JP7299822B2 - Apparatus for semantic analysis of trade transaction message, semantic analysis method and program - Google Patents

Description

The present invention relates to a semantic analysis device, a semantic analysis method, and a program for trade transaction messages.

Generally, in the payment settlement between the exporter and the importer in trade, as disclosed in Non-Patent Document 1, the payment is not made directly between the exporter and the importer, but the bank issued Settlement using a letter of credit (hereinafter referred to as L/C) is performed (L/C settlement). The L/C is prepared by the opening bank, which is the importer's bank. The created L/C is sent from the opening bank to the advising bank, which is the exporter's bank, and notified to the exporter.

In L/C settlement, the exporter prepares the necessary trade documents (required documents) based on the notified L/C, submits them together with the L/C to the advising bank (purchasing bank), and the purchasing bank By confirming that the documents match the contents of the L/C, the exporter can receive the payment from the advising bank. Then, the required documents submitted are sent from the advising bank to the opening bank, and the opening bank confirms that the documents match the contents of the L/C, and the opening bank pays the advising bank, After that, the advising bank collects the payment by presenting the relevant L/C and necessary documents to the importer. The required documents include the documents necessary for the importer to receive the goods. With such L/C settlement, the exporter can collect the export payment immediately after shipment, and the importer does not need to pay the import payment in advance, making the payment smooth and safe. It has the advantage of being able to

"Flow of L/C Payment", [Online], [Searched on October 16, 2019], Internet <URL: https://www.toishi.info/lc/index.html>

However, the L/C used in the L/C settlement described in Non-Patent Document 1 does not have a fixed description format, and includes items that can be arbitrarily input by the importer. Therefore, the exporter and bank who received the L/C required a great deal of effort to understand and check the contents of the L/C. In addition, there is a risk of transcription errors when the exporter prepares the necessary documents based on the L/C. Thus, from the viewpoint of facilitating procedures in L/C settlement, it is still insufficient.

SUMMARY OF THE INVENTION It is an object of the present invention to provide a semantic analysis apparatus, a semantic analysis method, and a program for trade transaction messages that are capable of facilitating L/C settlement procedures. and

In order to achieve the above object, the semantic analysis device according to the first aspect of the present invention includes:
a morphological analysis unit that divides trade transaction document data into morphemes;
a first entity extraction unit for extracting an entity, which is information necessary for settlement using a letter of credit, from the morphemes divided by the morpheme analysis unit based on a rule-based model predetermined by a user;
a second entity extraction unit that extracts the entity from the morphemes divided by the morphological analysis unit based on a learning model learned from past trade transaction document data;
a determination unit that determines whether or not the extraction result by the first entity extraction unit and the extraction result by the second entity extraction unit satisfy a predetermined criterion;
Analyzing to which classification in settlement using the letter of credit the entity included in the extraction result determined by the determining unit to satisfy the criteria belongs, based on a trade ontology learned according to predetermined conditions. a semantic analysis unit;
a storage unit that stores the entity analyzed by the semantic analysis unit in association with the classification;
an updating unit that performs learning based on the entity stored in the storage unit in association with the classification, and updates the contents of the learning model and the trade ontology;
Prepare.

an output unit that outputs the entity stored in the storage unit together with the classification associated with the entity;
The semantic analysis unit assigns a predetermined score value when analyzing which classification the entity belongs to,
The determination unit is
determining that the criterion is satisfied when the extraction result by the first entity extraction unit and the extraction result by the second entity extraction unit match;
When the extraction result by the first entity extraction unit and the extraction result by the second entity extraction unit do not match, the semantic analysis unit extracts entities included in the extraction result by the first entity extraction unit and the second entity extraction unit. Among the score values assigned by analyzing the entities included in the extraction results based on the trade ontology, the extraction results corresponding to the high score values are regarded as the extraction results that satisfy the criteria. judge,
When the extraction result is extracted only from one of the first entity extraction unit and the second entity extraction unit, the entity included in the extracted extraction result is added to the trade ontology If the score value assigned by analyzing based on is equal to or greater than a predetermined threshold value, the extraction result is determined to be an extraction result that satisfies the criterion;
You may do so.

The trade ontology is a letter of credit analysis that is a learning model for analyzing to which classification in settlement using the letter of credit an entity included in the extraction results determined by the determination unit to satisfy the criteria belongs. The ontology and the entity whose classification in the settlement using the letter of credit is analyzed by the semantic analysis unit based on the ontology and the letter of credit analysis ontology is the classification in the trade necessary data different from the trade transaction document data. a necessary information ontology, which is a learning model for analyzing whether it belongs to
The semantic analysis unit analyzes, based on the necessary information ontology, which classification in the trade necessary data the entity included in the data stored in the storage unit belongs to,
The storage unit stores the entity analyzed by the semantic analysis unit based on the necessary information ontology in association with the classification of the trade necessary data.
You may do so.

The classification in the trade necessary data is different for each trader,
The necessary information ontology indicates to which classification in the trade necessary data an entity whose classification in the settlement using the letter of credit is analyzed by the semantic analysis unit based on the letter of credit analysis ontology. The basic ontology of necessary information, which is a learning model for analysis, and the entity whose classification in the trade-required data is analyzed by the semantic analysis unit is different for each trader in the trade-required data. a necessary information individual ontology, which is a learning model for analyzing whether it belongs to a classification,
The semantic analysis unit analyzes, based on the necessary information individual ontology, to which class the entity analyzed based on the necessary information basic ontology belongs to, which differs for each trader in the necessary trade data, and
The storage unit stores the entity analyzed by the semantic analysis unit based on the individual necessary information ontology in association with the classification that differs for each trader in the necessary trade data,
You may do so.

