred together with identification information of each user, and returned it to the user after adding a viewpoint at the time and taking statistics. Content recognized by a terminal is converted so that it can easily take statistics by inference rules and machine learning. A favorable action plan, etc., are automatically recommended based on the statistics. The machine learning for the inference rules increases an element of a conditional set step by step so as to make the effect unique, and also increases or decreases the element of the conditional set step by step so as to determine the effect stochastically. Only inference rule, which is in the case the probability of the effect when the element of the conditional set is increased or decreased step by step changes, is precipitated. Constituent requirements that are generalized as a character string is precipitated by converting the same partial character string in a plurality of character strings constituting the same possible world into the same variable.SELECTED DRAWING: Figure 1

Description

The present invention relates to computer systems, in particular artificial intelligence (AI) using computer systems.

It may differ depending on the target domain (field) of the AI system, the specific site, and the user.
1) Data format (for example, natural language sentence pattern and style, schema in RDB / XML / JSON format, standard vocabulary, etc.),
2) Machine learning resources (for example, data sets (even if they include teacher signals), neural networks as a result of learning (hereinafter abbreviated as NN), hypotheses, statistics, etc.),
3) A set of inference rules (mainly used for deduction) that has been selected and edited (hereinafter referred to as customization) for the user (including the person) by a human being (knowledge engineer or system administrator for each site). ),
Because there is no system that can handle such things in an integrated manner (inference system, learning system, or hypothesis inference system that can do both)
Even if it is barely possible for a particular area (domain)
To new knowledge and philosophy in the "interdisciplinary" field (cross domain)
There was no system that could efficiently utilize the amount of information (including potential information) given by users.
Patent Document 1 made it possible to utilize the amount of information to some extent by preparing rules for standardizing terms and sentence patterns that may differ depending on the domain and site, but based on machine learning and the latest statistics of other people's information. We do not support the questionnaire and the request from the viewpoint of the phenomenon.
The main use was the information entered in the past.

Japanese Patent Application No. 2013-8708

Therefore, the subject of the present invention may differ depending on the target domain of the AI system, the specific site, and the user.
1) Data format,
2) Machine learning resources,
3) Customized set of inference rules,
To provide a system (inference system, learning system, or hypothesis inference system that can handle both) in an integrated manner.
In particular,
Not just what was entered in the past
It is to provide an integrated AI system that can request statistics on the results of questionnaires by others in the future and statistics on phenomena related to a specified viewpoint.

The present invention
The recognition result information of the questionnaire, which is based entirely or partially on the format or options derived from the input from the user set, and the recognition result information of the questionnaire acquired by the computer terminal in the field is related to the computer terminal or the terminal. The first means of transferring together with the user's identification information, the second means of automatically statistically processing the transferred recognition result information and the identification information from the viewpoint of the time of automatic recognition, and the automatic statistical processing result are automatically performed. Provided is a computer system having a third means of distribution.
: As an example of "questionnaire based entirely or partially on the format or choices derived from the input from the user set", for the questionnaire consisting of check boxes, radio buttons and free input fields, go to the past free input fields. The questionnaire format (of a certain period version) was created by automatically making the free texts that are frequently entered into check box items when presenting a new questionnaire, or by deleting the check box items that are infrequently checked. Here are two examples.
: In addition, any one of some contents entered by the user in multiple free input fields etc. will be used as a new questionnaire title, and some of the other input contents will be added check boxes (set). This also applies to a flexible questionnaire system based on a server = client system, etc., which allows you to specify (in an input form, etc.) that you want to create a new questionnaire that has been adopted (further upgraded).
: "Obtaining on-site" is mainly for users to answer questionnaires, but it is not limited to this.
> As a result, a dynamic questionnaire (with a certain amount of format) based on user input efficiency and timely statistical collection needs,
In addition to enabling related efficient statistical processing and machine learning (predicate logic level inference rules and conversion rules that abstract the same character string into variables can also be deposited),
Moreover,
It is also possible to deposit inference rules with posterior probabilities according to assumptions.
It is possible to provide a statistical / hypothesis inference system that can handle 1) data format, 2) machine learning resources, and 3) customized inference rule set, which may differ depending on the user of the AI system and the site.

