JP2020190857A - Computer program, method and device for inference using weight and general-purpose ai - Google Patents

Abstract

To achieve a general-purpose AI which can robustly utilize information by a plurality of weights and is active over knowledge of a plurality of fields and official/private information at a small cost of knowledge representation.SOLUTION: A computer program is provided, which can infer whether an effect is established in accordance with the magnitude of the sum of weights of established conditions about a Horn clause being a Horn clause of a logic programming language Prolog or a clause of a logic programming language for enabling equivalent declaratory interpretation to be able to add a weight to each condition.SELECTED DRAWING: Figure 1

Description

The present invention relates to computer programs, methods and devices for weighted inference and general purpose AI.
It also relates, in particular, to such computer programs that express knowledge in free text (including the entire natural language of each country).

Conventionally, neural networks (including deep learning), fuzzy sets, fuzzy PROLOG, etc.
Knowledge representation has been attempted to make more robust use of information by using multiple weights (strength of association, etc.).
There has been no integrated approach to extremely flexible and private knowledge and values expressed in free text, including the whole of natural language.
For this reason,
The cost of knowledge representation is high
It was a major obstacle to the realization of general-purpose AI that is active across knowledge in multiple fields and public and private information.

Japanese Patent Application No. 2019-078141 Japanese Patent Application No. 2013-087008 Japanese Patent Application No. 2017-192296 Japanese Patent Application No. 2017-092076 Japanese Patent Application No. 2017-092057 Japanese Patent Application No. 2017-092047 Japanese Patent Application No. 2016-165180 Japanese Patent Application No. 2016-09324 Japanese Patent Application 2015-234356 Japanese Patent Application 2015-219007

Therefore, the problem to be solved by the present invention is
Information can be used robustly by multiple weights,
The cost of expressing knowledge is small,
It is to realize a general-purpose AI that plays an active role across knowledge in multiple fields and public and private information.

In order to solve such a problem, the present invention
As claim 1, the horn clause of the logic programming language Prolog,
Or
A section of a logic programming language that allows for equivalent declarative interpretation
About the horn clause that can be weighted for each condition
Provided is a computer program capable of inferring whether or not an effect is satisfied according to the magnitude of the sum of the weights of the satisfied conditions.
: For example, "a section of a logic programming language capable of equivalent declarative interpretation" is a condition in which a sentence in which variables are embedded in a free sentence (including the entire natural language of each country) constitutes the body part of the horn clause. One and / or a section of a logic programming language having an effect constituting a head portion, etc. (see Patent Document 2).
: Note that the weight attached to each condition may be negative, and in this case, the establishment of the constituent requirements may be hindered (when a large threshold value is required) or promoted (small) depending on the negative weight. When a threshold is required), it becomes a condition.
> This allows the editor of a logical program to acquire a logical expressive power closer to that of a neural network.

The present invention also provides the computer program according to claim 1, wherein it is possible to infer whether or not the effect is satisfied according to the condition regarding the threshold value of the sum of the weights attached to the horn clause.
: This is logical programming, for example, assuming that the delimiter (usually ":-") that separates the effect (head part) and the condition part (body part) requires 80 or more, such as ":%> = 80". It is possible for a computer program, which is an interpreter, to interpret it by attaching it in a language.
: For example, logical programming (lines 2 to 12 correspond to rules and facts), query (? $ X likes sports), and answer (: $ X = Ichiro: $ X = Yuta) as follows. Although it is supposed to be obtained, the interpretation computer program can also be configured so that it can be logically programmed to require a sum smaller than -50, such as "% <-50".
===
1: 1:
2: $ X likes sports:% 70 40 / $ X likes tennis; 40 / $ X likes baseball; 40 / $ X likes swimming; 20 / $ X likes dance; 10 / $ X likes e-sports;
3: Ichiro likes baseball
4: Ichiro likes swimming
5: Ichiro likes tennis
6: Koji likes baseball
7: Koji likes e-sports
8: Yuta likes swimming
9: Yuta likes dance
10: Yuta likes e-sports
11: Hanako likes tennis
12: Hanako likes dance
===
------- 2019/05/11 12:26:53 ------- S

->? $ X likes sports

: $ X = Ichiro

: $ X = Yuta

Search time = 1357msec
===

> This allows the editor of a logical program to acquire a logical expressive power closer to that of a neural network at a lower cost.

The present invention also claims that claim 3.
The above section
So that it can be referred to as a function to find the sum of weights, or
The computer program according to claim 1, which can be called as being procedurally interpretable while obtaining the sum of weights.
: C language function call (return value is an integer) is the most similar concept, and the result of the call is compared syntax outside the standard Prolog grammar (example: "IF (happiness (Taro, marriage, $ Y))> Happiness (Taro, lover, $ Y)) {LINK "Taro should marry $ Y!";} ") Or substitution statement (example:" $ X = happiness (Taro, marriage, $ Y) ") It can be used for.
> As a result, one writing method can be used for multiple purposes (logical gates, functions, etc.), and the cost of writing knowledge can be saved.

