Information Processing Method

Abstract

To provide an information processing method which can estimate an account title of a character string expressed in a local language and indicating the account title.SOLUTION: An information processing method includes: estimating an account title, by using a learned model obtained by performing machine learning for a character string expressed in a first language, and the account title assigned as a correct label with respect to the character string; acquiring a sentence object character string which is the character string expressed in a second language different from the first language and indicates the account title in the second language (S2); acquiring a translation result by machine-translating an object character string to the first language (S7); and estimating the account title assigned as the correct label, corresponding to the object character string, by using the translation result and the learned model (S8).SELECTED DRAWING: Figure 5

Description

[Technical Field]

[0001] The present invention relates to an information processing method. [Background technology]

[0002] Companies that submit securities reports and large companies are required by the Companies Act to prepare consolidated financial statements. In consolidated financial statements, settlements are prepared for the entire corporate group, including domestic and overseas subsidiaries and affiliates, and consolidated financial statements such as balance sheets and profit and loss statements for the entire corporate group are prepared. The above consolidated financial statements are prepared by aggregating the individual financial statements of each company, but as a prerequisite for this aggregating process, it is necessary to unify the account items of the individual financial statements with the consolidated account items used in the consolidated financial statements. In recent years, there have been systems that have a correspondence table that associates keywords with account items, and that automatically journalize account items by referring to this table (see Patent Document 1). However, this does not automate the task of reconciling account items in such consolidated accounting, and the task of reconciling account items has been performed manually by workers. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Laid-Open No. 2014-182787 Summary of the Invention [Problem to be solved by the invention]

[0004] In particular, individual financial statements of overseas subsidiaries and affiliated companies may be written in the language of the local country (hereinafter referred to as the local language), making the task of reconciling the above account items extremely cumbersome.

[0005] Therefore, one object of the present invention is to provide an information processing method capable of estimating an account item from a character string that is expressed in a local language and indicates an account item. [Means for solving the problem]

[0006] One aspect of the present invention is an information processing method that causes a computer to estimate an account title using a trained model obtained by machine learning a string expressed in a first language and an account title assigned as a correct label for the string. The information processing method is characterized by: obtaining a target string that is expressed in a second language different from the first language and indicates an account title in the second language; obtaining a translation result of machine translating the target string into the first language; and using the translation result and the trained model to estimate the account title assigned as the correct label that corresponds to the target string. [Effects of the Invention]

[0007] According to the present invention, it is possible to provide an information processing method capable of estimating an account item of a character string that is expressed in a local language and indicates an account item. [Brief explanation of the drawings]

[0008] [Figure 1] FIG. 1 is a block diagram of a transaction processing system according to an embodiment of the present invention. [Figure 2] 1A is a diagram showing an example of reference data, FIG. 1B is a diagram showing a method for obtaining a character string expressed in English that indicates an account title, and FIG. 1C is a diagram showing another method for obtaining reference data. [Figure 3] FIG. 10 is a diagram showing an example of learning data. [Figure 4] FIG. 10 is a schematic diagram showing a display screen of a subject conversion master. [Figure 5] A diagram showing the estimation process for consolidated account items. DETAILED DESCRIPTION OF THE INVENTION

[0009] Hereinafter, an embodiment of the present invention will be described with reference to the drawings.

[0010] (Outline of accounting processing system) First, we will explain the configuration of an accounting processing system 1 according to this embodiment. The accounting processing system 1 is a system that provides users with accounting processing services, including a service for preparing consolidated financial statements, and is equipped with an accounting processing server 2, a translation server 3, and a terminal device 5. The accounting processing server 2, translation server 3, and terminal device 5 are communicably connected via a network 6 such as the Internet, and the above accounting processing service is provided as a so-called cloud service when a user accesses the accounting processing server 2 via the Internet.

[0011] The terminal device 5 is, for example, a user terminal such as a personal computer or mobile terminal, and can access the transaction processing server 2 via a web browser or a dedicated app. The translation server 3 is an external computer system that provides a translation service in the form of a cloud service, and is configured to translate character strings sent from the transaction processing server 2 into a first language (English in this embodiment) and send the translation results to the transaction processing server 2.