The present invention also
A computer system that transfers recognition result information of voice information, imaging information, or both automatically recognized by a computer terminal regardless of whether the first means is in the form of a questionnaire, together with identification information of the computer terminal or the related user. (Device) is provided.
: The recognition result information can include not only the classification but also the estimated value and the calculation result by combining them.
: The "second means" in this case is
Not only the server and cloud immediately after the above transfer
It may exist in another user terminal to which it is delivered as pre-statistical data.
In this case, as edge computing, statistical analysis and machine learning (including the precipitation of hypotheses by inductive reasoning) are performed.
> By doing this, the amount of extra information and fluctuations that are automatically recognized (having unique properties) and exist in the on-site voice and images can be reduced (by empirical rules according to the unique properties), or background knowledge and statistical knowledge. It is possible to perform efficient transfer and efficient statistical processing by saving only the amount of information about the standard information compressed by. This makes it easier to use the mail protocol, FTPS / HTTPS protocol, and the like.
>"Uniquenature", for example, regarding language, automatic recognition setting (language (Japanese / English / ...), dialect ("mass"->"let'sdo"), unique Nouns ("teruko"->"mother"), etc. can be considered.
> In addition, uniform processing of recognition result information (particularly linguistically standardized) by the server side (timely notification of a request for a person who wishes to board a regional shared bus, etc.) becomes easy. In the case of a shared bus, etc., the recognition result information or identification information for sending to the central WEB server (including the cloud; the same applies hereinafter) includes information on the installation location (= transmission location) of the computer terminal. It is advantageous to link them in advance.
: The recognition result information sent to the central WEB server can be forwarded by e-mail to a supporter (the caregiver in charge or the person in charge of the area of the shared bus company) that is further defined based on the content. ..

The present invention also
Provided is a computer system having means for performing a predetermined format conversion or a predetermined knowledge inference (deduction) so that the recognition result information matches the format or the option of claim 1.
: "Predetermined knowledge reasoning (deduction)" can be carried out by, for example, the logic programming language Prolog or a system that enables it in Japanese.
> This facilitates statistical processing and statistical machine learning (including those based on a few examples) that absorbs simple notational fluctuations and semantic marking fluctuations (when equivalent meanings can be obtained based on constituent requirements). You can get the data.
: In addition, as machine learning based on "small examples", a set of logical possible worlds consisting of only a few examples (each possible world consists of, for example, a set of character strings representing logic) is a so-called "closed world". It is advantageous to machine-learn the constituent requirements (including the constituent requirements with the occurrence probabilities in each closed world) that can be deposited as "" (claim 11).

The present invention also
For recognition result information or identification information
Provide a computer system that includes information about the installation location (= transmission location) of a computer terminal or is linked in advance.
: "Regarding the installation location" may be an address, IP address, MAC address, latitude / longitude, etc. In the case of an IP address, etc., service cooperation is performed by utilizing an information table that links it with latitude / longitude and address data. Is done.
> This makes it easier to link statistics from data and learning results to various regional physical services.

The present invention also
Automatically recommend a favorable or unfavorable action plan (part or all of 5W1H or a series thereof) to the user related to the computer terminal to the user or the associated user based on the result of the automatic statistical processing. To provide a computer system characterized by.
: "Related user (the user)" refers to a user who has moved into a long-term care facility, a user who lives alone in a general house, a tourist who is near a place where a computer terminal is installed, and the like.
: "Associated users" refers to family members, care contractors, and government agencies that promote tourism, etc., for long-term care users and users living alone.
> This makes it possible to connect hypotheses inductively obtained by statistics and machine learning to concrete actions (for myself and for support) and deduction (synthesis of judgments and ideas).

