JP7549760B2 - Representation of deep case meanings using labeled Petri nets - Google Patents

Description

In general, an inference system is composed of a knowledge base that is the premise of inference and an inference engine that extracts answers from the knowledge base. The present invention relates to an inference system based on natural language, that is, an inference system in which both the knowledge base and the query language are composed of natural language.

Since its birth in the 1950s, research into artificial intelligence (AI) has focused on the "symbol processing approach," which attempts to handle human-controlled words and symbols on a computer. This approach was most actively studied in the 1980s, when AI experienced its second boom. Expert systems (Non-Patent Document 1), which attracted the attention of researchers at the time, expressed expert knowledge as "if-then" rules, making it possible to perform logical inference by combining a large number of rules. Expert systems use formal knowledge expressed in words to deal with problems that require deep inference and are difficult to describe using existing algorithms, and can be said to be the first AI system to combine words and inference.

However, this expert system has several implementation problems. The first problem is the problem of describing the rules. The expression format of the knowledge (rules) of an expert system is an if-then type formal language (text), which is far from the natural language used by humans. Building the knowledge that an expert system handles requires a great deal of effort from experts with programming knowledge, and it has been impossible to escape the so-called "knowledge acquisition bottleneck problem."
The second problem is the efficiency of inference. In the forward inference carried out by expert systems, facts are explosively generated in the working memory. When the inference is deep, this creates serious implementation problems. For these reasons, expert systems did not bear much fruit in terms of social implementation, and research into them gradually fell out of favor.

In their attempts to overcome the problems mentioned above, subsequent AI researchers have taken paths that can be broadly divided into two, depending on their methodology.
The first approach is to manually create a knowledge base from natural language. The knowledge base referred to here is not the flat text format used in expert systems. It is a database that is "structured" in the same way as the cellular networks that represent knowledge in the human brain, and with the aim of building such a network-format knowledge base, text sentences are tagged with pointers (links) that represent the relationships between words and concepts in a human-wave manner. This has been the subject of active research as a means of structuring web knowledge on the Internet, which has been in widespread use since around 2000, and continues to this day. This research is summarized in Table 1.
In the table, the superiority or inferiority of each method in each category is indicated with ◎ (very superior), ○ (superior), △ (not superior), or × (inferior).

The pioneering research of this approach is Semantic Networks (SN) (Non-Patent Documents 1, 3). SN constructs a network in a relatively simple and naive way, by connecting concept nodes with edges that represent relationships such as 'is' and 'has'. The expressive power of SN is equivalent to that of first-order predicate logic (FOL), but the construction of SN is ad-hoc and difficult to predict, and inference is limited to a simple process of structure matching between a knowledge graph and a query graph.
The "Conceptual Graph" (CG) (Non-Patent Document 4) proposed by Sowa is a more sophisticated version of SN, and succeeded in eliminating edge labels by providing two types of nodes, concept nodes and concept relation nodes representing cases. This improved the homogeneity of the graph and clarified the correspondence with FOL, but it has problems such as the need for graph structuring (nesting) to express implication, for example. However, the method of describing clauses by conjunctions of (1) argument predicates (concepts) and (2) argument predicates (concept relations), introduced in CG, became a standard for converting natural language to FOL. CG is also the root of the recent Google "Knowledge Graph" (Non-Patent Document 8).

As an extension of this research, today, several projects (DBpedia, Freebase, UMBEL, etc.) are running to build knowledge bases based on text information on the Web. The main description tools used in these projects, Resource Description Framework (RDF) and Web Ontology Language (OWL), set simple logical units such as 'SVO triples' as units of expression in order to construct graphs, and therefore take the position that they do not extract 'deep' meanings from texts. In addition, Description Logic (DL) (Non-Patent Documents 2, 7), which provides the theoretical foundation of OWL, limits the description target to a part of first-order predicate logic (FOL), i.e., logic that does not include predicates with three or more arguments or function symbols, while providing a rich symbol library to improve user friendliness. In this sense, today's human-wave approach to constructing natural language knowledge bases sacrifices some of the ability to express logic in order to describe structured (graph) databases in a simple and easy-to-understand way.

The second way to solve the bottleneck of knowledge acquisition in expert systems is research into "natural language processing," which aims to automatically process natural language using computers. So-called "natural language understanding" has attracted attention since the early days of AI research, and prototypes have been created that can converse with humans in certain limited situations, but it is only relatively recently that natural language processing has attracted significant social attention as a feasible technology. This approach does not necessarily involve building a structured database from natural language, but instead converts natural language sentences into a unique expression format that corresponds to the method. The flow of this research is summarized in Table 2.

In natural language processing based on "shallow semantic representation," statistical information on words and keywords is extracted from natural language sentences and analyzed (Ji jkoun & Ri jke 2005, among many others). This is largely due to the "connectionism approach," which has been attempted separately from the "symbol processing approach," and its extensions, such as machine learning (Non-Patent Document 1) and deep learning.
Text written in natural language is converted into mathematical expressions such as word frequency vectors and feature vectors, and statistical analysis is performed using tools such as R. Commercial text mining tools that implement this technology are highly robust in extracting information, but the black-box knowledge is not explicit and the deeper meaning of the text cannot be understood. These tools are excellent at analyzing free-form text such as reviews, questionnaires, and word-of-mouth, and roughly grasping the contents and trends contained therein, and there are many examples of their application.