The present invention also claims that claim 4.
For the reference or call of claim 3.
The computer program according to claim 3, which returns or displays a bound value set corresponding to an unbound variable set in the head and a return value as the function corresponding to each bound value set.
: For example, it becomes as follows.

===
[Logical program]
$ X likes $ Y:% 70 50 / $ X wants to be near $ Y; 70 / $ X and $ Y are in love; 30 / $ X is a mammal;
Taro wants to be near Hanako Jiro wants to be near Hanako Taro is a mammal Jiro is a robot Taro and Hanako are in love

[Call]
?? $ X likes $ Y

[display]
{○ = 150/70 ($ X = Taro, $ Y = Hanako) ,? = 50/70 ($ X = Jiro, $ Y = Hanako)}
===

: Above, "○" indicates the establishment, 150 indicates the return value, 70 indicates the threshold value, and "?" Indicates the failure (insufficient evidence).
> As a result, it is possible to convey to the context of the caller the degree and reason why it is established or not for each solution, or to show the user.

The present invention also claims 5
The computer program according to any one of claims 1 to 4, wherein the section expresses a theory regarding the cause of liking, happiness, health or success.
: "Theory about the cause" is, for example, "Success: -I will not stop until I succeed;""$ X is happy:-$ X has a partner who understands; $ X has a job to devote to ; ”And so on.
> This makes many people interested in the inference results and provides real information to the questions from the system.

The present invention also claims that claim 6.
The weight or threshold value in the logical program is represented by a variable.
When the expressed variable is numerically bound under the condition,
The computer program according to any one of claims 1 to 5, which adopts the numerical value as a weight or a threshold value, respectively.
: As an example of "the weight is represented by a variable (for example, $ <capital letter>)", for example,
"$ X100g contains $ Yg of sugar and is sweet:% 10 $ Y / $ X100g contains $ Yg of sugar; ($ Z / 5) / $ X100g contains $ Zg of mandarin orange juice;" and the like. If $ X100g contains 10g or more of sugar, it is sweet by itself, and even if it contains only 5g of sugar, if it contains 25g or more of "tangerine juice", the weight becomes 5+ (25/5) = 10. The head will be able to concisely infer that "$ X 100g is sweet with 5g of sugar".
: As an example of "the threshold value is represented by a variable (for example, $ <capital letter>)", for example,
"$ X100g contains $ Yg of sugar:% $ Y $ M / $ X100g contains $ Mg of sugar; ($ N / 5) / $ X100g contains $ Ng of mandarin orange juice;" and the like. When the sugar content (M + N / 5) contained in sugar and mandarin orange juice is Y or more, the effect of the head can be inferred concisely.
> These will further improve the complete expressiveness using the weights and thresholds of inference rules.

The present invention also claims that claim 7.
The sum of the weights in a given rule, or the sum of the weights of a given weighted value set,
Provide a computer program that conducts conversations (including questions to users, responses to questions, etc.) that can update the context about users so that they become more or less.
: "Sum of weights in a given rule" means, for example, "Do you have a fever?" So that the sum of each given weight of symptoms (fever, cough, runny nose, ...) that is evidence of a cold is large. By asking the user "Do you have a cough?""Do you have a runny nose?" ... and add the sentence "You have a fever" to the context, "You have a cold." It is possible to have a conversation that suggests a solution (for example, "maybe I should go to internal medicine") in connection with the diagnosis (reasoning result).
: On the other hand, "sum of weights of predetermined weighted value sets" is based on the following information stored in the setting file registered in advance by the user.
So that the sum of the weights (70, 70, 50) is greater than (eg)
Look for a set of inference rules that can prove the greater weight of "the existence of a good understanding person" and "the existence of work that is passionate about" (ie,
From the already main sources of knowledge (main programs (facts and rules in the case of Prolog),
Main program candidate URL list preset as a candidate knowledge source,
Search from external site URLs that can search them),
A series of questions that can prove the conditional part of each inference rule directly, indirectly, or recursively (retroactively in combination with other inference rules) is implemented for the user (that is, in a user-specific context). (Adding new facts).

===
<% *
70 / Existence of a good understanding person 70 / Existence of work to work with passion 50 / Existence of a good working environment>% *
===

> This makes it easier to perform automatic dialogue with the user according to the "predetermined rule" and the "predetermined weighted value set".

The present invention also claims that claim 8.
The computer program according to claim 7, which asks a user a question so as to increase or decrease the sum as soon as possible or with a small number of questions.
: This can be obtained by simulating the truth of the question result corresponding to the condition in advance with a search tree or the like.
: If the simulation takes a long time, it is possible to trim various prunes to reduce the calculation time to ask a question within, for example, 6 seconds.
> As a result, it is possible to provide the user with an efficient question sequence with low cost.