In order to achieve the above object, the semantic analysis method according to the second aspect of the present invention comprises:
A semantic analysis method by a semantic analysis device comprising a morphological analysis unit, a first entity extraction unit, a second entity extraction unit, a determination unit, a semantic analysis unit, a storage unit, and an update unit,
a morphological analysis step in which the morphological analysis unit divides the trade transaction document data into morphemes;
The first entity extracting unit, based on a rule-based model predetermined by a user, extracts an entity that is information necessary for payment using a letter of credit from the morphemes divided by the morphological analysis step. an extraction step;
a second entity extraction step in which the second entity extraction unit extracts the entity from the morphemes divided by the morphological analysis step based on a learning model learned from past trade document data;
a determination step in which the determination unit determines whether or not the extraction result of the first entity extraction step and the extraction result of the second entity extraction step satisfy a predetermined criterion;
The semantic analysis unit learns, according to a predetermined condition, to which classification in settlement using the letter of credit the entity included in the extraction result determined to satisfy the criteria in the determination step belongs. a semantic parsing step that parses based on the ontology;
a storage step in which the storage unit stores the entity analyzed by the semantic analysis step in association with the classification;
an updating step in which the updating unit performs learning based on the entities stored in the storing step in association with the classification, and updates the contents of the learning model and the trade ontology;
including.

In order to achieve the above object, the program according to the third aspect of the present invention is
to the computer,
Morphological analysis processing for dividing trade transaction document data into morphemes;
a first entity extraction process for extracting an entity, which is information necessary for settlement using a letter of credit, from the morphemes divided by the morphological analysis process based on a rule-based model predetermined by a user;
a second entity extraction process for extracting the entity from the morphemes divided by the morphological analysis process based on a learning model learned from past trade transaction document data;
a determination process for determining whether or not the extraction result of the first entity extraction process and the extraction result of the second entity extraction process satisfy a predetermined criterion;
Analyzing to which classification in settlement using the letter of credit the entity included in the extraction results determined by the determination processing to satisfy the criteria belongs, based on a trade ontology learned according to predetermined conditions. a semantic analysis process;
a storage process for storing the entity analyzed by the semantic analysis process in association with the classification;
an update process for performing learning based on the entity stored by the storage process in association with the classification, and updating the contents of the learning model and the trade ontology;
to run.

ADVANTAGE OF THE INVENTION According to this invention, the procedure in L/C settlement can be facilitated.

It is a figure which shows an example of L/C data. 1 is a block diagram showing an example of an information processing system according to an embodiment of the present invention; FIG. 1 is a block diagram showing an example of a semantic analysis device and a terminal according to an embodiment; FIG. 4 is a diagram showing an example of a master format stored in a storage unit; FIG. It is a figure which shows an example of the trade ontology memorize|stored in a memory|storage part. (A) is a flowchart showing an example of structuring processing, (B) is a diagram showing an example of unstructured data included in L/C data, and (C) is an example of structured data It is a figure which shows. (A) is a diagram showing an example of an extraction result of a morphological analysis unit; (B-1) is a diagram showing an example of an extraction result of a first entity extraction unit; (B-2) is a diagram showing an example of a first entity extraction unit; FIG. 10 is a diagram showing an example of an extraction result of a two-entity extraction unit; 6 is a flowchart illustrating an example of entity extraction processing; FIG. 10 is a diagram showing an example of a result of processing by the semantic analysis unit; (A) is a diagram showing an example of a trade ontology according to a modification, and (B) is a diagram showing an example of I/V data. (A) is a diagram showing an example of an I/V ontology included in a trade ontology according to a modification, and (B) is a diagram showing an example of input data and output data of I/V data.

A semantic analysis device according to an embodiment of the present invention will be described in detail with reference to the drawings. In addition, in the figure, the same code|symbol is attached|subjected to the same or corresponding part. A case where the semantic analysis apparatus of the present invention is applied to a semantic analysis system used by an exporter (consignor) for L/C settlement will be described below as an example. Specifically, the L/C prepared by the opening bank, which is the importer's bank, is sent from the opening bank to the advising bank, which is the exporter's bank, and the advising bank notifies the exporter. A case in which the semantic analysis system is applied will be described as an example.

FIG. 1 is a diagram showing an example of L/C data prepared by an opening bank and input to a semantic analysis system by an exporter via an advising bank. As shown in the figure, the L/C data contains the information (item data) required for trade transactions, such as the product name, quantity, unit price, currency type and unit used, etc. include. These item data are arbitrarily input by the importer in random order. In this way, L/C data in which item data is arbitrarily input by the importer is called unstructured data. Also, L/C data in which item data is described based on a predetermined format is referred to as structured data. The L/C data shown in FIG. 1 is an example of unstructured L/C data.

The semantic analysis system according to the present embodiment extracts item data, which is information necessary for L/C settlement, from unstructured L/C data as entities, and based on a pre-designed ontology, Assign meaning to each extracted entity (classify what each entity means). As a result, the semantic analysis system 1 classifies (also referred to as structuring) each item data arbitrarily input by the importer, and generates structured data. By doing so, each item data included in the unstructured data is classified, so that it becomes easy to understand and check the description contents of the L/C, and the procedure for L/C settlement can be facilitated.

Here, the entity in the present embodiment is item data that is information necessary for L/C settlement. It is information that is the subject of trade transactions. In addition to those shown in Fig. 1, it also includes proper nouns such as product model numbers, country names of import destinations, country names of export destinations, and numerical expressions such as dates. An ontology in the present embodiment is a concept defined for giving meaning to entities, and is composed of entity definitions (classes), relationship information indicating relationships between entities, and the like.

The semantic analysis system 1 includes a semantic analysis device 100 and a terminal 200, as shown in FIG. The semantic analysis device 100 is an arbitrary computer device (information processing device) such as a mainframe, a workstation, or a PC (Personal Computer). The terminal 200 is an information terminal (so-called computer) such as a mobile phone, a smart phone, a tablet, or a PC, and constructs a distributed network 210 such as a client-server type. Note that the semantic analysis system 1 is not limited to a client-server system, and may be a cloud computing system, for example. Semantic analysis device 100 generates structured data from unstructured data included in L/C data transmitted from terminal 200 . The terminal 200 accepts input of contract document data such as L/C data sent by the advising bank and purchase order (P/O) data sent by the importer from the exporter, and sends it to the semantic analysis device 100 Output. Semantic analysis apparatus 100 and terminal 200 are communicably connected via network 210 .