Meanwhile, IBM's Watson is based on shallow semantic representation, but also incorporates semantic correction and learning functions, and is able to provide answers to questions with priorities set by answer parameters based on document information provided by the customer. However, from the perspective of practical implementation, many technical issues remain, such as the large amount of resources and time required for adjusting machine learning parameters and building a teaching corpus, as well as the high frequency of incorrect answers, and in actual applications, it is only used to discover and extract knowledge from large volumes of reports such as papers and patent documents.

At the other extreme, natural language processing based on "deep semantic representation" (CCG + Montague semantics) has the ability to express deep logic using higher-order logic such as intensional logic and typed dynamic logic as a method of expressing the meaning of language, but the sequential inference processing based on lambda expressions lacks robustness (Montague 1974, Toji 2006, "Can a Robot Get into Tokyo University?" Project Mathematics Department Team 2011, etc.). In addition, the lambda expressions used for semantic representation are not structured.
As described above, "shallow semantic representation" and "deep semantic representation" have advantages and disadvantages in terms of expressive ability and inference, but there is a different method called "simplified semantic representation" called Concept Dependency Theory (Schank 1975, etc.). In this theory, the meaning of a sentence can be expressed independent of the language type by graphing the meaning contained in a natural language using only 11 basic verbs. Although the data format of Concept Dependency Theory is graph-structured, it cannot express logically deep meaning, and the inference rules are also limited.

Ichiro Kobayashi: Fundamentals of Artificial Intelligence. Science Press (2008) Baader, F. , Horrocks, I. , Sattler, U. :Chapter 3 Description Logics. In: van Harmelen, F. , Lifschitz, V. , Porter, B. (eds.): Handbook of Knowledge Representation. Elsevier (2007) Randal, D. M. :Semantic Networks. In: Ringland, G. A. , Duce, D. A. (Eds.): Knowledge Representation-An Introduction. Research Studies Press, Ltd. (1988) Chapter 3, pp. 45-80 Sowa, J. F. : Conceptual Structures: Information Processing in Mind and Machine. Addison-Wesley: Reading, MA. (1984) Suzuki, H. , Yoshida, M. , Sawai, H. :A network representation of first-order logic that uses token evolution for interference. Journal of Information Science and Engineering (JISE) 30(3) (2014) 669-686 Hideaki Suzuki: A Petri net-based inference model that commonly expresses logic and language. Proceedings of the Society of Instrument and Control Engineers, Systems and Information Division Academic Conference (SSI-2014). (2014) GS11-6, pp. 254-260 Akama Seiki: Introduction to the Semantic Web. Cut System (2011) Zhang, L. :Knowledge graph theory and structural parsing. Thesis, Twente University Press (2002) Harada, M., "Semantic Analysis Opens Up the World of Natural Language Processing," NICT-EDR Symposium, 8, National Institute of Information and Communications Technology (NICT) (April 2007). Hideaki Suzuki: Logical Networks Representing FOL: An Attempt at Transformation from Dependency Structures. Proceedings of the 21st Annual Conference of the Association for Natural Language Processing (2015) 976-979 Kenji Yoshimura: Fundamentals of Natural Language Processing. Science Press (2000)

Representing natural language in a graph-structured database has two major advantages: it is easy to read and parallel and distributed. A graph structure is an ideal tool for describing semantic connections between words or concepts, and all processing on a graph structure can be decomposed into local operations on nodes and arcs, which can be carried out in a parallel and distributed manner. The natural language processing approaches we have seen so far have been developed without necessarily focusing on this graph structure, and therefore have problems such as knowledge not being explicit and deep inference not being possible (shallow semantic representation), and inference being sequential and vulnerable to errors (deep semantic representation). If we could automatically convert natural language into some kind of structured database and perform logical inference on it, we could create a system that solves both the non-explicitness of knowledge and the vulnerability of inference.

When trying to graph the meaning of natural language and create a structured database, the first method that comes to mind is the generation of a tree graph using deep cases (Non-Patent Document 11). Harada et al. have proposed a method that converts natural language sentences into a semantic graph that includes relationships between words, taking into account not only simple dependency structures but also deep cases, and also considers vocabulary from a large-scale dictionary and deep case similarities between words in addition to comparing structures based on graph structure matching for inference (Non-Patent Document 9). However, these methods do not reduce the deep meaning of language to logic, but are merely auxiliary tools for advancing inference through structure matching by visualizing case relationships in a graph. Methods that perform inference by matching cannot handle deep logical inference, and have problems in that they cannot perform deductive inference, such as combining multiple pieces of universal knowledge to derive an answer.

On the other hand, the inventors of the present invention have recently proposed a description format called pKTN (Petri-net-based Knowledge Transitive Network) as a graph representation of a Horn clause set, which is a part of first-order predicate logic (FOL) (Non-Patent Documents 5, 6). pKTN is a Petri net in which constants, variables, function symbols, and predicates of FOL are expressed as places, and deduction on it (backward inference) can be formulated as a search for a subgraph in the Petri net expanded from pKTN and consistency between constraints (unification). The inventors have also proposed an algorithm for solving this problem, called ELISE (ELiminating Inconsistency by SELECTION), which uses local parallel propagation of tokens on a network to find a unified solution together with confidence (performing a kind of ambiguous inference) (Non-Patent Document 5). It has been reported that when there is no redundancy in the network, ELISE converges in a time that is approximately proportional to the number of variables.