The present invention also claims as 9.
A computer program that is used by switching from the previous one at the time of execution according to the contents of the memory for each user (for individual definition, for storing the context at the time of execution, etc.), or
Provided is a computer program characterized by determining a computer program to be used in addition to the conventional ones at the time of execution.

The present invention also claims 10.
Search at runtime according to the contents of the memory for each user (for individual definition, memory of runtime context, etc.), switch to the previous one, or add to the previous one Provided is a computer program to be executed or a computer program characterized in that a search range of the computer program is defined.
: Depending on the content of "Memory for each user", switching and addition can be used properly, and the memory for each user can include the last adopted computer program, so it is a candidate for adoption. The programmer of a computer program can specify whether it should be switched or added to the computer program. This also applies to the search range in each case of switching / adding, and the programmer can directly or indirectly specify the candidate computer program to which each computer program is switched, as in the case of HTML. The same applies to the range to be additionally executed.
> This makes it possible to specify a preferable setting and deployment range for each user and each programmer of the computer program.

The present invention also claims 11.
Relatively or absolutely weighted by relative or absolute definition of value weights in memory for each user or in the running computer program, to fill the value that should be prioritized faster according to the priority of the value. Provide a computer program for chatbots that asks questions and guides people.
: The following is a relative definition example, where the ";" in the same line indicates equivalence and the previous line has a high relative value.

===
<% *
Existence of a good understanding person; Existence of work that works with passion Existence of a good work environment>% *
===

: On the other hand, the following is an absolutely defined example, showing "weight / content", and those with the same weight are regarded as equivalent.

===
<% *
70 / Existence of a good understanding person 70 / Existence of work to work with passion 50 / Existence of a good working environment>% *
===

Use of the "condition part" (body part of the horn clause) of the "rule of inference" (horn clause, etc.) with each value ("existence of a good understanding person", etc.) as the "effect" (head of the horn clause, etc.) Other inference rules that develop for conditions that do not exist as facts (eg, "today is sunny") in the human context (eg, "today is sunny:-no clouds now; today's weather forecast is sunny") If there is no such thing as "the weather forecast is good;"), there is an inference rule that develops by directly asking "is it sunny today?", And most of the conditions (for example, "the weather forecast is sunny today"). If "weather forecast is good") is registered as a fact in the user context, for example, "there is no horn" using the remaining one condition, "there is no horn". Addition of a computer program that allows chatbots to ask users the question "?", Indirectly prove the value of "Today is sunny", and have other suggestive conversations such as "Why don't you go fishing?" And can be connected to switching.
> This allows for more important value
For example, suggesting conditions that the user lacks earlier (what the user can meet himself (for example, brushing his teeth before going out in the context of going on a date, etc.)) and then satisfying himself. ("1: Polished, 2: Still"), etc. to input and add as a fact to the user context,
Check the relevant facts that are already sufficient,
You can configure a chatbot or similar user interface that allows you to have tactical recommended conversations such as "If you do □□, you can do XX!".
: In addition, for automatic confirmation and automatic questioning whether the conditions of each inference rule related to whether the value is achieved are met, the condition part of the inference rule equivalent to PROLOG is recursively expanded and the question is asked. Only the necessary parts can be specified and performed.
: Also, by controlling one of the conditions of the inference rule to actually propose to the user in the chatbot conversation in the form of "proposed XX", the priority value has already been realized. It can also have the effect of a flag that means that the conversation is controlled to achieve a "next" priority value rather than that value.
: In this case, for example: "Proposal implementation 1:-Runny nose; swelling; fever; headache;" inference rule that "proposed to go to internal medicine" and, for example: "Implement various proposals" : -40 / Proposal Implementation 1; 30 / Proposal Implementation 2; ”by the inference rule or equivalent definition of value
Proposal implementation 1 to confirm whether it is a cold and recommend going to the internal medicine department, etc.
Proposal implementation 2 that confirms whether it is sunny today and recommends going fishing today is in one of the value definitions (user definition file and user context memory (including a copy of an external computer program)). It is possible to coexist in one session with the user efficiently by coexisting in one session with the user.

===
<% *
40 / Proposal implementation 1
30 / Proposal implementation 2
＞% *
===

The present invention also provides a method using such a computer program.

The present invention also provides a device comprising such a computer program.

It is a figure which shows the example of the interaction of the knowledge source (logical program) by this invention. It is a figure which shows the simple operation example of the system by this invention. It is a figure which shows the architecture of the general-purpose AI system aimed at by this invention.

