JP7313646B2 - Method, computer program and apparatus for processing knowledge representation - Google Patents

Description

Application of Patent Act Article 30(2) MITOU Advanced Business Project Proposal

Application of Article 30, Paragraph 2 of the Patent Law 71st Awagasaki Business Salon Supplementary Materials

Patent Act Article 30, Paragraph 2 applied Slides for the 71st Awagasaki Business Salon

The present invention relates to methods, computer programs and devices for processing knowledge representations.

The present inventor has researched and developed a logical programming language for AI (knowledge expression language) that allows character strings containing variables as literals and its processing system (Patent Document 1).
This has the advantage of being easier to describe in natural language than conventional PROLOG.
Like other logic programming languages for AI such as PROLOG,
Most of the preparation of knowledge representation that serves as the basis for reasoning and judgment is
Because it is attributed to the human logical programming cost (knowledge representation cost),
There is a need to make logic programming languages for AI easier to understand and concise, and to reduce the amount of description.
However, including the logical programming language for AI in Patent Document 1 above,
The PROLOG-based logic programming language for AI, which is designed and tuned for backward reasoning "closer",
When forward-looking inference processing (diagnosis, production system, machine control, user questioning based on more hypothetical, environmental sensing, etc.) is required,
It is hard to think of only forward inference using procedural interpretation from single-headed matches of Horn clauses,
When it is necessary to make efficient forward-looking inferences that make full use of various information as opportunities and information,
Significant human effort (programming cost) and amount of description were required.

Patent application 2013-087008

The present invention, in view of such problems,
1) To make it possible to use the knowledge for backward reasoning for forward reasoning (in an easy-to-understand, concise, and small amount of description), reduce the cost of logic programming for AI, and make it easier to spread 2) To enable the above 1) with knowledge expression in a format that is easy to express/match conclusions, effects, conditions, facts, etc. in both text and speech, and to make the above 1) possible, the cost of preparing fixed facts (FACT) by the general public in the logic program, the cost of writing the head and body of the rule by the general public, and the cost of writing the head and body of the rule, and the cost of forward reasoning. The purpose is to reduce the cost of inputting temporary facts (temporary context) by ordinary people, further reduce the total cost of logical programming for AI and program operation, and make it easier to spread.

In order to solve such problems,

Claim 1 of the present invention is
A computer program that processes a rule that assumes that a given proposition is established when some or all of a plurality of conditions are established, and when new information that meets some of the multiple conditions of the rule is obtained,
To provide a computer program for automatically searching for the establishment of other conditions of the rule and automatically confirming the establishment of the predetermined proposition of the rule.
Thereby, the rules for backward inference can be diverted to forward inference with no conversion cost or very little conversion cost.
Examples of "rules" in such a format include PROLOG and the Horn clause of a logic description language that is an extension or modification of PROLOG.
The program may be configured so that, when "the given proposition is established", the meaning of the proposition is rendered (animation, still image, voice, text display, etc.), and further recursive search (this may be called "recursive forward-looking inference"), analysis, database operation, etc., starting from the establishment (= "matching information is newly obtained").
As a result, it is possible to provide an automatic dialogue or the like with more expressive power (meaning, image, etc.) and logical development power.
It is advantageous for the efficiency of inference to configure the rules used for forward inference among the rules for backward inference so that the programmer can change the type of delimiter (“:-” in PROLOG) (for example, use “:=”) to distinguish the rules also used for forward inference from others.

In addition, claim 2 of the present invention is
if said condition and said given proposition are:
Expressed in a form that allows variables to be embedded in natural language,
The matching information is
A computer program is provided in the form of a natural language or in a form in which variables are embedded in the natural language.
This makes it more compatible with natural language speech synthesis and speech recognition, and allows even lower grades of elementary school who are unfamiliar with functions and logical expressions (there are many parentheses, commas, and symbols) to express inference rules (original/general way of thinking) without discomfort.
Examples of formats that allow variables to be embedded in natural language include the examples in the aforementioned patent documents.
By allowing matching information to be given in a form in which variables are embedded in a natural language, even if information about a case corresponding to DONTCARE (a value that is not essential for driving forward-looking inference and can be any value. For example, whether it is 0.01 seconds off from noon or whether the cause of the boom is a weak yen or an increase in population) is not given, unification with a character string that expresses it as a variable can drive forward-looking inference as a result.