In pKTN, a graph equivalent to the clauses of predicate logic is constructed around the predicate verbs of natural language sentences, and is therefore considered to have a high affinity with the dependency structure of natural language sentences. Non-Patent Document 10 takes advantage of this fact to show a simple example of converting a dependency structure graph (a dependency network with a verb predicate as the root and other words as the leaves and branches) extracted from natural language sentences into a homogeneous pKTN without a nested structure. However, this has not yet accurately extracted the deep meanings (deep cases) of natural language and converted all natural language sentences into a network. In addition, the problem of spelling variations in various languages has not been taken into consideration, and it has not been a method of absorbing the differences in various ways of expressing the same meaning and concisely extracting logic equivalent to FOL.
The present invention has been made to solve the problems described above, and its purpose is to provide a method for constructing a structured database that expresses the deeper meaning of natural language sentences in the form of deep cases by automatically processing natural language sentences using a computer, and then performing logical inference on that database.

In one aspect of the present invention, the goal is to create a database of all knowledge expressed in natural language sentences, and knowledge is described using a Petri net that corresponds one-to-one with natural sentences, but expresses meaning in deep cases rather than surface cases of natural sentences. Deep cases such as subject, verb, object, locative case, and temporal case are given to clauses based on information such as surface cases, parts of speech, destination, and aspect, and the meaning of the sentence is expressed in detail using a pKTN graph that describes the relationship between them. By creating a semantic graph using deep cases, the logical structure becomes clear, and it becomes possible to deal with paraphrasing of natural sentences using conversion rules. Inference from structure matching or sentence dependency cannot be made when the expressions are different even if the same meaning is expressed, but on a pKTN constructed using deep cases, judgments can be made based on logical inference, and it is not affected by differences in the expressions of the sentences. Furthermore, by ranking and graphing not only a single sentence but also the document structure, inference taking into account the context may be possible.

According to this invention, logic expressing deeper meanings is extracted from sentences written in natural language that contain a mixture of mathematical expressions, tables, and itemized lists, and is stored in a network-type database with deep cases as labels.
Furthermore, according to the present invention, it is possible to carry out deduction and calculations on the constructed database, thereby obtaining a desired answer to a question.

FIG. 1 is a block diagram showing the configuration of a calculation processing device 100 according to the present embodiment. FIG. 1 is a block diagram showing the functional configuration of a calculation processing device 100 according to an embodiment of the present invention. FIG. 13 is a diagram showing a list of labels prepared in a predicate word place as aspects. A diagram showing a list of labels prepared for arcs as deep cases. "Corresponding function" is the meaning of the function symbol used in the implicit Petri net that represents a tautological formula. This is a diagram showing the FOL expression/logic program (1) expressed in pKTN. This is a diagram showing the FOL expression/logic program (2) expressed in pKTN. This is a diagram showing the FOL expression/logic program (3) expressed in pKTN. This is a diagram showing the FOL expression/logic program (4) expressed in pKTN. FIG. 13 is a diagram showing an example of dividing a sentence into phrases and converting the surface cases contained in the phrases into deep cases. FIG. 13 is a diagram showing an example of converting a surface case into a deep case based on target information; This is a diagram showing the process of storing natural language sentences as knowledge in the pKTN database. FIG. 13 is a diagram showing how pKTN links are established during inference. This is a diagram of pKTN in which a question sentence pKTN and a knowledge sentence pKTN are connected by a link. This is a diagram of a stretched pKTN in which a question sentence pKTN and a knowledge sentence pKTN are connected by a link and oriented. This is a diagram of a stretched pKTN in which a question sentence pKTN and a knowledge sentence pKTN are connected by a link and oriented. FIG. 13 is a diagram showing an example of a fact default (an implicit Petri net representing a tautology). This is a diagram of pKTN in which a question sentence pKTN and a knowledge sentence pKTN are linked to a fact default. FIG. 13 is a diagram showing an example of rule defaults. This is a diagram of pKTN in which a question sentence pKTN is linked to a knowledge sentence pKTN via a rule default. This is a diagram of pKTN in which a knowledge sentence pKTN and a question sentence pKTN expressed using the outer relative case are linked together. This is a diagram of pKTN, which shows the temporal order within a complex sentence by connecting the temporal cases of the complex sentence through inequality predicates. FIG. 1 is a diagram showing the structure of a noun ontology. This is a diagram of pKTN in which the constant places of the knowledge sentence pKTN are linked to the ontology graph so that they can be matched with superordinate words in the question sentence pKTN. FIG. 1 illustrates a document structure ontology. This is a diagram showing one of the pKTNs derived from a sentence with a conjunction expressing OR. Hanako is the negated condition. This is a diagram showing one of the pKTNs derived from a sentence with a conjunction expressing OR. Taro is the negated condition. This is a diagram of pKTN that makes inferences by comparing the conditions of the knowledge sentence pKTN connected to the question sentence pKTN with other knowledge. This is a pKTN diagram in which the phrase "A's B" is expressed in the "desu" aspect, treating "no" as a function. FIG. 13 is a diagram showing an example of an implicit Petri net for expressing the abbreviated meaning of the phrase "A's B." This is a diagram of pKTN showing an example in which a question sentence pKTN that asks about the meaning implied in the phrase "A's B" can be answered by linking it to a knowledge sentence pKTN or an implicit Petri net. This is a diagram of pKTN in which a knowledge sentence pKTN in which units are expressed by functions and a question sentence pKTN are linked together. FIG. 1 illustrates the basic concept of a data flow graph. FIG. 13 is a diagram showing an example in which calculations can be performed within pKTN by assigning arithmetic functions to operators and numerical values to arguments.