Although only one terminal 200 is shown in FIG. 2, the semantic analysis system 1 includes, in addition to the terminal 200 of the exporter, the terminal 200 of the advising bank, the terminal 200 of the opening bank, the terminal 200 of the importer, and the like. A plurality of terminals 200 may be provided so that any terminal 200 can communicate with the semantic analysis apparatus 100 . For example, the L/C data may be directly input to the semantic analysis apparatus 100 via the network 210 from the terminal 200 provided at the advising bank. In this embodiment, the semantic analysis system 1 is composed of the semantic analysis device 100 and the terminal 200 , but the terminal 200 may have the functions of the semantic analysis device 100 . That is, terminal 200 having the functions of semantic analysis device 100 may function as semantic analysis device 100 .

Next, configurations of semantic analysis apparatus 100 and terminal 200 will be described with reference to FIGS. 3 to 5. FIG. First, the configuration of the semantic analysis device 100 will be described. As shown in FIG. 3, the semantic analysis apparatus 100 includes a storage unit 110, a control unit 120, a communication unit 130, and a system bus (not shown) interconnecting these units.

The storage unit 110 includes a ROM (Read Only Memory), a RAM (Random Access Memory), and the like. The RAM is used as a work area for a CPU (Central Processing Unit) of the controller 120 . The ROM stores programs executed by the CPU of the control unit 120, and stores various data used when executing the programs.

Specifically, storage unit 110 stores document data 111 , master format 112 , rule base model 113 , learning model 114 , and trade ontology 115 .

The document data 111 is data input by the user when generating a learning model 114 used by the second entity extraction unit 125b, which will be described later. The document data 111 is, for example, past L/C data, P/O data, and the like.

The master format 112 is data that serves as a template for generating structured data, and is stored in advance in the storage unit 110 . An example of the master format 112 is shown in FIG. As shown in FIG. 4, the master format 112 is composed of classification names of entities to be extracted from L/C data and given meanings, such as "Products", "Quantity", and "Unit Price". The semantic analysis system 1 generates structured data by inputting the extracted and semantic entities into the master format 112 .

The rule base model 113 is a model indicating rules for extracting entities by the first entity extraction unit 125a, which will be described later. Item data to be extracted from the L/C data is registered in the rule base model 113 by the user. It is Also, in the rule-based model 113, for example, a predetermined character string such as "JPY" or "USD" in the unstructured data is given a currency tag, and numbers before or after the character string to which the currency tag is given are added. is written by the user for extracting entities such as extracting as "amount". The rule base model 113 is stored in the storage unit 110 in advance.

The learning model 114 is a model used when a second entity extraction unit 125b, which will be described later, extracts entities. The learning model 114 is generated by performing machine learning based on the past L/C data, P/O data, etc. stored in the document data 111 . Machine learning is performed by a supervised learning algorithm such as NCRF (Neural Conditional Random Fields). The L/C data stored in the document data 111 is used as input data in the machine learning, and the P/O data is used as correct data such as product names and model numbers.

The trade ontology 115 includes data such as entity definitions (classes) in L/C settlement and relationship information indicating relationships between classes. FIG. 5 shows an example of L/C's trade ontology 115 . As shown in FIG. 5, the trade ontology 115 describes classes such as "Products", "Quantity", and "Unit Price", and "Products" includes "Category", "Product Name", and "Model". Relational information of belonging is represented by a tree structure. Trade ontology 115 is designed in advance and stored in storage unit 110 . The trade ontology 115 is used when the semantic analysis unit 122, which will be described later, gives meaning to entities.

Control unit 120 includes a CPU, operates according to a program stored in storage unit 110, and executes processing according to the program. The control unit 120 includes an entity extraction unit 121 , a semantic analysis unit 122 and an output unit 123 as main functional units provided by the programs stored in the storage unit 110 .

The entity extraction unit 121 extracts unstructured data contained in the L/C data, performs morphological analysis, decomposes the unstructured data into morphemes, and extracts entities from the decomposed morphemes. It is a functional part that performs The entity extraction unit 121 further includes a morphological analysis unit 124, a first entity extraction unit 125a, a second entity extraction unit 125b, and a determination unit 126. These functional units cooperate to implement the functions of the entity extraction unit 121 .

The morphological analysis unit 124 is a functional unit that divides unstructured data into morphemes. Data for each divided morpheme is called morpheme data. Further, the morphological analysis unit 124 may include part-of-speech information in each morpheme data, and the first entity extraction unit 125a and the second entity extraction unit 125b, which will be described later, may extract entities based on the part-of-speech information. good.

The first entity extraction unit 125a is a functional unit that extracts entities from the morphological data morphologically analyzed by the morphological analysis unit 124 based on the rule base model 113 defined by the user. Specifically, the first entity extraction unit 125a extracts morphological data that matches the item data registered in the rule base model 113 from the L/C data as an entity. Note that the rule base model 113 used for extraction by the first entity extraction unit 125a is defined by the user. It is possible to extract the item data such as the transaction target that has been changed.

The second entity extraction unit 125b is a functional unit that extracts entities from morphological data that has been morphologically analyzed by the morphological analysis unit 124 based on the learning model 114 based on machine learning. Since the learning model 114 used for extraction by the second entity extraction unit 125b is generated by performing machine learning based on past L/C data and P/O data, the first entity extraction unit 125a and In comparison, there is a possibility that the accuracy may be lowered in terms of extracting item data such as products and model numbers that are newly traded targets, and trade targets whose product names and model numbers have changed. However, since machine learning is performed based on past L/C data and P/O data, it is possible to prevent extraction based on an incomplete rule-based model 113 due to user input errors. can.