In addition, claim 3 of the present invention is
A computer program is provided that searches for the establishment of the predetermined proposition in a more restricted form by spreading the variable binding on the rule side obtained by unifying the matching information and the condition to match the rule.
For example, the binding of $X that unifies the new matching information ``Yuta has comprehensibility'' with the condition ``$X has comprehension'' is ``$X=Yuta'', and in the case of the rule ``$X is suitable for a leader: - $X has comprehension'', by spreading this constraint, it is possible to automatically confirm the establishment of a proposition ``Yuta is suitable for a leader'', which is more specific than ``$X is suitable for a leader''.
That is, it is possible to automatically confirm the establishment of a more specific proposition (more information and less ambiguity) reflecting the more specific content of the new information (for example, who the subject is) (and more specific rendering, command execution, etc.).
In addition, it leads to the limitation of the search space, which leads to the reduction of the amount of calculation and the speeding up of the response.

In addition, claim 4 of the present invention is
Variable binding on the matching information side obtained by unification matching the matching information and the condition,
A computer program is provided that returns the matching information to an information source that provided it.
Here, the "information source" may be, for example, a user or another computer.
For example, for the condition "$Y is grain" in the rule "$X needs $Y as staple food; $Y is grain;", unifying the "matching information""rice is $A" for the condition "$Y is grain", the rule side variable binding "$Y = rice", the matching information side variable binding "$A = grain", and the proposition "Japanese people need rice" are obtained. I can.
This makes it possible for the information source to obtain some sort of clue as to the grounds for establishing a new proposition based on new information, without referring to the whole inference process.

In addition, claim 5 of the present invention is
subject to said unification;
For a pair consisting of one matching information and one condition,
If multiple variable bindings are possible,
A computer program characterized by independently processing each of them is provided.
Here, the ``possibility of multiple variable bindings'' means, for example, that there are multiple bindings ({($A=Taro, $B=Jiro and Saburo), ($A=Taro and Jiro, $B=Saburo)}) that unify the condition that ``$A and $B are siblings'' and the information that ``Taro, Jiro, and Saburo are siblings'', for example.
As a result, it is possible to search for the sufficiency of a condition without losing multiple possibilities (backward search (inference)), and exhaustive forward search (enumeration of propositions that can be newly confirmed (proven)) based on the satisfaction set.

In addition, claim 6 of the present invention is
If we confirm the existence of the proposition,
With the proposition as newly obtained information,
By searching whether a part or all of each condition of one or more rules stored by a computer is matched,
Characterized by recursively repeating checking whether each proposition of each rule holds,
Provide a computer program.
Here, "one or a plurality of stored rules" may be the same rule as the rule used to confirm the establishment of the proposition or a different rule.
In addition, when the original proposition on the rule contains variables, it is preferable that the variables contained in the proposition are also bound (typically, fixed character strings representing natural language) as a result of binding variables in the process of unification with the conditions searched for matching before the establishment of the proposition. However, even if the variable is not necessarily bound and remains, it does not restrict propositions that assert tautology (for example, "$X is equal to $X") or propositions whose ambiguity (remaining variables) do not affect the essence of the meaning in the same way as fixed character strings (such as triggering recursive forward inference as newly obtained information).
This will allow us to see the meaning and impact of the new information more broadly and deeply.
Also, even if the way of giving new information is ambiguous or slightly off, it will lead to confirmation of the contents that may have an impact.