The present invention also
When a computer automatically recommends to a user or an associated user
The recognition result information itself used for the automatic statistics (which may or may not be narrowed down to the recognition result information about the user),
Provided is a computer system that deduces a combination of inference rules machine-learned from a plurality of recognition result information and / or inference rule sets input by humans.
: Here, the "human-input inference rule set" may be an inference rule set selected and edited by a human for the customer (the above-mentioned "customized inference rule set"), and is more general. It may include scientific causal relationships, semantic conversion rules, rank information (values) and taboo information from various viewpoints, and the like.
> By doing this, the facts, hypotheses, theories, and philosophies that form the basis of the recommendation can be obtained.
It is possible to answer the depth required for the chain of "why", "why" ... up to the general common sense that the user can understand.

The present invention also
Provided is a computer system that displays the content used as the basis for deduction of automatic recommendation as the basis for recommendation.
: Here, "deduction" includes the case where some of the hypothesis sets (fact form and rule form) obtained by machine learning are used assuming "true", so hypothesis reasoning is also included. However, the adoption of hypotheses that clearly contradict each other constitutes different solutions and ideas (displayed so that they can be understood) as multiple "possible worlds" to be separated.
> This enables safer and more effective operation by confirming the basis of dynamically generated timely reasoning regardless of whether it is based on a machine or a human.

The present invention also
When the user or the associated user edits a story such as a novel, a dissertation, a report or a proposal
Recognition result information that the computer is related to the user or the associated user,
Inference rules machine-learned from multiple recognition result information,
A set of inference rules entered by the user or the associated user,
Content in the middle of editing or
A combination of these,
Provides a computer system that automatically recommends the addition of deduced content from.
: For "addition", it may be recommended to recursively add a sentence at the beginning or end of the sentence being edited, or to add it as a sentence in the middle of the sentence being edited. In the simplest case, the default position to be inserted is set to the end of the sentence, and it is automatically recommended that it should be inserted at that position only when the user embeds $ X in the sentence as an exception. In the simplest case, for example, the sentence is regarded as a unit (gram), and the similarity of the surface character arrangement between two sentences (for example, the similarity of the d-bigram set of each character constituting the sentence) is regarded as the sentence matching degree. It can be recommended by the most probable selection based on sentence-based n-gram statistics and sentence-based d-bigram statistics.
> This makes it easier for anyone to write novels based on questionnaires on philosophical issues and tourism that they are interested in.
It becomes easier to change the experiences and sensibilities of people close to us into shareable values.

The present invention also provides the automatically recommended content of the computer.
Provided is a computer system that can be used as a material for dynamically generating a script of an adventure game or a role-playing game related to the facts, hypotheses or philosophies used in the deduction.
: For these games,
This includes real-life games in which real local citizens "play" local police and witnesses, table-talk games, and VR / AR / MR games linked to computer screens.
: In addition to real local citizens, the computer terminal of the tourist information center can be configured to play such a role.
> This makes it possible to enjoy social games with dynamic and rational new developments.

In the present invention, the recognition result information is also related to tourism.
Provided is a computer system that is a mixture of the geography of the tourist object and general philosophy related to history.
> This gives you the opportunity to come into contact with general philosophies in the process of understanding local geography and history.
> Also, the more you travel, the more general philosophies (history, technology, morals, languages, anthropology, related bonds) about other than the destination (not limited to locals, but the way of the heart). You can also study deeply) with people who visited).
: "Mixed" here means deeper understanding of the geography and history of the tourist object (the cultural concept necessary to understand and the philosophy not limited to the tourist object) through dialogue with the computer system. Or it is a comprehensive content with more generality and context that can be obtained by going back more broadly (there may be a time series due to the order of dialogue and the arrangement in the content) or its image.

The present invention also
The order of items that are easier for the user to study by interacting with the computer system (eg, morality, history, philosophy) or the order of items that are more accessible to the user.
Inferring from the statistics of dialogues with other users,
Provide a computer system that automatically controls the content of dialogue (choices, prompt expressions for free input, etc.) according to the estimation result.
> This allows for efficient "mixing".

The present invention also
The inference rule that is deposited is machine learning consisting of a conditional set (for example, a set of existence or nonexistence of a character string) and an effect (in the same example, the existence or nonexistence of a character string not included in the conditional set).
By gradually increasing the elements of the condition set (that is, tightening the data narrowing conditions)
Provided is a computer system that performs machine learning so that an effect (existence or absence of a character string) in a given context set (set of cases) is uniquely determined.
> This makes it possible to deposit inference rules with the minimum conditions (state of maximum abstraction).