The determination unit 126 is a functional unit that determines whether or not the extraction results are valid based on the extraction results of the first entity extraction unit 125a and the second entity extraction unit 125b.

Specifically, the determining unit 126 determines whether or not the entity extracted by the first entity extracting unit 125a and the entity extracted by the second entity extracting unit 125b match. , the extracted entity is determined to be valid.

If the determination unit 126 determines that the entity extracted by the first entity extraction unit 125a and the entity extracted by the second entity extraction unit 125b do not match, the determination unit 126 sends the respective extraction results to the semantic analysis unit 122. output to The semantic analysis unit 122 compares the extraction results of the first entity extraction unit 125a and the second entity extraction unit 125b with the trade ontology 115, calculates the total score value indicating the likelihood of the extraction result, and sends it to the determination unit 126. Output. A method of calculating the total score value by the semantic analysis unit 122 will be described later. The determination unit 126 determines that an extraction result with a high score value among the extraction results of the first entity extraction unit 125a and the second entity extraction unit 125b is appropriate.

Furthermore, when an entity is extracted by only one of the first entity extraction unit 125 a and the second entity extraction unit 125 b, the extraction result is output to the semantic analysis unit 122 . The semantic analysis unit 122 collates the extraction result with the trade ontology 115 to calculate a total score value and outputs it to the determination unit 126 . If the total score value is equal to or greater than a predetermined threshold value, the determination unit 126 determines that the extraction result is valid. When the total score value falls below the threshold, the extraction rule and threshold are relaxed, and the first entity extraction unit 125a and the second entity extraction unit 125b are allowed to re-extract for determination. Note that the threshold may be arbitrarily set by the user's input.

The semantic analysis unit 122 is a functional unit that gives meaning to entities extracted by the entity extraction unit 121 . Specifically, the semantic analysis unit 122 performs a process of assigning entities extracted by the entity extraction unit 121 to classes of the trade ontology 115 . Rules for assigning entities to classes in the trade ontology 115 are realized by a model that has learned correct data in advance by deep learning or the like. For example, in the semantic analysis unit 122, by inputting the extracted entity "ABC" into the trade ontology 115 shown in FIG. 5, the extracted entity is assigned to "Name". Assigned (classified) to one of the classes in the trade ontology 115 . This gives meaning to the entity.

The semantic analysis unit 122 also has a function of calculating a total score value indicating the likelihood of the extraction result by the entity extraction unit 121 when allocating an entity to a class of the trade ontology 115 . Each class of the trade ontology 115 has a predefined score value for each class. The semantic analysis unit 122 classifies all entities included in the extraction results into classes, and integrates the score values of the classified classes to calculate a total score value.

The output unit 123 generates structured data by writing the entities assigned meanings by the semantic analysis unit 122 to the corresponding classification names in the master format 112 .

These functional units cooperate to realize the function of generating structured data from unstructured data in the semantic analysis device 100 .

The communication unit 130 is composed of a device (network card, etc.) for performing data communication with other devices. The semantic analysis device 100 transmits and receives data to and from the terminal 200 via the communication unit 130 .

The above is the configuration of the semantic analysis apparatus 100 . Next, the configuration of terminal 200 according to the present embodiment will be described. The terminal 200 is an information terminal device such as a PC. As shown in FIG. 2, terminal 200 includes control unit 20, storage unit 21, communication unit 22, display unit 23, and input unit 24, and is communicably connected to semantic analysis apparatus 100. there is

The control unit 20 is composed of a CPU and the like. The control unit 20 operates according to a program stored in the storage unit 21 and executes processing according to the program.

The storage unit 21 is composed of ROM, RAM, etc., and stores programs executed by the CPU of the control unit 20 and necessary data.

The communication unit 22 is composed of a device (network card, etc.) for performing data communication with other devices. The terminal 200 transmits and receives data to and from the semantic analysis device 100 and the like via the communication unit 22 .

The display unit 23 is configured by a display such as a liquid crystal display or an organic EL (Electro-Luminescence), and displays L/C data or the like received from the semantic analysis apparatus 100 .

The input unit 24 is composed of a keyboard, mouse, touch panel, etc., and receives inputs such as L/C data.

The above is the configuration of the terminal 200 . Next, the operation of the semantic analysis device 100 will be described with reference to the drawings. First, structuring processing as an operation of the semantic analysis system 1 will be described with reference to FIG. In this example, in order to make the explanation easier to understand, the unstructured data included as the L/C data is the L/C data shown in FIG. 6(B), and the structured data shown in FIG. 6(C) is A case of generating will be described as an example. The L/C data shown in FIG. 6B is generated by SWIFT (Society for Worldwide Interbank Financial Telecommunications), which is a network system for international financial transactions between banks.

When the user operates terminal 200 and inputs L/C data and P/O data into input unit 24 , the L/C data and P/O data are transmitted to semantic analysis apparatus 100 . Execution of the structuring process is started when L/C data and P/O data are received from the terminal 200 . Execution may be started by a user's operation. In this example, the product name ("XXXX" in the figure) and model number ("V1") that have been traded in the past are learned by the learning model 114 as correct data, respectively. It is assumed that the target product name “YYYY” and model number “400” are not registered in the learning model 114 . It is also assumed that both are registered in the rule base model 113 . That is, in this example, a case where the extraction result by the first entity extraction unit 125a and the extraction result by the second entity extraction unit 125b are different will be described as an example.

When the structuring process shown in FIG. 6A is started, the entity extraction unit 121 acquires unstructured data from the L/C data (step S11). As described above, the L/C data includes L/C data of structured data and L/C data of unstructured data. By determining the identification information described in the portion marked with the unstructured data, the L/C data of the unstructured data is obtained. Specifically, “<45A: Description of Services>” shown in FIG. 6B is a message assigned by SWIFT, and “45A” corresponds to identification information. In step S11, the identification information is confirmed and it is determined whether the L/C data corresponding to the identification information is structured data or unstructured data. Note that it is only necessary to determine in advance whether the L/C data corresponding to which identification information is unstructured data. The acquired unstructured data is output to the morphological analysis unit 124 .