In the most advantageous form, the value weight can be defined on the logical program side that can be the main program (PG), and the user's definition file and context memory (which becomes a file when saved) can also be "weighted". ..
Questions and guidance from the system that increase the total score of user definition + context will be implemented.
As shown in the above-mentioned patent documents, the logical program has the ability to dynamically cite (LOR) external links in the program and dynamically rewrite the user context.
The main program (logical program) that is temporarily (dynamically) loaded for the execution of general-purpose AI has parameters such as startup frequency and importance in addition to values, and the frequency of questions to increase the value and It is now possible to adjust for conflicts between values.
There are the following types of specifications for the search range of the main program that is dynamically (switched / added) loaded.
1) URL list of main PG candidates 2) URL of an external server that returns the URL list of main PG candidates when the user definition is passed
3) The URL of the search server that returns the optimum main PG itself, PG set, or candidate URL list when the user definition and user context itself are passed.
When taking the form of a general-purpose AI chatbot, it is configured so that a user-specific program (for example, PYTHON, C #, etc.) can be preferentially implemented. Also, if the user is typing on the keyboard, consider it as an addition of context and refrain from thinking and outputting.
The main program asks questions according to the policy of the main program itself, changes the internal state of the AI chatbot, and accumulates the user context.
The main program proceeds to cause identification> problem solving proposal according to the order of the value to be solved.
Many of the main programs have the following value rankings that they want to solve.
<% *
$ {User} will be the happiest
The most important person of $ {user] lives happily for 100 years
$ {User} contributes to human happiness>% *
Many of the main programs have inference rules and assumptions that lead to the value you want to solve.

$ {User} will be the happiest:-$ X will keep working with the best partner for the rest of his life; $ X will live for 100 years;
$ X lives $ Y years:-$ X believes it will live $ Y years Users believe it will live 100 years

The main program asks questions and suggestions to users as chatbots, etc. so as to satisfy the highest possible "value to be solved".
Question formats are often choices (numbered), but you can also answer with choice content (substrings not found in other choices).
If the definition of value is "<&*", try to find multiple "values you want to solve" instead of one.
The user context is cleared in one day based on the time stamp added in the comment.
A typical chatbot for companies functions as a CFO, in-house educator, and general education system, but the general public is also used for chat.

FIG. 1 is a logical program different from the computer programs of claims 1 to 11 based on the contents of the user definition and the user context according to the computer program (not shown) of the embodiment of the present invention. ) Is a diagram for explaining the mechanism of general-purpose AI using weights by searching and switching and rewriting the definition / context.
FIG. 2 is a diagram occupying an example of simple logic using weights (however, based on English, which is a natural language).
FIG. 3 is a diagram showing a general-purpose AI architecture aimed at by the present invention.

Logical program that can be the PG main program (may include parts that should be interpreted procedurally)

Claims (13)

The horn clause of the logic programming language Prolog,
Or
A section of a logic programming language that allows for equivalent declarative interpretation
About the horn clause that can be weighted for each condition
A computer program that makes it possible to infer whether or not an effect is satisfied according to the magnitude of the sum of the weights of the satisfied conditions. The computer program according to claim 1, wherein it is possible to infer whether or not the effect is satisfied according to the condition relating to the threshold value of the sum of the weights attached to the section. The above section
So that it can be referred to as a function to find the sum of weights, or
The computer program according to claim 1, which can be called as being procedurally interpretable while obtaining the sum of weights. For the reference or call of claim 3.
The computer program according to claim 3, wherein a bound value set corresponding to an unbound variable set in the head and a return value as the function corresponding to each bound value set are returned or displayed. The computer program according to any one of claims 1 to 4, wherein the section expresses a theory regarding the cause of liking, happiness, health or success. The weight or threshold value in the logical program is represented by a variable.
When the expressed variable is numerically bound under the condition,
The computer program according to any one of claims 1 to 5, which adopts the numerical value as a weight or a threshold value, respectively. The sum of the weights in a given rule, or the sum of the weights of a given weighted value set,
A computer program that conducts conversations (including questions to users, responses to questions, etc.) that can update the context of users so that they are more or less. The computer program according to claim 7, wherein the user is asked a question as soon as possible or with a small number of questions so that the sum is large or small. A computer program that is used by switching from the previous one at the time of execution according to the contents of the memory for each user (for individual definition, for storing the context at the time of execution, etc.), or
A computer program characterized by determining which computer program to use in addition to the previous ones at run time. Search at runtime according to the contents of the memory for each user (for individual definition, memory of runtime context, etc.), switch to the previous one, or add to the previous one A computer program to be executed or a computer program characterized in that a search range of the computer program is defined. Relatively or absolutely weighted by relative or absolute definition of value weights in memory for each user or in the running computer program, to fill the value that should be prioritized faster according to the priority of the value. A computer program for chatbots that asks questions and guides you.
The method using the computer program according to any one of claims 1 to 11.
A device comprising the computer program according to any one of claims 1 to 11.

Images