The present invention also
The inference rule that is deposited is machine learning consisting of a conditional set (for example, a set of existence or nonexistence of a character string) and an effect (in the same example, the existence or nonexistence of a character string not included in the conditional set).
By increasing or decreasing the elements of the condition set in stages (that is, by tightening or loosening the data filtering conditions)
Provided is a computer system that performs machine learning so that an effect (existence or absence of a character string) in a given context set (set of cases) is stochastically determined.
> This makes it possible to make inferences for the degree of probability.

The present invention also
Provided is a computer system in which only inference rules are precipitated when the probability of an effect when the elements of a conditional set are increased or decreased stepwise changes.
> This makes it possible to deposit an effective set of representative rules while suppressing the amount of information of the set of inference rules to be deposited.

The present invention also
Constituent requirements generalized as character strings by converting the same substrings in multiple character strings that make up the same possible world (context) into the same variable (variables common to each fixed character string that makes up the head and body) Provided is a computer system characterized by precipitating a horn clause, etc. expressed in an embedded form.
> This will increase the number of cases where generalization can be easily achieved in both edge computing and the central server (or cloud) to achieve the purpose.

The present invention also provides computer programs for such computer systems.

The present invention also provides processing methods for such computer systems.

It is a figure which shows the example of the AI system by this invention which is operated mainly in the observation and control of the phenomenon including the user in the form which each of a user and the supporter manages a terminal. It is a figure which shows the example of the AI system operated with a focus on the control of the questionnaire by this invention which a user who does not have a specific supporter starts to collect by revising while responding. It is a figure which shows the example of the AI system by this invention which performs the setting and inference mainly on the terminal side for the dialogue with the user who does not have a specific supporter. It is a figure which shows the general structure of the conventional AI speaker. It is a figure which shows the feature of programming (Japanese Patent Application No. 2013-87008 etc.) for performing full-scale thinking in the terminal of this invention. It is a figure which shows the easy-to-understand example 1 of learning (precipitation of an inference rule) which regarded a few cases (two contexts) as a closed world. It is a figure which shows the easy-to-understand example 2 of learning (precipitation of an inference rule) which regarded a few cases (five contexts) as a closed world.

It may differ depending on the target area of the AI system, the specific site, and the user.
data form,
Machine learning resources,
Various forms for providing a system that can handle a customized set of inference rules in an integrated manner will be described below, but the outline is as follows.
The questionnaire implementation result information derived from the input from the user is transferred together with the identification information of each user, and the viewpoint at the time is added to the statistics and returned to the user. In addition to user input, voice and images are also automatically acquired. The content recognized by the terminal is converted so that it can be easily statisticalized by inference rules and machine learning. A favorable action plan, etc. is automatically recommended based on the statistics. Not only questionnaires and actions, but also the development of the story being written and the script of the game are automatically recommended based on statistics and machine learning. Regarding tourism, we will deal with a mixture of geography and history and general philosophies related to tourism. For the purpose of studying, the order in which it is easy to study is automatically recommended. Machine learning of inference rules makes the effect unique by increasing the elements of the conditional set step by step. Alternatively, similarly, the elements of the condition set are increased or decreased stepwise so as to be stochastically determined. Only the inference rule when the probability of the effect when the elements of the conditional set are increased or decreased stepwise changes is precipitated. By converting the same sub-character string in multiple character strings that make up the same possible world into the same variable, the constituent requirements that are generalized as a character string are precipitated.