In addition, claim 7 of the present invention is
To provide a computer program in which a rule allows listing of a plurality of propositions whose establishment can be confirmed in response to matching of the same or similar conditions.
“Allowable in rules” means, for example, allowing rules in a format that allows multiple propositions and effects to be listed in the head portion, instead of the traditional Horn clause, which limits the “predetermined proposition” to one.
For example, "$X lives in the Orient & $X lives in the Northern Hemisphere:-$X lives in the Japanese archipelago;" lists two propositions "$X lives in the Orient" and "$X lives in the Northern Hemisphere" that can be deduced from "$X lives in the Japanese archipelago" separated by "&".
It is also possible to reverse the right and left sides of the conditions and propositions (effects) and express special delimiters, such as "$X lives in the Japanese archipelago: <$X lives in the East; $X lives in the Northern Hemisphere;"
As a result, it is possible to effectively avoid duplication of calculations regarding whether the same or similar conditions are satisfied, and it is possible to increase the calculation efficiency regarding whether conditions are satisfied, which is an element of the efficiency of forward inference.
Note that "similar conditions" are, for example, when the part corresponding to the right side (body) of the Horn clause or part of it is a list of OR conditions for the purpose of absorbing notational variations ("$X dissolves $Y | $X dissolves $Y"), a list of equivalent OR conditions ("$X likes ramen | $X likes udon | $X likes soba | $X likes spaghetti"), or falls under the category of a specific regular expression. Various cases can be assumed, such as when

Moreover, claim 8 of the present invention is
Taking the common part of the similar conditions as the root or a branch close to the root,
Taking the different parts of similar conditions as branches farther from the root,
8. The computer program according to claim 7, wherein the rules are expressed in a branched tree with propositions that can be regarded as established by checking the conditions from the root to the corresponding branches as leaves.
An example of such a representation would be the following, representing an equivalent tree by corresponding nesting of parentheses:
A: <{B {C: E1} {D: E2}} {E {F {G: E3} {H: E4}}} {I: E5} {J {K: E6} {L: E7} {E8}};
here,
Unlike the usual horn theory,
A is the root condition,
": <" is a syntax delimiter and a delimiter that indicates the type,
The path (necessary condition set) to reach the establishment of proposition (effect) E1 is A, B, C,
The path (necessary condition set) to reach the establishment of proposition (effect) E2 is A, B, D,
The path (necessary condition set) to reach the establishment of proposition (effect) E3 is A, E, F, G,
The path (necessary condition set) to reach the establishment of proposition (effect) E4 is A, E, F, H,
The path (necessary condition set) to reach the establishment of proposition (effect) E5 is A and I,
The path (necessary condition set) to reach the establishment of proposition (effect) E6 is A, J, K,
A, J, and L are the paths (necessary condition set) to reach the establishment of proposition (effect) E7.
A is the path (necessary condition set) up to the establishment of proposition (effect) E8.
Thereby, the calculation cost used for confirming the sufficiency of the common part of the condition set is not wasted, and the calculation of the sufficiency of the difference part becomes possible, and the sufficiency confirmation of the condition set having the common part is made efficient and speedy.
The "part of the condition" to be unified with the "matching information" can be only A, which is the root, or any of the intermediate branch conditions. Also, by setting this A to "$X" (asterisk) or the like with no pattern restrictions, it is possible to try to match everything with each branch.
It should be noted that the grammar and interpretation of ordinary Horn clauses and the above-mentioned patent documents are the same in that they support multi-length matching, that backtracking or equivalent exhaustive processing can be performed, and that procedural interpretation is possible.

In addition, claim 9 of the present invention is
8. The computer program according to claim 7, wherein a list of a plurality of propositions whose validity can be confirmed corresponding to matching of the same or similar conditions allowed in the rules follows a preferred procedure order.
As a result, procedural interpretation makes it possible to reproduce the procedural order of occurrence among multiple propositions whose validity can be confirmed.
In addition, by programming such rules so that only one matching rule is selected, each step of the procedure can be recursively developed in a tree form. In such cases, steps consisting of complex nests of various granularities can be defined in a tree form, and at the time of execution, it is possible to teach a procedure corresponding to a procedure specification in a complicated order and to control the situation.