[0012] The accounting processing server 2 is an accounting processing system capable of creating consolidated financial statements by importing individual financial statements of domestic and overseas subsidiaries and affiliates, and is equipped with a control unit 10, a storage unit 20, and a communication unit 30. The communication unit 30 is a communication interface for carrying out encrypted communication with the terminal device 5 and the translation server 3 via the network 6.

[0013] The storage unit 20 stores programs to be executed by the control unit 10, various accounting data, etc., and is composed of RAM (Random Access Memory), ROM (Read Only Memory), etc. The storage unit 20 includes an individual accounting data storage unit 21, a consolidated accounting data storage unit 22, a reference data storage unit 23, a learning data storage unit 24, a trained model storage unit 25, and a master storage unit 26. The individual accounting data storage unit 21 stores data necessary for creating consolidated financial statements, such as individual financial statement data and transaction detail data for each group company.

[0014] For example, in the case of data for an overseas subsidiary, the individual financial statement data and transaction detail data are stored as individual financial statement data and transaction detail data expressed in a local language (a second language different from the first language) output from the accounting software used by the overseas subsidiary. Furthermore, as will be described in detail later, when individual financial statement data and transaction detail data expressed in a local language are imported, the individual accounting data storage unit 21 stores translation data, in which the text data of this individual financial statement data and transaction detail data is translated into English by the translation server 3, in association with the import data.

[0015] The consolidated accounting data storage unit 22 stores consolidated financial statement data created based on the individual financial statement data and transaction detail data of each company stored in the individual accounting data storage unit 21.

[0016] The reference data storage unit 23 stores data that serves as the source of learning data when performing machine learning. More specifically, in this embodiment, a large number of data sets are stored, each containing a character string expressed in English (a character string expressed in a first language) and a consolidated account item into which the English character string is classified. Note that the consolidated account item is an account item used when preparing consolidated financial statements and is an account item that is pre-registered in the consolidated account item master in the master storage unit 26. In this embodiment, the consolidated account item is used as a correct answer label in machine learning. In the following description, the account items used in the individual financial statement data and the transaction detail data are referred to as individual account items. As such, the consolidated account items and individual account items have different ranges (groups) of use. For example, if the consolidated account items are considered to be account items in a first group, the individual account items may include account items from multiple different groups, such as account items from a second group, account items from a third group, and so on, for each company.

[0017] The above-mentioned character strings expressed in English can be collected from account item data in publicly available financial statement data. That is, the character strings expressed in English can be account item data from financial statement data published in English. Also, character strings machine-translated into English account item data from financial statement data published in a language other than English can be used.

[0018] For example, FIG. 2(a) shows an example of reference data stored in the reference data storage unit 23. The reference data includes information about English account titles as character strings written in English. The character string "Bank Deposits" in the account title (English) is obtained by machine translating account titles in financial statements published in Vietnamese into English, as shown in FIG. 2(b). That is, the account title (English) in FIG. 2(a) can be considered to represent the result of machine translation. As shown in FIGS. 2(a) and 2(b), the English character string obtained by machine translating Vietnamese account titles does not necessarily accurately represent the account titles in English. In the machine learning of this embodiment, the English character string obtained by machine translation is assigned a correct label representing a Japanese account title as a consolidated account title. In other words, in this embodiment, multiple Vietnamese character strings, each representing a Vietnamese account title, are machine-translated into English. According to the machine learning of this embodiment, a trained model can be generated by assigning a Japanese account title as a correct answer label to each of multiple character strings resulting from machine translation into English. While the above example illustrates training by machine translating character strings indicating Vietnamese account titles into English, the present disclosure should not be limited to this example. For example, data from the summary field in transaction data may be used instead of character strings indicating account titles. Specifically, as shown in FIG. 2(c), the character string "Subway," which is an English translation of the description in the Vietnamese summary field, and the corresponding consolidated account title name "Travel Expenses" may be paired and stored as reference data in the reference data storage unit 23.

[0019] Additionally, the account items of financial statement data published in English and in languages ​​other than English may be automatically collected by web scraping or the like. Furthermore, the account items of web-scraped financial statement data published in languages ​​other than English may be automatically translated by the translation server 3. Furthermore, data on account items previously used by the user can also be used as reference data. For example, public data such as the EDINET account item list can also be used as reference data.