After executing the process of step S11 shown in FIG. 6A, the morphological analysis unit 124 analyzes the unstructured data and divides it into morphological data for each morpheme (step S12). As described above, in the process of step S12, each divided morpheme data may be given a part of speech. The analysis result by the morphological analysis unit 124 is output to the first entity extraction unit 125a and the second entity extraction unit 125b. Specifically, in the process of step S12, for the L/C data shown in FIG. , “QTY”, “:”, and “2004” (see FIG. 7A), and output to the first entity extraction unit 125a and the second entity extraction unit 125b. When giving a part of speech, for example, each morpheme data may be associated with a part of speech such as "noun", "preposition", or "other".

Next, upon receiving the data morphologically analyzed by the morphological analysis unit 124, the first entity extraction unit 125a and the second entity extraction unit 125b perform processing for extracting entities from the morphologically analyzed data (step S13). Details of the entity extraction process (step S13) will be described later.

The semantic analysis unit 122 performs a semantic analysis process in which the entities extracted by the entity extraction process in step S13 are assigned to classes of the trade ontology 115 based on the rules of the model in which correct data has been learned in advance by deep learning or the like (step S14). Details of the processing in step S14 will be described in step S107 of the entity extraction processing (step S13) described later. The semantic analysis unit 122 assigns meaning to the entity by assigning the entity to a class that indicates the definition of the entity. Note that the processing of step S14 may be executed when the processing of step S105 in FIG. 8 described later is executed. Since each entity is assigned to each class, the process of step S14 may be skipped and the process of step S15 may be performed. The semantic analysis unit 122 stores, in the storage unit 110, the analysis result indicating the entity and its corresponding classification, which is obtained by the processing of step S14 (the same applies when the processing of steps S109 and S112 of FIG. 8 is performed). Let In addition, when the semantic analysis unit 122 stores the analysis result indicating the entity and its corresponding classification in the storage unit 110, it performs learning based on the analysis result, and updates the contents of the learning model 114 and the trade ontology 115. .

Next, the output unit 123 writes the entity assigned meaning by the semantic analysis unit 122 to the corresponding classification name in the master format 112 to generate the structured data shown in FIG. (step S15), and the process ends.

Next, the entity extraction process (step S13) executed in step S13 of FIG. 6 will be described with reference to FIG.

In the entity extraction process shown in FIG. 8, when data (morphological data) morphologically analyzed by the morphological analysis unit 124 is input, the first entity extraction unit 125a converts the rule base model 113 pre-stored in the storage unit 110 into Based on this, the entity is extracted (step S101). The first entity extraction unit 125a outputs the extraction result to the determination unit 126 when the entity can be extracted. Specifically, in step S101, when the data shown in FIG. 7A is input, data that matches the data registered in the rule base model 113 is extracted. In this example, as described above, the traded product names "YYYY" and "XXXX" and model numbers "400" and "V1" are registered in the rule-based model 113. Therefore, step S101 As the extraction result in , data corresponding to each product name and model number are included as shown in FIG. 7(B-1).

In addition to the process of step S101, when the data morphologically analyzed by the morphological analysis unit 124 is input, the second entity extraction unit 125b extracts the entity based on the learning model 114 stored in advance in the storage unit 110. is extracted (step 102). The second entity extraction unit 125b outputs the extraction result to the determination unit 126 when the entity can be extracted. Specifically, in step S102, when the data shown in FIG. 7A is input, data that matches the data registered in the learning model 114 is extracted. In this example, as described above, the learning model 114 registers the product name "XXXX" and the model number "V1" to be traded, but the product name "YYYY" and the model number "400" are not registered. Therefore, the extraction result in step S102 includes data corresponding to the product name "XXXX" and model number "V1" as shown in FIG. 7B-2, but the product name "YYYY" and model number Data corresponding to "400" will not be included.

Next, the determination unit 126 checks whether or not extraction results have been received from the first entity extraction unit 125a and the second entity extraction unit 125b (step S103). When the determination unit 126 confirms that the extraction results have been received from both the first entity extraction unit 125a and the second entity extraction unit 125b (step S103; received from both), the first entity extraction unit 125a and the second entity extraction unit 125a It is compared whether the extraction result of the part 125b is the same (step S104).

If the determination unit 126 determines that the extraction results of the first entity extraction unit 125a and the second entity extraction unit 125b are the same (step S104; Yes), it determines that each extraction result is valid (step S105). , the extraction result is output to the semantic analysis unit 122 (step S106).

If the determination unit 126 determines in step S104 that the extraction results of the first entity extraction unit 125a and the second entity extraction unit 125b are different (step S104; No), the first entity extraction unit 125a and the second entity extraction unit 125a Each extraction result of the unit 125 b is output to the semantic analysis unit 122 . In this example, as shown in FIGS. 7(B-1) and (B-2), since the extraction results of the first entity extraction unit 125a and the second entity extraction unit 125b are different, the respective extraction results are semantically analyzed. It will be output to the unit 122 . The semantic analysis unit 122 compares the entities included in each extraction result with the trade ontology 115, and calculates the total score value of each extraction result of the first entity extraction unit 125a and the second entity extraction unit 125b (step S107). .

Specifically, in step S107, the semantic analysis unit 122 performs a process of assigning each entity extracted by the first entity extraction unit 125a and the second entity extraction unit 125b to a class of the trade ontology 115. For example, using a deep learning model learned based on predetermined rules, such as the numeric data on the right side of "QTY", which is a character string representing quantity, is the "value" of the "quantity" class, Assign each data to each class. A score value is set for each class, and the semantic analysis unit 122 adds the score value of the assigned class by assigning the entity to the class.