In the following, with reference to the drawings, a method for converting various expressions into pKTN/FOL will be described using Japanese natural language as an example, as a typical embodiment of the present invention.

Modalities are what support the meaning of natural language sentences at the very base. Modalities are equivalent to auxiliary verbs in English, and in Japanese, they usually appear at the end of a sentence to affirm/negate/modify the meaning of the sentence. Figure 3 shows an example of modalities prepared as pKTN. In this embodiment, modalities are prepared as predicates that express truth or falsehood.

Deep case: Like predicates in ordinary first-order logic, modal predicates have one or more arguments, all of which are assigned a deep case.
Figure 4 shows an example of a deep case.
Deep cases are broadly divided into predicative cases and rhetorical cases. Predicative cases represent elements such as subject, object, and action that express the meaning of a sentence, while rhetorical cases represent additional information equivalent to prepositions in English.

We will now show an example of converting the meaning of a Japanese sentence into a FOL formula/logic program using the above modalities and deep cases.
(1) "do" ([Sb] "Taro", [V] "iku", [Plc] "umi"). (Taro goes to the sea.)
(2) "can" ([Sb] "Taro", [V] "make", [Od] "hat"). (Taro can make a hat.)
(3) "De aru" ([Sb] "Taro", [Od] "Otoko"). (Taro is a man.)
(4) "Should" ([Sb] "Taro", [V] "protect", [Od] "family"). (Taro should protect his family.)
These are also simultaneously represented as Petri net type KTN (pKTN).

向である。後に述べる推論はこのｐＫＴＮ上で行なわれる。 The same example is converted to pKTN in Figures 5 to 8. In the figures, ○ indicates an argument place, and ◇ indicates a predicate place.

The inferences described later are carried out on this pKTN.

- Analysis of natural language In order to convert Japanese natural sentences into pKTN, this embodiment assumes that the results of morphological analysis and dependency structure analysis of the natural sentences have been obtained. In a typical example, the natural sentences are divided into phrases by morphological analysis, and the morphemes in the phrases are tagged with parts of speech (nouns, verbs, ...), attributes (numbers, times, place names, ...), and surface cases (ga case, ni case, ...). In addition, dependency relationships between phrases are extracted by dependency structure analysis.
Figure 9 shows the result of morphological analysis of the sentence "Taro meets Hanako at 9 o'clock." As shown in this example, by preparing appropriate conversion rules, it is possible to recognize deep cases from parts of speech, attributes, and surface cases, and the deep cases obtained in this way are shown after the second arrow in each line in Figure 9. It is noteworthy that the two dyadic phrases "at 9 o'clock" and "to Hanako" are converted to Tim and Oi cases, respectively, but these can only be clearly distinguished by taking into account not only the surface cases but also additional information such as parts of speech and attributes.

As an example of how deep case recognition becomes possible only after taking into account the results of dependency structure analysis, we will next take the example of passive sentence conversion. The active sentence "Hanako confessed to Taro" and the passive sentence "Taro was confessed to by Hanako" both express the same meaning, so it is desirable to convert them into the same pKTN if possible. For this reason, in this embodiment, the passive modal "sareru" is not included in the modal list (Figure 3), and when a passive sentence is read, the active sentence of that sentence is inferred and the deep case therein is recognized by the conversion rules.
Figure 10 shows an example of this. In this example, the target of the ga-case "Taro is" is "confessed" in the passive voice, so the deep case of the ga-case is recognized as the Oi-case. Also, the target of the ni-case "Hanako ni" is "confessed" in the passive voice, so the ni-case is recognized as the Sb-case. This example shows that the ga-case/ni-case do not simply become the Sb-case/Oi-case, but can be recognized as different deep cases depending on the attributes of the target.

As can be seen from this example, when attempting to accurately extract meaning from natural language sentences and apply it to logic in the present invention, it is essential to prepare detailed conversion rules that also include various exception handling procedures.

FIG. 11 shows an exemplary implementation for doing this.
Detailed conversion rules are not written directly into the conversion program, but the conversion program reads a separately prepared conversion rule base to perform exception processing based on case distinctions.
By manually preparing a set of conversion rules using an appropriate interface, the network generation unit can extract logic accurately.

Links for inference To make inferences using pKTN, it is first necessary to create pKTNs for the knowledge sentence and question sentence. All pKTNs created at this point are undirected. The question sentence pKTN is created by determining the deep case of the interrogative word through morphological analysis, dependency structure analysis, and rules, and then entering "_q" as the question variable in the deep case argument.
Next, a link is established between the predicate place of the created question sentence pKTN and the corresponding knowledge sentence pKTN. To select the corresponding knowledge sentence pKTN, the case that takes a constant argument in the question sentence pKTN is used as a reference, and the knowledge sentence pKTN whose modality and deep case argument match is selected. At this time, it is not required that the number of deep cases match.