[0020] Learning data generated based on the reference data is stored in the learning data storage unit 24. Figure 3 shows an example of the learning data, which includes information on the features of the individual account items to be classified and the consolidated account items.

[0021] The trained model storage unit 25 stores trained models that are machine-learned using the training data stored in the training data storage unit 24.

[0022] The control unit 10 is composed of a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), etc., and by executing programs stored in the memory unit 20, it functions as an accounting data acquisition unit 11, a translation data generation unit 12, a feature generation unit 13, an account item estimation unit 15, an item conversion master generation unit 16, a data editing unit 17, a machine learning execution unit 18, and a consolidated financial statement generation unit 19 (see Figure 1).

[0023] The accounting data acquisition unit 11 is configured to acquire data such as individual financial statement data and transaction detail data of subsidiaries and affiliated companies. Specifically, the accounting data acquisition unit 11 accepts data such as individual financial statement data and transaction detail data input from the terminal device 5 using CSV data or the like, and stores the data in the individual accounting data storage unit 21. Furthermore, if the individual financial statement data or transaction detail data is input as image data (for example, including PDF data that does not include text data or cannot be read), the accounting data acquisition unit 11 performs OCR processing on the input data and stores the converted text data in the individual accounting data storage unit 21.

[0024] The translation data generation unit 12 determines whether the individual financial statement data and transaction detail data acquired by the accounting data acquisition unit 11 are written in a language other than English. If the individual financial statement data and transaction detail data are written in a language other than English, the translation data generation unit 12 transmits the text data contained in the individual financial statement data and transaction detail data written in a language other than English to the translation server 3 and receives and acquires the text data translated into English (translated data, translation result) from the translation server 3. The acquired translation data is stored in the individual accounting data storage unit 21. Note that the data sent to the translation server 3 does not necessarily need to be all of the text data contained in the individual financial statement data and transaction detail data; it is sufficient if it includes at least data related to individual account items. The translation data may also be acquired using an API (Application Program Interface) provided by the translation server 3.

[0025] The feature generation unit 13 generates features of character strings expressed in English in the reference data storage unit 23 (such as the data in the account title (English) column in Figure 2(a)) and individual account title names expressed in English included in the translation data acquired by the translation data generation unit 12. Specifically, in generating features, the feature generation unit 13 first performs morphological analysis on the character strings expressed in English and divides the character string into words. Next, the feature generation unit 13 calculates the weight of each divided word and sets this calculated weight of each word as the feature of the account title.

[0026] Fig. 3 is a diagram showing an example of training data stored in the training data storage unit 24. The features of character strings expressed in English in the reference data storage unit 23, which are obtained by the feature generation unit 13 described above, are associated with the corresponding consolidated account items, as shown in Fig. 3, and stored as training data in the training data storage unit 24. In addition, the features of individual account item names expressed in English included in the translation data are treated as the features of the individual account items to be classified.

[0027] Any method may be used to calculate the weight of a word, but the weight may be calculated based on the frequency of the word, such as tf-idf (term frequency inverse document frequency), or the weight of a word may be calculated based on the frequency of each word in the entire reference data.

[0028] The account item estimation unit 15 classifies individual account items into specific consolidated account items (i.e., estimates the corresponding consolidated account items) by inputting the features of the individual account items generated by the feature generation unit 13 into the learned model memory unit 25.

[0029] The account conversion master generation unit 16 generates account conversion master data for converting individual account items into consolidated account items based on the consolidated account items estimated by the account item estimation unit 15. The account conversion master data includes information such as the account item code, the input individual account item name, the individual account item name translated into English, the corresponding consolidated account item name, and whether the account item name is registered (see FIG. 4), and is stored in the account conversion master in the master storage unit 26.

[0030] Furthermore, the item conversion master data is not only used to convert individual account items into consolidated account items, but also to allow the user to confirm whether the consolidated account items estimated by the account item estimation unit 15 are correct and to determine the correspondence between the correct individual account items and consolidated account items. Figure 4 is an example of the display screen of the item conversion master, and for consolidated account items that have not yet been registered by the user, that is, consolidated account items that have been estimated by the account item estimation unit 15 but not confirmed by the user, an icon indicating an unregistered state (a cloud in the example of Figure 4) is attached to the beginning of the consolidated account item name.