For example, as shown in FIG. 9, the semantic analysis unit 122 assigns "YYYY" to the "product name" class of the trade ontology 115 from the extraction result of FIG. is assigned to the "model number" class and added by 3 points, and "2004" is assigned to the quantity "value" class and added by 3 points. The semantic analysis unit 122 calculates the total score value of the extraction result of the first entity extraction unit 125a by adding all the score values of the classes to which the entities are assigned. The semantic analysis unit 122 performs similar processing on the extraction result of the second entity extraction unit 125b, and calculates the total score value of the extraction result of the second entity extraction unit 125b. The semantic analysis unit 122 outputs the calculated total score value for the extraction results of the first entity extraction unit 125 a and the second entity extraction unit 125 b to the determination unit 126 . In this example, the product name "YYYY" and the model number "400" are included in the extraction result of the first entity extraction unit 125a, but are not included in the extraction result of the second entity extraction unit 125b. Therefore, the total score value of the extraction result of the first entity extraction unit 125a is higher than that of the extraction result of the second entity extraction unit 125b.

Returning to FIG. 8 , after executing the process of step S<b>107 , the determination unit 126 receives the total score value of each extraction result from the semantic analysis unit 122 . Next, the determination unit 126 compares whether the total score value of each extraction result received is equal to or greater than a predetermined threshold, and determines whether the total score value of any extraction result is equal to or greater than the threshold ( step S108). When determining that the total score value of any of the extraction results is equal to or greater than the threshold (step S108; Yes), the determination unit 126 determines that the extraction result with the high total score is appropriate (step S109). Then, the determination unit 126 outputs the extraction result determined to be valid to the semantic analysis unit 122 (step S106). In this example, the determination unit 126 determines that the extraction result of the first entity extraction unit 125a with a high total score value is valid (step S109), and sends the extraction result of the first entity extraction unit 125a to the semantic analysis unit 122. Output (step S106).

Returning to step S108, when the determination unit 126 determines that both the extraction results of the first entity extraction unit 125a and the second entity extraction unit 125b are below the threshold (step S108; No), for example, the rule base The model 113 is updated, the threshold is lowered, and the extraction rule and threshold are relaxed (step S113), the process returns to steps 101 and 102, and the process is executed again.

Returning to step S103, when the determination unit 126 determines that the extraction result has been received from only one of the first entity extraction unit 125a and the second entity extraction unit 125b (step S103; received from one), the received extraction result to the semantic analysis unit 122 . The semantic analysis unit 122 compares the entities included in the received extraction results with the trade ontology 115 to calculate a total score value (step S110). The total score value is calculated by the same method as in step S107 described above. Upon receiving the total score value of the extraction result from the semantic analysis unit 122, the determination unit 126 determines whether or not the received total score value is equal to or greater than a threshold value predetermined by the user (step S111).

If the determination unit 126 determines that the total score value is equal to or greater than the predetermined threshold value (step S111; Yes), it determines that the extraction result is valid (step S112), and outputs the extraction result to the semantic analysis unit 122. (step S106).

In step S111, when the determination unit 126 determines that the total score value is less than the predetermined threshold (step S111; No), it relaxes the extraction rule and the threshold (step S113), and performs steps 101 and 102. Go back and try again.

Returning to step S103, when the determination unit 126 determines that the extraction result has not been received from either the first entity extraction unit 125a or the second entity extraction unit 125b (step S103; not received), the determination unit 126 outputs no extraction result. and terminates the process (step S114). If there is no extraction result, the processing of steps S14 and S15 in FIG. 6 is not executed, and an output to the effect that an abnormality has occurred is output to the terminal 200 as error processing.

As described above, the semantic analysis device 100 extracts entities from the unstructured data in the L/C data, assigns meaning to the extracted entities based on the trade ontology 115, and generates structured data. can be done.

(Modification)
Although the embodiments of the present invention have been described above, the present invention is not limited to the above embodiments, and various modifications and applications are possible. For example, the semantic analysis device 100 does not have to have all the technical features shown in the above embodiments. It may have a configuration of parts. Moreover, at least a part of each of the following modified examples may be combined.

In the above embodiment, the output unit 123 generates structured data in the process of step S15 after the process of step S14 of FIG. 6 by the semantic analysis unit 122, but the present invention is not limited to this. As mentioned above, L/C data includes both structured and unstructured data. Therefore, for example, when structured data includes "total amount" and its amount data, whether or not the amount data of the analysis result by the semantic analysis unit 122 matches the amount data of the structured data is checked. A confirmation process may be included before the process of step S15. That is, for information important in trade transactions such as amount data, the correctness of the processing of step S14 by the semantic analysis unit 122 may be verified by comparing it with the structured data included in the L/C data. . Then, for example, when the data assigned as "total amount" in the process of step S14 matches the total amount indicated by the structured data included in the L/C data, the process of step S15 is executed. If they do not match, the fact that an abnormality has occurred may be output to the terminal 200 together with the content of the abnormality. Note that the above-described confirmation processing is not limited to important information in trade transactions, and may be executed for all analysis results by the semantic analysis unit 122. Information related to L/C settlement may be checked for information important to the user. A degree (priority) may be defined in advance so that only information with a high degree of importance is executed.

In the above embodiment, in step S109 of FIG. 8, the total score value, which is the sum of the score values of all the entities extracted by the first entity extraction unit 125a, and the total score value of all the entities extracted by the second entity extraction unit 125b Although the example of comparing the total score value, which is the sum of the score values (that is, the example of comparing the total score values), this is just an example. In addition, for example, the score value of one entity extracted by the first entity extraction unit 125a is compared with the score value of one entity extracted by the second entity extraction unit 125b corresponding to the entity. Score values may be compared for each entity, and an entity with a high score value may be determined as a valid entity. Then, combine valid entities to generate structured data. According to this, since validity is determined for each entity, structured data with higher accuracy can be generated.