In addition, claim 10 of the present invention is
10. The computer program according to claim 7 or 9, characterized by automatically requesting externally that each proposition should be satisfied in the order according to a list of a plurality of propositions whose establishment can be confirmed corresponding to matching of the same or similar conditions allowed in the rule, and waiting until the satisfaction is confirmed.
As a result, it becomes possible to entrust the computer program to some extent with supervising work of sequentially requesting and confirming the satisfaction of the requested content.
Such a rule can recursively expand the sequence of external requirements and satisfaction confirmations in a tree by narrowing down the matching rule to one. In such a case, tasks consisting of complex nesting of various granularities can be defined in a tree form, and during execution, it is possible to wait for external requests and confirmations of satisfaction corresponding to requirements satisfaction specifications in a complicated order.
"External" may be a person, an animal, a robot, a control device, or an environmental element (water level, etc.), and confirmation of sufficiency may be made by cues from them (signal, gesture, physiology, physical phenomenon, natural phenomenon, etc.).

In addition, claim 11 of the present invention is
When the number of satisfied conditions is a predetermined number of 2 or more (the rule may be specified as ": 2" etc.),
Provided is a computer program that satisfies the predetermined proposition.
If the rule delimiter is ":2", it may be 2 or more, if it is ":3", it may be 3 or more, and so on.
This makes it possible to concisely describe a logical program for diagnostic forward reasoning, such as diagnosing "probably cold" when two or more conditions such as "runny nose", "sore throat", "sore head", and "coughing/sneezing" are met.

In addition, claim 12 of the present invention is
If the number of conditions satisfied is exactly 1,
Provided is a computer program characterized by automatically confirming the establishment of the predetermined proposition.
This corresponds to the definition of exclusive OR, and the rule delimiter may be (:1) or the like.
As a result, it is possible to code forward-looking reasoning at low cost, which can simply express conditions requiring exclusive OR, which is a premise of forward-looking reasoning.

In addition, claim 13 of the present invention is
A computer program characterized by automatically confirming the establishment of a predetermined proposition when a value calculated from a satisfied condition and each weight associated with each satisfied condition is equal to or more than a predetermined threshold value, equal to or less than a threshold value, or exceeds or is less than a threshold value.
This enables not only binary (provable/unprovable or true/false) and ternary (provable/unknown/provable false) but also more quantitative and precise reasoning that takes into account the degree of certainty/probability along the way.
Here, the threshold, the hierarchical relationship with the threshold, and the weight related to each condition (for example, the confidence or probability of the condition itself, or the weighting factor to be multiplied by them in the rule) can be added to the rule or fact (or temporary context) in a predetermined format.
For example, "cold/80; influenza/10: % 50 < runny nose/20; sore throat/30; headache/30; cough/sneezing/40;" (50 x 20 + 30 x 30 + 30 x 90 + 40 x 60) / 100 = 10 + 9 + 27 + 24 = 72 satisfies "greater than 50" meaning "% 50 <", so the establishment of the proposition is automatically confirmed with a certainty of 72, and "cold" degree: 80 × 72/100 → "cold" degree: 58 (rounded off) and "flu" degree: 10 × 72/100 → "flu" You will get a new fact (temporal context) of degree: 7 (rounded off). It should be noted that here, although it is a confidence format, it may be more appropriate to apply X + Y - XY / 100 for the confidence of the same fact obtained from another event or rule (in a related temporal context set for the same patient, etc.), as in the probability format.

In addition, claim 14 of the present invention is
14. The computer program according to claim 13, characterized in that, when automatically confirming the establishment of the predetermined proposition, for one or more related propositions, a value obtained from the weight of each associated proposition, the weight associated with each condition, or both is stored in association with each proposition.
This makes it possible to easily define the certainty of one or more automatic confirmation propositions corresponding to different viewpoints (condition sets) for each rule.
Influenza/10: %50< runny nose/20; sore throat/30; headache/30; coughing and sneezing/40;
In addition, in order to prevent the same rule from firing recursively to form an infinite loop based on the same grounds and unduly increasing or diverging the certainty, it is also possible to store a history of which rule has fired based on which fact, and restrict the firing of the same rule based on the same fact.