Embodiment 1 of FIGS. 1 and 2
In particular, we provide not only those entered in the past but also those that can request statistics on the results of questionnaires by others in the future and statistics on phenomena related to a specified viewpoint.
The statistical processing and learning processing in this form are based on the data (information and knowledge, statistics / learning resources) accumulated in the storage device 3.
It is performed by a group of servers that make up the management cloud, WebAPI, and so on.
In this example, many users respond to questionnaires defined by other users as shown in Fig. 1, and information (mainly the management cloud, etc.) is sent to the system (mainly the management cloud, etc.) by words and actions (images, sounds, etc.) at the site. It only provides (the basis of statistics and learning), but some users enter a response to the terminal that independently defines the questionnaire for which they want to collect statistics, as shown in Fig. 2. As a result, it is possible to obtain a questionnaire result (simply a data set) from the viewpoint of wanting to obtain statistics and a tendency of statistics / learning from the questionnaire result from another user set.
This "independent definition" is to "reply" a new questionnaire request and item definition that you want to be conducted via the cloud via a terminal near the site using XML text (including a unique questionnaire name ID). Alternatively, it is carried out by increasing or decreasing the items of the questionnaire defined by other users or the questionnaire prepared by default, giving a different questionnaire name ID, and "replying".
When a "mere data set" is obtained from a cloud or the like, machine learning such as that related to FIGS. 6 and 7, 2018-175748 (automatic precipitation of horn clause (inference rule) having variables) is performed on the terminal side. If such learning is performed on the cloud side, the inference rules deposited in the cloud will be obtained as a terminal.
In addition, for the items of the questionnaire and the set of choices for each item (and how to arrange them), various users recursively increase / decrease the items and the questionnaire name ID (this system) based on the predetermined default questionnaire definition and this. It is also possible to generate a questionnaire definition (for example, an XML file) that is automatically optimized based on the results of statistics and machine learning by the cloud or a terminal as well as the one given (by operating). The generated survey definition can be shared and reused with appropriate access privileges.
Some of the questionnaire items can be defined to be automatically filled by automatically collecting images and sounds for each user (or for each site of the related terminal) by the terminal function. At this time, the viewpoints of "time", "co-occurrence", and "order" related to the data guarantee the cohesiveness, identity, simultaneity, and relevance of the context, but systems and questionnaires that do not require such guarantee. In the definition, it can be configured to process automatic statistics and machine learning, ignoring the viewpoints of "time", "co-occurrence", and "order".
In the system of this embodiment, the statistical results of the questionnaire, etc. are automatically distributed only to the users who responded to the questionnaire and the users who have permitted the automatic acquisition of images and sounds, and the users can view or perform additional machine learning, etc. I am trying to do it.
As a result, the user can easily collect the latest analysis results and the like regarding better lifestyle habits.
The data set automatically distributed here can be automatically or manually narrowed down to data from other user sets that match the age, gender, and other attributes replied by the users. In the case of manual operation, the user can inform the management cloud of the narrowing characteristics of the desired data set by turning on the check box of the item set to be narrowed down at the time of "reply".
"Information derived from input from a set of users" is comprehensively accumulated through the above process, and when each is set as shared content in the cloud or the Internet, it can be said to be the property of humankind. , Will form a vast set of "recognition ontology" across a wide range of fields.
In many cases, each user reads useful recognition ontology, questionnaire definition (choices and blanks), inference rules, learned neural network (NN), etc. into a terminal near the site via the net. Then, the result of automatic recognition (including explicit "reply") based on the ontology etc. will be automatically transferred to the management cloud etc. together with the user identification information etc.
The inference rule not only summarizes the meanings expressed in multiple questionnaire items according to the constituent requirements, but also normalizes the item expressions and recognition formats (eliminating notational fluctuations, etc.) so that statistical processing is easy. Neural networks are useful, and neural networks are particularly useful for classifying images and sounds with a large amount of information into predetermined texts, digitizing them into predetermined units, and performing statistical processing.