In the above embodiment, trade ontology 115 is an L/C ontology, but it is not limited to this. Trade ontology 115 may include an ontology corresponding to required documents in addition to L/C. Specifically, I/V (Invoice), P/L (Packing List), S/I (Shipping Instruction), I/P (Insurance Policy), B/L (Bill of Lading), B/E (Bill of Exchange), S/A (Shipping Application), A/N (Application for Negotiation of Documentary Bill), and the like, and may further include ontologies for each required document to define the class relationships between the documents. FIG. 10 shows an example where trade ontology 115 comprises an I/V ontology and an L/C ontology. As shown in FIG. 10(A), the "Order Detail" class of the L/C ontology and the "Goods Detail" class of the I/V ontology have the same meaning according to the relationship information of "sameAs" indicating that the classes have the same meaning. be related. As a result, information classified as "Order Detail" in L/C can be classified as "Goods Detail" in I/V. From the structured L/C data shown in 6(C), the I/V data shown in FIG. 10(B) can be created. In this case, it is sufficient that the I/V master format 112 is stored in the storage unit 110 . Based on the trade ontology 115 including the relationship information between the L/C ontology and the I/V ontology, the semantic analysis unit 122 generates structured data such as "YYYY", which is the classification of the "Product Name" shown in FIG. 6(C). The information of the L / C data obtained is assigned to "name" which is a class under the "Goods" class of the I / V ontology, and the output unit 123 writes the assigned information to the I / V master format 112 generates I/V data (“Goods name” shown in FIG. 10B). According to this, the necessary documents can be easily generated from the generated structured data, and the procedure for L/C settlement can be facilitated. The L/C data is an example of trade transaction document data in the scope of claims, and the I/V data is an example of necessary trade data in the scope of claims. The L/C ontology is an example of a letter of credit analysis ontology in the scope of claims, and the I/V ontology is an example of a necessary information ontology in the scope of claims.

The trade ontology 115 may also include an individual ontology corresponding to the format of trade transaction document data for each individual company (trader), and relationship information for associating the trade ontology 115 with the class of the individual ontology. For example, an individual ontology is an ontology for I/V and is composed of classes defined by a classification name for each individual company. FIG. 11(A) shows an example in which the trade ontology 115 includes I/V ontologies of companies α and β in addition to the I/V ontology shown in FIG. 10(A). As shown in FIG. 11(A), the "Goods Detail" class of the I/V ontology, the class name "001" of the I/V ontology of company α, and the class name of "Product Detail" of the I/V ontology of company β. are related by the relationship information of "sameAs" indicating that the two are the same. As a result, the information classified as "001" in the I/V data of Company α and the information classified as "Products Detail" in the I/V data of Company β are classified as "Goods Detail" in the I/V ontology. can be defined to have the same meaning as the information provided. As a result, when the formats of the I/V data of company α and company β are different, as shown in FIG. 11B, I/V data matching the format of company β is generated from the I/V data of company α. becomes possible. In this case, the individual ontology and related information may be generated by machine learning from the past trade transaction document data for each individual company stored in the document data 111, and the I/V master format for α company and β company 112 is stored in the storage unit 110 . The semantic analysis unit 122 assigns the information of the I/V data of company α to the I/V ontology class of the trade ontology 115 associated with the relationship information. The output unit 123 generates I/V data by writing the assigned information into the I/V master format 112 of the β company. In this way, even if each company uses required documents with different formats and item names, the required documents data of one company can be converted into the format of another company through the trade ontology 115. data can be generated. The I/V ontology is an example of the necessary information basic ontology in the scope of claims, and the I/V ontology of company α and company β is an example of the individual necessary information ontology in the scope of claims.

In the above embodiment, the function of the output unit 123 in the semantic analysis device 100 writes the entity given meaning by the semantic analysis unit 122 to the corresponding classification name in the master format 112. Although an example of creating structured data was shown in , this is just an example. For example, the structured data may not be generated by the semantic analysis device 100 but may be generated by the terminal 200 side. In this case, the master format 112 is stored in the terminal 200, the semantic analysis device 100 outputs (transmits) the entity given meaning by the semantic analysis unit 122 to the terminal 200, and the terminal 200 receives the entity. Entities can be written to the corresponding taxonomy in master format 112 . Alternatively, terminal 200 may generate structured data by displaying the classifications and entities included in the analysis result of semantic analysis unit 122 in association with each other, regardless of master format 112 .

Further, in the above embodiment, an example was shown in which L/C data was used as input data and P/O data was used as correct data. If the semantic analysis unit 122 can derive a solution without using the P/O data by referring to the list as a dictionary, the P/O data need not be input.

Also, in the above embodiment, the L/C is written in English, but it may be written in Japanese or another language.

Also, in the above embodiment, the case where the advising bank and the negotiating bank are the same has been described, but they may be different banks.

Further, in the above embodiment, the trade ontology 115 is represented by a tree structure showing a one-to-many relationship, but it is not limited to this, and may be represented by a structure showing a many-to-many relationship. .

Note that the semantic analysis apparatus 100 and the terminal 200 can be realized using a normal computer without using a dedicated apparatus. For example, the semantic analysis apparatus 100 and the terminal 200 that execute the above processes may be configured by installing the program in the computer from a recording medium that stores the program for executing any of the above processes. Also, a single semantic analysis apparatus 100 or terminal 200 may be configured by a plurality of computers working together.

Further, when the above-described functions are realized by sharing the responsibility between an OS (Operating System) and an application, or by cooperation between the OS and an application, only parts other than the OS may be stored in the medium.

It is also possible to superimpose a program on a carrier wave and distribute it via a communication network. For example, the program may be posted on a bulletin board (BBS, Bulletin Board System) on a communication network and distributed via the network. These programs may be activated and executed in the same manner as other application programs under the control of the operating system so that the above processes can be executed.