Figure 12 shows the pKTN for the question "What did Taro buy?" In the question pKTN, "what" is an interrogative + wo case, and "_q" is inserted into the Od case, with "Taro" in the Sb case and "buy" in the V case. The condition for the knowledge sentence to be linked is that the aspect, Sb case, and V case match. Figure 13 shows the knowledge sentence pKTN with this condition and the link. The knowledge sentence here is "Taro gives the ring he bought to Hanako."

Directed and Undirected Petri Nets In one aspect of the present invention, logical inference on pKTN is a backward inference starting from a question, which is formulated as a breadth-first search on an AND/OR tree with the question pKTN as the root. The tree is oriented from the root to the leaves, so to make inference on pKTN, it is necessary to change all the arcs of the linked question and knowledge sentences pKTN to directed ones.
Examples are shown in Figures 14 and 15. In Figures 14 and 15, the knowledge sentence pKTN corresponding to the compound sentence (a sentence having multiple verbs/predicates) "Taro gives the ring he bought to Hanako" is oriented in the opposite direction when it is connected to different question sentences pKTN, "What did Taro buy?" and "What did Taro give?" As shown here, in the present invention, a compound sentence in a natural language, which generally has an unclear condition-consequence relationship, is made into a knowledge sentence pKrN as an undirected Petri net, and is made oriented after being connected to the question sentence pKTN by a link.

・Fact default Backward inference in first-order predicate logic/Horn clause logic cannot be completed without a fact clause that makes the atoms in the body of the rule clause true. In the real world that natural language sentences target, there are many common knowledge and assumptions that are not explicitly stated in the natural language, and pKTN inference can only be completed by preparing fact clauses and rule clauses that represent them as implicit Petri nets (defaults).

Figure 16 is an example of a default prepared for this purpose. This means that "x does y to [object that x does y to]", and it holds true no matter what words that fit the deep case are placed in the variables x and y (always true). By connecting this default to Figure 14, the pKTN in Figure 17 is generated, and an inference that starts with "_q" on this is completed with the knowledge being proven correct by default. The answer obtained is "q = ring" or "q = object ([Sb] Taro, [V] give) (what Taro gave)."

・Rule default In natural language sentences, by making some verbs antonyms and replacing the deep case, the meaning of the sentence becomes the same as before the replacement. For example, if the sentence "Taro gives Hanako a ring" is converted into a FOL formula/logic program,
'To do' ([Sb] 'Taro', [V] 'to give', [Od] 'ring', [Oi] 'Hanako').
However, by changing the deep case V case from "give" to "receive" and switching it from "Oi to Sb" and "Sb to Src", we can create "do" ([Sb] "Hanako", [V] "receive", [Od] "ring", [Src] "Taro").
means "Hanako receives a ring from Taro," and the meaning of the sentence remains the same as before it was changed. A situation in which such a rephrasing is necessary occurs, for example, when the knowledge "Taro gives a ring to Hanako" is used and the question "What did Hanako receive from Taro?" is asked.
In this case, simply connecting the knowledge sentence pKTN and the question sentence pKTN will not be enough to answer the question because the deep cases will not match. Therefore, in order to exchange the deep cases, an implicit Petri net (rule default) is prepared that expresses a paraphrase of "y gives z to x ⇔ x receives z from y" as shown in Figure 18.

By inserting the default in FIG. 18 between the question sentence pKTN and the knowledge sentence pKTN as shown in FIG. 19, the pKTN inference is completed.

・Expression of Complex Sentences Using "External Relatives" Natural language contains many complex sentences. "Taro gives the ring he bought to Hanako." shown in Figure 12 is an example of a complex sentence. In this case, the pKTNs corresponding to the two predicates "buy" and "give" are connected via the variable x, which is connected to the concept of "ring", and are expressed as a single pKTN.
However, there are many compound sentences that do not share a common word. For example, the sentence "Taro hears that Hanako is getting married" does not share a common word.
In cases like this example, by using the "outer relative case", it is possible to express the compound sentence in a logic program without losing its meaning. The "outer relative case" (Out case) has the function of grouping together all of the predicates that contain it. Therefore, a compound sentence can be expressed by preparing an Out case for the sub-sentences that make up the compound sentence, and a case that receives it in the main sentence, and connecting them with a common variable. In the above example, by setting an Out case for the sub-sentence "Hanako is getting married" and receiving it with the objective case of "Taro listens", the FOL formula/logic program becomes:
``To do'' ([Sb] ``Taro'', [V] ``to listen'', [Od]_x1), ``To do'' ([Sb] ``Hanako'', [V] ``to get married'', [Out]_x1).
This is expressed as:
FIG. 20 shows the state in which the pKTN representing this is connected to the question pKTN "What will Taro hear?".

The external relative case can be combined with various rhetorical cases such as time, place, means, condition, cause, and purpose in addition to the Od case, and can express various semantic connections between the main and subordinate clauses. For example, if the sentences are connected with the conjunction "if...", a conditional case can be used, and if the sentences are connected with "therefore...", a complex sentence can be expressed by connecting the causal case and the external relative case.
There are also compound sentences that express a time order. When expressing a time order, a pKTN is created by combining an inequality predicate with an Rg case and an Lf case. For example, in the sentence "Taro drinks beer before going to bed," a pKTN is created as shown in Figure 21.
By connecting the Tim cases of each verb with an inequality sign, the order in which the actions were performed can be expressed. Here, "_t1" and "_t2" are examples of variable places that represent numerical values.