In this embodiment, in addition to the user selecting the recognition ontology by himself / herself, in case the user is a support target person such as long-term care, a mobile supporter performs a recognition ontology of a terminal near the site, etc. Is explicitly manually set so that the automatic recognition of (supporting) on-site phenomena related to the user can be automatically operated over multiple continuous and discrete "time". The statistical result (B) related to the recognition ontology, etc. set here is automatically distributed to the supporter / site control terminal of the mobile supporter, not to the user / terminal near the site, and is displayed to the mobile supporter, etc. It is configured to enable further edge-type machine learning.
The "blank" in the transmission form or the like displayed on the terminal is useful for allowing a reply with a free ontology (free sentence), and can also be used for inputting the questionnaire name ID.
The recognition ontology is more flexible than the parsing / semantic analysis system for the standard language of a given language, and can automatically recognize dialects, new words, coined words, corporate codes, etc., and perform wider statistical processing by inference rules, etc. It can be converted into an expression.
By doing so, each user, a set of users with the same attributes as each user, each user's mobile supporter's intention to improve their habits, objective attitude, and automatic precipitation of the latest improvement method. And recommendations can be implemented in an integrated manner at different sites.
That is, in the present embodiment, the format is converted or the predetermined knowledge inference (deduction) is performed so that the recognition result information matches the questionnaire format or options.
In addition, in the case of a long-term care recipient or the like, information on the installation location of the computer terminal is linked in advance to the recognition result information or identification information.
In addition, an action plan for related users is automatically recommended based on the results of statistical processing. For example, if statistical results are obtained that elderly men with similar attributes feel a lot of happiness when they join a ballroom dance circle, then an elderly man who is not in a ballroom dance circle will be assigned to a ballroom dance circle. Configure computer programs for the management cloud and devices to automatically advise you to enter.
In making this automatic recommendation, the recognition result information itself used for automatic statistics, inference rules machine-learned from these, and / or inference rule sets input by humans are combined and deduced. At this time, by displaying the content used as the basis of deduction and clarifying the basis of advice and recommendations, it is easy for users and supporters to ignore the advice and respond to exceptions and arrangements. I have to.
FIG. 3 is a diagram showing an example of an AI system (“personal AI speaker”) according to the present invention in which settings and inferences for dialogue with a user who does not have a specific supporter are mainly performed on the terminal side.
In this embodiment, by installing the dedicated software according to the present invention on a personal computer equipped with a camera and a microphone, voice recognition, image recognition, full-scale inference (Japanese Patent Application No. 2013-8708), and necessary are required at the terminal (edge). Dynamic quotations of the latest data and inference rules (logic) from the net (Japanese Patent Application No. 2015-219007) are performed.
On the other hand, in the cloud, big data accumulated by manually programmed facts and inference rules (Japanese Patent Application No. 2013-8708), questionnaire results and automatic recognition results in shared folders that are open to the Internet or shared only by organizations and the like. Hypotheses (qualitative logic, etc.), statistical values (calculation logic, etc.), NNs (neural nets, deep learning results, etc.), procedural knowledge, multimedia data, corpus, and other context sets and values learned by subsets. The definition of taboo (Japanese Patent Application No. 2015-234356, Japanese Patent Application No. 2017-092057) is included.
On the other hand, in the past, as shown in Fig. 4, AI in the cloud performed recognition and inference, so the degree of freedom for each user, site, and supporter and the degree of involvement in the entire AI system were low, and the elderly.・ It was difficult to obtain a “motivation to live” as a caregiver / system improver.
FIG. 5 is a diagram showing the features of programming (Japanese Patent Application No. 2013-87008, etc.) for performing full-scale thinking in the terminal of the present invention. The advantages of FIG. 3 and the features that expand the industrial benefits are described.
FIG. 6 is a diagram showing an easy-to-understand example 1 of learning (precipitation of inference rules) in which "small cases" (two contexts) due to the large "narrowing down" by the user are regarded as a closed world. Learning based on such a small number of cases is easy to utilize the learning ability (calculation ability) having a small edge, and therefore has a particularly high affinity with the system according to the present invention.
FIG. 7 is a diagram showing an easy-to-understand example 2 of learning (precipitation of inference rules) in which a small number of cases (five contexts) are regarded as a closed world. In this case, it can also be shown that it is possible to learn an inference rule with a constituent condition of "A and B". Also, by replacing "Kanazawa" with a variable such as "$ X", it becomes possible to precipitate generalized inference rules.