In addition, claim 15 of the present invention is
When new information that meets some of the multiple conditions of the rule is obtained,
To provide a computer program characterized by confirming with a user the truth of a condition whose truth cannot be proved in the same rule, or confirming with the user a set of variable bindings to be true or false.
As a result, especially in a diagnostic system, etc., from a state in which there is little temporary context about the patient (that is, a state in which it is difficult to prove the truth of the condition),
It is possible to gradually increase the number of questions to be confirmed by the user for each rule,
It is possible to avoid communication costs and inference costs due to unnecessary question-and-answer (user confirmation).
For example, if such a rule were expressed in a PROLOG-like syntax separated by ":2?"
This rule is activated according to the input of information (temporary context fact) that matches any of the four conditions,
Until you can confirm the establishment of two or more conditions,
"Automatic proof -> user confirmation if not possible (addition of context may also occur)" will be performed in order from the front.
"Cold: 2? runny nose; sore throat; headache; coughing and sneezing;"

In addition, claim 16 of the present invention is
When new information that meets some of the multiple conditions of the rule is obtained,
A computer program characterized by confirming with the user the truth or falsehood of corresponding conditions or propositions (both of which cannot be automatically proved true or false) in the same rule, or confirming with the user the set of variable bindings that are true or false.
This allows, particularly in diagnostic systems, from low temporal context about the patient to
It is possible to gradually increase the number of questions to be confirmed by the user for each rule,
It is possible to avoid communication costs and inference costs due to unnecessary question-and-answer (user confirmation).
For example, in the example of "A:<{B{C:E1}{D:E2}}{E{F{G:E3}{H:E4}}}{I:E5}{J{K:E6}{L:E7}{E8}};"
"A:<{?B{?C:E1}{?D:E2}}{E{?F{G:E3}{H:E4}}}{?I:E5}{J{K:E6}{L:E7}{?E8}};"
It should be noted that it is also possible to arrange and implement program language specifications for confirming the user for all conditions that cannot be automatically proved by delimiting rules that include "?", such as ":?<".

In addition, claim 17 of the present invention is
An apparatus using a computer program according to any one of claims 1 to 16 is provided.

In addition, claim 18 of the present invention is
A method using a computer program according to any one of claims 1 to 16 or according to Fig. 2 is provided.

FIG. 10 is a diagram illustrating an exemplary relationship between a logic program (fixed facts and rules) of the present invention, a temporary context already added by the user, data newly added by the user (temporary context), and a proposition (variables are bound) that is output (speech, text display, etc.) as a conclusion of forward inference and that can be a seed of recursive forward inference. 2 is a diagram for explaining the most basic processing flowchart (steps as a method) of a computer program for implementing FIG. 1 and the like; FIG. It is an explanatory view showing a general configuration as a computer to implement.

FIG. 1 is a diagram illustrating an example relationship between the logic program (fixed facts and rules) of the present invention, the temporary context already added by the user, the data newly added by the user (temporary context), and the propositions (variables are bound) that are output (speech, text display, etc.) as the conclusion of forward inference and that can be the seed of recursive forward inference. The forward reasoning step of FIG. 2 is activated by matching the new information "A child has a sore throat" entered by the user with the underlined "$X has a sore throat".

FIG. 2 is a diagram illustrating a processing flowchart (steps as a method) of the most basic forward inference of a computer program for implementing FIG. 1 and the like. In fact, in step 3-2, it is advantageous in various aspects to perform special unification such as multi-length matching and backward inference including backtracking, which are described in detail in Patent Document 1 (published application).

FIG. 3 is an explanatory diagram showing a general configuration as a computer that implements. Since there are no unique features of the present invention here, it will not be described in detail.