- Ontology In the previous examples, words such as "Taro,""ring," and "buy" were written directly in the constant places of pKTN. However, if the knowledge sentence "Taro has a car" is asked "Does Taro have a car?" and an answer cannot be given because of a word mismatch, this is inconvenient. Similar problems occur not only between synonyms such as "car" and "car," but also between hypernyms and hyponyms. For example, when the knowledge sentence "Taro drives an electric car" is asked "Does Taro drive a car?", it is problematic that an answer cannot be given because "electric car" and "car" are not recognized as being related.
In the present invention, this problem is solved by treating the terms that fit into the constant places of pKTN as concepts, and preparing a synonym list that represents the correspondence between concepts and words, and an ontology that represents the superior/subordinate relationship of concepts. An example of an ontology (which is represented here as words for simplicity) is shown in FIG.

As can be seen from FIG. 22, "automobile" is a hyponym of "car", and "electric car" is a hyponym of "car", so it can be seen that "electric car" is a member of the group of "car".
As a method for using this ontology in pKTN inference, this embodiment shows a method in which a network database representing the ontology is prepared separately and inference is carried out while referring to this.
Figure 23 shows the pKTN when the question "What kind of car did Taro drive?" is asked for the knowledge sentence "Taro drove an electric car." Although "car" and "electric car" are not the same word, when tracing the ontology of "electric car" upwards, "car" exists. Therefore, it is clear that "electric car" is a member of the "car" family, and the question can be answered with "driving an electric car."

Furthermore, when referencing an ontology, you always trace back to the top on the knowledge sentence side. This is because tracing back will result in an incorrect inference. For example, consider the case where the knowledge sentence is "Taro drove a car," and you ask the question "Did Taro drive an electric car?" If you trace back from "electric car" to the higher level "car" on the question side, you will get the incorrect answer "drew."

- Title Case In the various knowledge sentences pKTN created by reading documents, if it is possible to retain information about which document the pKTN was created from, it is possible to specify the range of the knowledge sentences and make inferences.
According to the present invention, for structured documents such as document titles, chapters, paragraphs, lists, and tables, a special case called a title case is added to the deep case, making it possible to provide information about which sentence the knowledge sentence pKTN was generated from.

Title cases can be used for range searches by specifying documents and paragraphs, and are also effective for displaying the title, paragraph, and sentence that are the source of the knowledge when providing an answer. The document structure referenced by title cases is expressed as a tree-structured graph like that shown in Figure 24, just like ontologies. The place where the title case argument of each knowledge sentence pKTN goes points to the terminal leaf in Figure 24, and where it is included in the document hierarchical structure can be specified by tracing the tree in Figure 24 upwards.
The method of referencing the document structure ontology for the title case is the same as that for the concept ontology. When matching the title case arguments of the question sentence pKTN with those of the knowledge sentence pKTN, it is not enough to simply match symbols, but the document structure ontology is traced upwards on the knowledge sentence side to see if there is anything that matches. This makes it possible to search for knowledge sentences in a wide range specified from the question sentence side.

・Logical AND/OR
In natural language, logical expressions such as negation, AND, and OR appear. In this embodiment, negation is expressed in the predicate place by including it in the modality as shown in FIG. For AND and OR, they are expressed as a plurality of Horn clauses that are equivalently converted into a logical formula.
First, consider the case where nouns and verbs are connected by a logical relationship.

といった等価関係を用いると複数のホーン節へと変換することができる。ここに⇒の後に示した式は、Ｐｒｏｌｏｇプログラムの記法に習い、用意する節を「．」（ピリオド）で区切って表したものである。In the nominative case:
"A and B run" = "A runs" ∧ "B runs."
"A or B runs" = "A runs" ∨ "B runs."
For verbs:
"Taro does A and also does B" = "Taro does A" ∧ "Taro does B".
"Taro does A or Taro does B" = "Taro does A" ∨ "Taro does B".
For time cases:
"Taro eats at A and B" = "Taro eats at A" ∧ "Taro eats at B."
"Taro eats at A o'clock or B o'clock" = "Taro eats at A o'clock" or "Taro eats at B o'clock."
As shown in the above examples, when a deep case argument has words connected with logical relations ∧ and ∨, those logical relations can be externalized and the meaning of the sentence can be described as a logical formula connecting multiple propositions. Then,

Using such an equivalence relation, it can be converted into multiple Horn clauses. The expression shown after the ⇒ is a Prolog programming notation in which the clauses to be prepared are separated by "." (period).

Figures 25 and 26 show pKTNs derived from the natural sentence "Taro or Hanako is in charge." In this example, it is converted into two knowledge sentences pKTNs that express the meanings "If Hanako is not in charge, Taro is in charge" and "If Taro is not in charge, Hanako is in charge." In these pKTNs, Horn clause rules are created by connecting the "conditional case" with the "external relative case."