1 semantic analysis system, 100 semantic analysis device, 200 terminal, 210 network, 110 storage unit, 111 document data, 112 master format, 113 rule base model, 114 learning model, 115 trade ontology, 121 entity extraction unit, 122 semantic analysis unit , 123 output unit, 124 morphological analysis unit, 125a first entity extraction unit, 125b second entity extraction unit, 126 determination unit, 130 communication unit, 20 control unit, 21 storage unit, 22 communication unit, 23 display unit, 24 input Department

Claims (6)

a morphological analysis unit that divides trade transaction document data into morphemes;
a first entity extraction unit for extracting an entity, which is information necessary for settlement using a letter of credit, from the morphemes divided by the morpheme analysis unit based on a rule-based model predetermined by a user;
a second entity extraction unit that extracts the entity from the morphemes divided by the morphological analysis unit based on a learning model learned from past trade transaction document data;
a determination unit that determines whether or not the extraction result by the first entity extraction unit and the extraction result by the second entity extraction unit satisfy a predetermined criterion;
Analyzing to which classification in settlement using the letter of credit the entity included in the extraction result determined by the determining unit to satisfy the criteria belongs, based on a trade ontology learned according to predetermined conditions. a semantic analysis unit;
a storage unit that stores the entity analyzed by the semantic analysis unit in association with the classification;
an updating unit that performs learning based on the entity stored in the storage unit in association with the classification, and updates the contents of the learning model and the trade ontology;
A semantic analysis device for trade transaction messages. an output unit that outputs the entity stored in the storage unit together with the classification associated with the entity;
The semantic analysis unit assigns a predetermined score value when analyzing which classification the entity belongs to,
The determination unit is
determining that the criterion is satisfied when the extraction result by the first entity extraction unit and the extraction result by the second entity extraction unit match;
When the extraction result by the first entity extraction unit and the extraction result by the second entity extraction unit do not match, the semantic analysis unit extracts entities included in the extraction result by the first entity extraction unit and the second entity extraction unit. Among the score values assigned by analyzing the entities included in the extraction results based on the trade ontology, the extraction results corresponding to the high score values are regarded as the extraction results that satisfy the criteria. judge,
When the extraction result is extracted only from one of the first entity extraction unit and the second entity extraction unit, the entity included in the extracted extraction result is added to the trade ontology If the score value assigned by analyzing based on is equal to or greater than a predetermined threshold value, the extraction result is determined to be an extraction result that satisfies the criterion;
2. The apparatus for semantic analysis of trade transaction messages according to claim 1. The trade ontology is a letter of credit analysis that is a learning model for analyzing to which classification in settlement using the letter of credit an entity included in the extraction results determined by the determination unit to satisfy the criteria belongs. The ontology and the entity whose classification in the settlement using the letter of credit is analyzed by the semantic analysis unit based on the ontology and the letter of credit analysis ontology is the classification in the trade necessary data different from the trade transaction document data. a necessary information ontology, which is a learning model for analyzing whether it belongs to
The semantic analysis unit analyzes, based on the necessary information ontology, which classification in the trade necessary data the entity included in the data stored in the storage unit belongs to,
The storage unit stores the entity analyzed by the semantic analysis unit based on the necessary information ontology in association with the classification of the trade necessary data.
3. The apparatus for semantic analysis of trade transaction messages according to claim 1 or 2. The classification in the trade necessary data is different for each trader,
The necessary information ontology indicates to which classification in the trade necessary data an entity whose classification in the settlement using the letter of credit is analyzed by the semantic analysis unit based on the letter of credit analysis ontology. The basic ontology of necessary information, which is a learning model for analysis, and the entity whose classification in the trade-required data is analyzed by the semantic analysis unit is different for each trader in the trade-required data. a necessary information individual ontology, which is a learning model for analyzing whether it belongs to a classification,
The semantic analysis unit analyzes, based on the necessary information individual ontology, to which class the entity analyzed based on the necessary information basic ontology belongs to, which differs for each trader in the necessary trade data, and
The storage unit stores the entity analyzed by the semantic analysis unit based on the individual necessary information ontology in association with the classification that differs for each trader in the necessary trade data,
4. The apparatus for semantic analysis of trade transaction messages according to claim 3. A semantic analysis method by a semantic analysis device comprising a morphological analysis unit, a first entity extraction unit, a second entity extraction unit, a determination unit, a semantic analysis unit, a storage unit, and an update unit,
a morphological analysis step in which the morphological analysis unit divides the trade transaction document data into morphemes;
The first entity extracting unit, based on a rule-based model predetermined by a user, extracts an entity that is information necessary for payment using a letter of credit from the morphemes divided by the morphological analysis step. an extraction step;
a second entity extraction step in which the second entity extraction unit extracts the entity from the morphemes divided by the morphological analysis step based on a learning model learned from past trade document data;
a determination step in which the determination unit determines whether or not the extraction result of the first entity extraction step and the extraction result of the second entity extraction step satisfy a predetermined criterion;
The semantic analysis unit learns, according to a predetermined condition, to which classification in settlement using the letter of credit the entity included in the extraction result determined to satisfy the criteria in the determination step belongs. a semantic parsing step that parses based on the ontology;
a storage step in which the storage unit stores the entity analyzed by the semantic analysis step in association with the classification;
an updating step in which the updating unit performs learning based on the entities stored in the storing step in association with the classification, and updates the contents of the learning model and the trade ontology;
Semantic analysis method for trade transaction messages including to the computer,
Morphological analysis processing for dividing trade transaction document data into morphemes;
a first entity extraction process for extracting an entity, which is information necessary for settlement using a letter of credit, from the morphemes divided by the morphological analysis process based on a rule-based model predetermined by a user;
a second entity extraction process for extracting the entity from the morphemes divided by the morphological analysis process based on a learning model learned from past trade transaction document data;
a determination process for determining whether or not the extraction result of the first entity extraction process and the extraction result of the second entity extraction process satisfy a predetermined criterion;
Analyzing to which classification in settlement using the letter of credit the entity included in the extraction results determined by the determination processing to satisfy the criteria belongs, based on a trade ontology learned according to predetermined conditions. a semantic analysis process;
a storage process for storing the entity analyzed by the semantic analysis process in association with the classification;
an update process for performing learning based on the entity stored by the storage process in association with the classification, and updating the contents of the learning model and the trade ontology;
program to run.

Images