FIG. 1 shows one embodiment of the present invention. This knowledge expression is used to implement the invention of Patent Document 1, and the delimiter between conditions is ";". Also, the division between the head and the body is ":3" indicating that three or more conditions must be met. In step S1, the information "Child A has a sore throat" is entered, in step S2 matching conditions are retrieved from the condition parts of all rules, and in step S3-1, matching constraints are propagated to all $X of the rules as "$X = child A". In step S3-2, since "Japan" already exists in the logic program as a fixed fact in this case, the computer judges that this rule can be fired when it is confirmed by searching (within the scope of the logic program and the temporary context) that the remaining two or more conditions out of the three in ":3" are met. will become the seed of

4 General computer used in the present invention (user-dedicated personal computer, shared file server, web server)
4a-4k general components of a general computer 4 (user-dedicated personal computer, etc.) 11 logic program (fixed facts and rules)
12 data (temporary context already added to array memory adjacent to the logic program)
13 New facts (data) (temporary context added by the user)
14 Rule meaning that if three or more conditions on the right side are satisfied, the establishment of the proposition on the left side is automatically confirmed (automatically added & notified) 15 New conclusion (automatically added and notified to the user, temporary context input as a seed for recursive forward reasoning)
16 Paths Input as Seeds for Recursive Forward Inference (via Queues)

Step S1: Step of accepting input of new data A and storing it as a temporary context S2: Step of searching for a condition that matches A for each rule in the logic program: Step S3: Step of executing each rule including matching conditions (because it may be in sequence): Step S4: If it is not determined whether data exists in the queue, return to S1: Step S5: Return to S2 using data A′ existing in the queue as new data (fact) A;

Claims (9)

A computer program that processes a rule that assumes that a given proposition is established when all of a plurality of conditions are established.
A computer program that automatically searches for the establishment of other conditions of the rule and automatically confirms the establishment of the predetermined proposition of the rule,

if said condition and said given proposition are:
Expressed in a form that allows variables to be embedded in natural language,
The matching information is
be given in the form of natural language or in the form of embedding variables in natural language;

obtained by unification of matching the matching information and the condition;
Spreading the variable binding on the rule side to the rule,
searching for the holding of the given proposition in a more restricted form;

If we confirm the existence of the proposition,
With the proposition as newly obtained information,
By searching whether all the conditions of one or more rules stored in the computer are met,
recursively repeating checking whether each proposition of each rule holds;
A computer program characterized by 2. The computer program according to claim 1, wherein the rule allows a list of a plurality of propositions whose establishment can be confirmed in response to matching of the same or similar conditions. Taking the common part of the similar conditions as the root or a branch close to the root,
Taking the different parts of similar conditions as branches farther from the root,
3. The computer program according to claim 2 , wherein the rules are expressed in a tree-like manner with branches, with propositions that can be regarded as established by checking the conditions from the root to the corresponding branches as leaves. When the number of satisfied conditions is a predetermined number of 2 or more (the rule may be specified as ": 2" etc.),
2. The computer program according to claim 1, wherein said predetermined proposition is established. If the number of conditions satisfied is exactly 1,
2. The computer program according to claim 1, wherein the establishment of said predetermined proposition is automatically confirmed. If the value calculated from the satisfied condition and each weight associated with each satisfied condition is greater than or equal to a predetermined threshold value, less than or equal to a threshold value, or greater than or less than a threshold value,
2. The computer program according to claim 1, wherein the establishment of said predetermined proposition is automatically confirmed. When automatically confirming the establishment of the predetermined proposition,
For one or more related propositions,
7. The computer program according to claim 6, wherein a value obtained from the weight of each associated proposition and/or the weight associated with each condition is stored in association with each proposition. When new information that meets some of the multiple conditions of the rule is obtained,
2. The computer program according to claim 1, further comprising confirming with the user whether the conditions in the same rule cannot be proved to be true or false, or confirming with the user a set of variable bindings to be true or false. When new information that meets some of the multiple conditions of the rule is obtained,
2. The computer program according to claim 1, which confirms with the user whether the corresponding conditions or propositions (both of which cannot be automatically proved true or false) in the same rule are true or false, or confirms with the user a set of variable bindings that are true or false.

Images