これらの式を用いると、例えば「ＢまたはＣが死ぬとＡが走る」という文は、「Ｂが死ぬとＡが走る」「Ｃが死ぬとＡが走る」という２つの複文知識文を表すｐＫＴＮへと変換される。 If, in addition to this knowledge, there exists knowledge that "Hanako is not in charge," and the question "Who will be in charge?" is posed, the graph in FIG. 27 is generated, making it possible to answer the question.
As an example of conversion to multiple Horn clauses, the conversion formula for a compound sentence containing a condition-consequence relationship expressed by conjunctions/conjunction particles and containing a logical product/logical formula is shown below.

Using these formulas, for example, the sentence "When B or C dies, A runs" is converted into a pKTN that represents two compound knowledge sentences, "When B dies, A runs" and "When C dies, A runs."

・A's B
The phrase "A's B" that frequently appears in Japanese natural text is a significant abbreviation. For example, the phrase "Taro's hat" implies the meaning "Taro owns a hat." In the present invention, in order to explicitly express the previously implied meaning as a means of generating a network from "A's B," a special implicit Petri net (default) is also generated in addition to the knowledge sentence Petri net. In the following embodiment, "Taro's hat" is taken as an example. If this phrase is found in a given knowledge sentence, it is interpreted as "The hat is Taro's." and "no" is treated as a function to create a pKTN as shown in FIG. 28. Here, the dotted circle represents a function symbol place. Meanwhile, an implicit Petri net (FIG. 29) is generated that expresses the equivalence of two sentences including arbitrary variables x and y, "x is y's" and "y owns x," and these are used for inference.

Fig. 30 shows a pKTN that is referenced when the question "What does Taro own?" is asked in response to the knowledge sentence "Hanako made Taro's hat." In Fig. 30, the question sentence pKTN is connected to the knowledge sentence pKTN via an implicit Petri net that represents a paraphrase, and is further connected to an implicit Petri net that represents a tautology, so that the answer "What Hanako made, the hat" can be given.

Units and suffixes In natural language, frequently used units and suffixes can be more accurately extracted logically by treating them separately from the word base. In the present invention, by treating units and suffixes as functions with the word base as an argument, "word base + unit" is regarded as an argument and assigned one argument place. Figure 31 shows an example of a knowledge sentence pKTN corresponding to "Sulfuric acid is acidic" created in this way, connected to a question sentence pKTN corresponding to "What is the nature of sulfuric acid?" In the case of this question sentence, by interpreting "what nature" separately from "what + nature", a contradiction occurs with other knowledge sentences pKTN, such as "sulfuric acid is a toxic substance," and the connection is not possible.

In order to integrate pKTN reasoning and formula calculation, the present invention introduces the idea of data flow graphs to express formulas and algorithms into pKTN. Figure 32 shows an example of a data flow graph.
A data flow graph is a graph that allows any mathematical formula or algorithm to be expressed by assigning numerical values to the edges and operators to the nodes. In Figure 32, numerical values flow through the graph, making it possible to obtain the calculation result of 6 x ((3 + 9) - 7) = 30.
Another notable feature of dataflow graphs is that they enable solutions to numerical equations to be found by flowing data in the reverse direction while performing local inverse calculations at each operator node.

In the example of Figure 32, for example, by flowing numerical values from left to right, it is possible to find the solution x = 9 that satisfies the equation 6 x ((3 + x) - 7) = 30. Something similar can be done in pKTN as well. To achieve this, in this invention, function symbols that correspond to mathematical symbols such as + and - are prepared, and the transitions that bring these together are given the function of numerical calculation.
Figure 33 shows the pKTN when the question "How many years before 2017 did Taro live in Yokohama?" is asked for the knowledge "Taro lived in Yokohama until 2000." The equation x = 2017 - q is embedded in the lower left part of Figure 33, and by performing the reverse calculation q = 2017 - x during inference, the above question can be answered as "q = 17 (until 17 years ago)."

Examples of applications of the logical inference system according to the present invention include the following.
(1) Operation support/agent system for transactions and settlements in the financial and securities fields. Based on documents used by customers in their operations (manuals for each institution, laws, etc.), customers can quickly and accurately obtain confirmation information on the content of certain transactions by querying the system without having to manually check a huge number of manual documents. Even inexperienced operators can accurately retrieve all manual information, which is expected to reduce human error and alleviate labor shortages.
(2) An inquiry and information provision system based on know-how documents, manuals, etc. at factories and production sites. At manufacturing sites, a huge amount of manual documents are accumulated, from research and development to production, sales, distribution, and maintenance. If a database that integrates these documents and allows inquiries is created, it will be possible to effectively utilize these documents and eliminate human error.
(3) By incorporating it as an inference engine for a chatbot with voice IF, it can handle inquiries at call centers.
(4) "Integration of medical information" throughout the local community through logical database organization of medical information (electronic medical records, test and diagnosis information, etc.).
(5) Labor savings and improved service quality achieved by delegating customer service duties to local government offices and commercial facilities.