Further, in another embodiment 2,
Recognition result information related to the user or the associated user when the computer edits a story such as a novel, a dissertation, a report or a proposal by the user or the associated user.
Inference rules machine-learned from multiple recognition result information,
A set of inference rules entered by the user or the associated user,
Content in the middle of editing,
Alternatively, it is automatically recommended to add the content deduced from these combinations.
In the second embodiment, the user can input by voice or text box.
Give the text in the middle of editing to the terminal,
Likelihood of the contents that can be surely obtained by the terminal or cloud by deducing the inference rule set from the above various sources and this sentence (consisting of multiple sentences) and some candidate contents with high probability. Recommendations will be displayed in descending order of. On the other hand, when the user approves or selects, a new input to the terminal is confirmed, which leads to further recommendation display.
Also, in another related embodiment 2A, this automatically recommended content is
To dynamically generate scripts for role-playing games related to a specific range of philosophies (philosophy that is not the whole of publicly shared philosophies) and (consistent) worldviews used in these deductions. Used as a material.

Further, in the third embodiment,
The recognition result information is related to tourism,
It is a mixture of the geography and history of the tourist object and the general philosophy associated with it.
As an example of recognition result information mixed with general philosophy,
"Kenrokuen tour""It was fun""The lessons of the Maeda family were the samurai's simple and strong and cultural recommendations""The attitude of the Sengoku warlords ...""War is an extension of diplomacy ...・ ”“ Diplomatic is one of the means of politics ... ”“ Politics is to decide the decision of the group ”can be answered in one questionnaire or accumulated in the cloud. As one of the contextual units, it is collectively the subject of statistics and machine learning, and each text is the subject of knowledge advice and action recommendation to users.

In the fourth embodiment
The order of items that the user can easily study by interacting with the computer system or the order of items that are suitable for the user to touch by dialogue.
Infer from the statistics of dialogue with other users, and automatically control the content of the dialogue according to the estimation result.
Specifically, in the form of FIG. 3, the optimum learning order in which statistics are taken with the viewpoint of time is reproduced is reproduced.
The optimum learning order is determined by the learning order of the user with the closest knowledge at present described in Japanese Patent Application No. 2012-218899.

In the above embodiments 1 to 4,
When the inference rule to be deposited precipitates an inference rule consisting of a condition set and an effect,
By gradually increasing the elements of the conditional set
Machine learning is performed so that the effect in a given context set is uniquely determined. An example is shown below.

Context 1:
The season is spring It rained in Kanazawa It rained in Komatsu

Context 2:
The season is summer It rained in Kanazawa It didn't rain in Komatsu

Context 3:
The season is spring It rained in Kanazawa It rained in Komatsu

Stage 1 It rained in Kanazawa ⇒ It rained in Komatsu (2/3)
Stage 2 It rained in Kanazawa & spring in the season ⇒ It rained in Komatsu (2/2 = unique effect)

In yet another embodiment
When the inference rule to be deposited precipitates an inference rule consisting of a condition set and an effect,
By increasing or decreasing the elements of the conditional set in stages
Machine learning is performed so that the effect in a given context set is stochastically determined.
This is useful when making recommendations based on the above stochastic reasoning.

In yet another embodiment
Similarly
Only the inference rule when the probability of the effect when the elements of the conditional set are increased or decreased stepwise changes is precipitated.

In any embodiment
By converting the same sub-character string in multiple character strings that make up the same possible world into the same variable, the constituent requirements that are generalized as a character string are precipitated. An example is shown below.

Stage 1 It rained in Kanazawa & spring in the season ⇒ It rained in Komatsu (2/2 = unique effect)
Stage 2 $ X fell in Kanazawa & Spring in season ⇒ $ X fell in Komatsu (2/2 = unique effect)

1 Setting files for users / terminals near the site (ontology for voice / image recognition (list of recognition terms, files for cascade classifier, etc.), questionnaire specification set, terminal learning / learning result data, selective editing by supporters (on the Internet) Contains inference rule sets and fact data selected from shared libraries and / or edited as necessary)
2 Setting files for supporters / field control terminals (recognition ontology provided to users, questionnaire specification sets (templates, etc.), terminal learning / learning result data, inference rule sets and fact data selected and edited by supporters, etc. Is stored)
3 Setting files shared / limited to the cloud or in-house LAN (ontology for voice recognition, questionnaire specifications and result set (versioned template set and each corresponding questionnaire result set), cloud learning / learning result data, etc. Is stored)