Claims (11)

A logical inference system for performing inference based on knowledge expressed in natural language or mathematical expressions,
The logic and algorithms extracted from the knowledge
A Petri net in predicate logic form,
The system comprises a plurality of places, a plurality of transitions, and a plurality of arcs respectively connecting between the places and the transitions,
The plurality of places include a predicate place corresponding to the truth or falsity of the natural language sentence, a term place corresponding to various concepts and quantities of the natural language, and a function symbol place corresponding to various connections between concepts of the natural language, numerical operations, and units;
the plurality of arcs have labels corresponding to deep cases expressing deep meanings of the natural language;
The description word place includes modal places corresponding to various modalities and negations of the natural language, and concept places corresponding to various concepts,
The term places include constant places corresponding to various concepts and quantities in the natural language, and variable places corresponding to various pronouns, indefinite articles, and variables in the mathematical formula.
A logical inference system expressed using knowledge Petri nets. A logical inference system for performing inference based on knowledge expressed in natural language or mathematical expressions,
The logic and algorithms extracted from the knowledge
A Petri net in predicate logic form,
The system comprises a plurality of places, a plurality of transitions, and a plurality of arcs respectively connecting between the places and the transitions,
The plurality of places include a predicate place corresponding to the truth or falsity of the natural language sentence, a term place corresponding to various concepts and quantities of the natural language, and a function symbol place corresponding to various connections between concepts of the natural language, numerical operations, and units;
a logical inference system expressed using a knowledge Petri net, in which the plurality of arcs have labels corresponding to deep cases that represent deep meanings of the natural language,
A rule for determining modality and deep case from morphological analysis and dependency structure analysis of a natural language sentence,
Using rules for determining aspects and deep cases by referring to attribute information including parts of speech, particles, destinations, and time and place for each clause of the natural language sentence,
a network creation unit that converts the natural language sentence into the knowledge Petri net;
Logical inference system. A method for resolving conflicts and performing numerical calculations based on local symbol contradictions on the knowledge Petri net;
By performing a breadth-first search on a virtual AND/OR graph created by expanding the knowledge Petri net,
3. The logical inference system according to claim 1, further comprising an inference execution unit that performs backward inference on the logic extracted from said knowledge. As part of the knowledge Petri net,
A network creation part converts complex sentences in natural language, which do not have a clear premise-conclusion relationship, into undirected Petri nets, whose arcs have no direction;
a network extraction unit for receiving a question, extracting related sub-Petri nets from the knowledge Petri net, and providing a direction to the extracted sub-Petri nets;
The logical inference system of claim 3 . A Petri net representing a tautology in predicate logic,
It generates implicit Petri nets that correspond to the assumptions and common knowledge implicit in natural language sentences,
A network creation unit for connecting the implicit Petri net with the knowledge Petri net is further provided.
The logical inference system of claim 3 . As part of the knowledge Petri net,
Various complex sentences in natural language expressed with conjunctions, conjunctions, and relative pronouns,
The outer relative case, which is a special deep case,
A network creation unit converts the data into an undirected Petri net including deep cases expressing objects, time, place, scope, actors, means, partners, opponents, conditions, assumptions, presuppositions, causes, reasons, purposes, topics, sources, positions, metaphors, quantities, frequencies, and descriptions, or predicate places for comparing quantities.
The logical inference system of claim 3 . As part of the knowledge Petri net,
Natural language sentences in documents structured by chapters, paragraphs, lists, and tables,
A network creation unit is further provided for converting the Petri net into a Petri net including a case for the title specially prepared as a deep case.
The logical inference system of claim 3 . As part of the knowledge Petri net,
a network creation unit that performs logical conversion of a natural language sentence including a logical OR relation and converts the sentence into a Petri net corresponding to a plurality of Horn clauses;
The logical inference system of claim 3 . As part of the knowledge Petri net,
Generate an implicit Petri net corresponding to the implicit knowledge contained in the abbreviated natural language expression "A's B";
A network creation unit for connecting the implicit Petri net with the knowledge Petri net is further provided.
The logical inference system of claim 3 . A logical inference method for performing inference based on knowledge expressed in natural language or mathematical expressions, comprising:
The logic and algorithms extracted from the knowledge
A Petri net in predicate logic form,
The system comprises a plurality of places, a plurality of transitions, and a plurality of arcs respectively connecting between the places and the transitions,
The plurality of places include a predicate place corresponding to the truth or falsity of the natural language sentence, an argument place corresponding to various concepts and quantities of the natural language, and a function symbol place corresponding to various connections and units between concepts of the natural language,
the plurality of arcs have labels corresponding to deep cases expressing deep meanings of the natural language;
The description word place includes modal places corresponding to various modalities and negations of the natural language, and concept places corresponding to various concepts,
The term places include constant places corresponding to various concepts and quantities in the natural language, and variable places corresponding to various pronouns, indefinite articles, and variables in the mathematical formula.
A logical inference method expressed using Petri nets. A logical inference method for performing inference based on knowledge expressed in natural language or mathematical expressions, comprising:
The logic and algorithms extracted from the knowledge
A Petri net in predicate logic form,
The system comprises a plurality of places, a plurality of transitions, and a plurality of arcs respectively connecting between the places and the transitions,
The plurality of places include a predicate place corresponding to the truth or falsity of the natural language sentence, a term place corresponding to various concepts and quantities of the natural language, and a function symbol place corresponding to various connections between concepts of the natural language, numerical operations, and units;
The method of logical inference is expressed using a knowledge Petri net, in which the plurality of arcs have labels corresponding to deep cases that represent deep meanings of the natural language,
A rule for determining modality and deep case from morphological analysis and dependency structure analysis of a natural language sentence,
Using rules for determining aspects and deep cases by referring to attribute information including parts of speech, particles, destinations, and time and place for each clause of the natural language sentence,
a network creation unit that converts the natural language sentence into the knowledge Petri net;
Logical reasoning methods.

Images