Claims (17)

The recognition result information of the questionnaire, which is based entirely or partially on the format or options derived from the input from the user set, and the recognition result information of the questionnaire acquired by the computer terminal in the field is related to the computer terminal or the terminal. The first means of transferring together with the user's identification information, the second means of automatically statistically processing the transferred recognition result information and the identification information from the viewpoint of the time of automatic recognition, and the automatic statistical processing result are automatically performed. A computer system having a third means of delivery. A claim for transferring recognition result information of voice information, imaging information, or both of automatically recognized by a computer terminal regardless of whether the first means is in the form of a questionnaire, together with identification information of the computer terminal or the related user. The computer system according to 1. The computer system according to claim 2, further comprising means for performing a predetermined format conversion or a predetermined knowledge inference (deduction) so that the recognition result information of claim 2 matches the format or options of claim 1. In the recognition result information or identification information
The computer system according to any one of claims 1 to 3, which includes information on the installation location (= transmission location) of the computer terminal or is linked in advance. Automatically recommending a preferable or unfavorable action plan (part or all of 5W1H or a series thereof) to the user related to the computer terminal to the user or the associated user based on the result of the automatic statistical processing. The computer system according to any one of claims 1 to 4, wherein the computer system comprises. When the computer automatically recommends to the user or the associated user,
The recognition result information itself used for the automatic statistics (which may or may not be narrowed down to the recognition result information about the user),
The computer system according to claim 5, wherein an inference rule machine-learned from a plurality of recognition result information and / or a set of inference rules input by a human is combined and deduced. The computer system according to claim 6, wherein the content used as the basis for the automatic recommendation deduction is displayed as the basis for the recommendation. When the user or the associated user edits a story such as a novel, a dissertation, a report or a proposal

Recognition result information that the computer is related to the user or the associated user,
Inference rules machine-learned from multiple recognition result information,
A set of inference rules entered by the user or the associated user,
Content in the middle of editing,
Or
A combination of these,
The computer system according to claim 6 or 7, which automatically recommends the contents deduced from. The automatic recommendation of the computer,
The computer system according to claim 7, wherein a script of an adventure game or a role-playing game related to the facts, hypotheses or philosophies used in the deduction can be used as a material for dynamically generating a script. The recognition result information is related to tourism,
The computer system according to any one of claims 1 to 9, which is a mixture of the geography of the tourist object and general philosophy related to history. The order of items that are easier for the user to study by interacting with the computer system (eg, morality, history, philosophy) or the order of items that are more accessible to the user.
Inferring from the statistics of dialogues with other users,
The computer system according to claim 10, wherein the content of the dialogue (choices, prompt expressions for free input, etc.) is automatically controlled according to the estimation result. The inference rule that is deposited is machine learning consisting of a conditional set (for example, a set of existence or nonexistence of a character string) and an effect (in the same example, the existence or nonexistence of a character string not included in the conditional set).
By gradually increasing the elements of the condition set (that is, tightening the data narrowing conditions)
The computer system according to any one of claims 1 to 11, wherein machine learning is performed so that an effect (existence or absence of a character string) in a given context set (set of cases) is uniquely determined. The inference rule that is deposited is machine learning consisting of a conditional set (for example, a set of existence or nonexistence of a character string) and an effect (in the same example, the existence or nonexistence of a character string not included in the conditional set).
By increasing or decreasing the elements of the condition set in stages (that is, by tightening or loosening the data filtering conditions)
The computer system according to any one of claims 1 to 11, wherein machine learning is performed so that the effect (existence or absence of a character string) in a given context set (set of cases) is stochastically determined. The computer system according to claim 13, wherein only the inference rule when the probability of the effect when the elements of the condition set are gradually increased or decreased changes. Constituent requirements generalized as a character string by converting the same substring in multiple character strings that make up the same possible world (context) into the same variable (variable common to each fixed character string that makes up the head and body) The computer system according to any one of claims 12 to 14, wherein a horn clause or the like expressed in a form in which the above is embedded is deposited. The computer program for the computer system according to any one of claims 1 to 15. The processing method for a computer system according to any one of claims 1 to 15.

Images