[0031] The data editing unit 17 registers or modifies the consolidated account items estimated by the account item estimation unit 15. For example, when the user clicks the "Register" icon on the screen in FIG. 4, the data editing unit 17 changes the data indicating whether the account item name is registered from "No" to "Yes," thereby finalizing the correspondence between the individual account item and the consolidated account item. If the user changes the suggested consolidated account item, the data editing unit 17 modifies the correspondence between the individual account item and the consolidated account item. When the consolidated account item is registered, the data editing unit 17 stores a dataset of the registered English account item name and the consolidated account item name in the reference data storage unit 23.

[0032] The machine learning execution unit 18 generates a trained model for classifying individual account items into consolidated account items by executing machine learning based on the training data stored in the training data storage unit 24. Any machine learning algorithm may be used, and examples of such algorithms include SVM (Support Vector Machine) and Random Forest.

[0033] The consolidated financial statement generation unit 19 combines the individual financial statement data and transaction detail data, which have been converted from individual account items to consolidated account items using the item conversion master, to generate a consolidated settlement sheet, consolidated financial statement data, and consolidated transaction detail data, and stores them in the consolidated accounting data storage unit 22.

[0034] (Estimation of consolidated account items) Next, the estimation process for consolidated account items will be explained using the example of importing individual financial statement data and transaction detail data (hereinafter referred to as individual accounting data) expressed in a local language from an overseas subsidiary. As shown in Figure 5, when creating consolidated financial statements, when the individual accounting data of an overseas subsidiary is imported by the user (S1, S2 in Figure 5), the accounting processing server 2 first determines whether the individual financial statement data and transaction detail data are expressed in a language other than English.

[0035] Then, if the individual financial statement data and transaction detail data are expressed in a local language, the accounting processing server 2 sends data including the text of the individual account items in the individual financial statement data and transaction detail data to the translation server 3 (S3).

[0036] When the translation server 3 receives the text data of the individual account items expressed in the local language (S4), it translates the received text data of the individual account items into English (S5) and sends the translated data of the individual account items to the accounting processing server 2 (S6).

[0037] When the accounting processing server 2 receives the data of the individual account items translated into English (S7), it performs morphological analysis on the target character strings that are expressed in English and indicate the individual account items, and generates features of the individual account items from the weights of each word resulting from the morphological analysis.Then, by inputting the obtained features of the individual account items into the trained model, it classifies the individual account items expressed in the local language into specific consolidated account items from the consolidated account items stored in the consolidated account item master (S8).

[0038] When consolidated account items corresponding to the individual account items expressed in the local language are estimated, the accounting processing server 2 registers the correspondence between the individual account items and consolidated account items in the account conversion master (S9). When the account conversion master is created / updated, the user opens the account conversion master and checks whether the correspondence between the suggested individual account items and consolidated account items is correct, and if it is correct, the correspondence between the individual account items and consolidated account items is confirmed, and if it is incorrect, the corresponding consolidated account items are corrected and the correspondence between the individual account items and consolidated account items is confirmed (S10).

[0039] More specifically, taking the example of the account conversion master display screen shown in Figure 4, the user determines whether or not the consolidated account item name prefixed with a cloud symbol is suitable as the consolidated account item name for the original individual account item name, while referring to the account item name translated into English. If it is suitable as the corresponding consolidated account item name, the user clicks the "Register" icon to register it; if it is not suitable, the user clicks the cell for the consolidated account item name to be corrected, selects an appropriate consolidated account item name from the pull-down candidates, corrects it, and then registers it.

[0040] As described above, in this embodiment, when classifying individual account items expressed in a local language into consolidated account items, the individual account items expressed in the local language are translated into English using machine translation as a preprocessing step. Therefore, even if the individual account items are expressed in a local language that is different from English, which is the language used to train the trained model, the trained model can be used to estimate the corresponding consolidated account items.

[0041] In particular, with the recent trend toward globalization, it is not uncommon to consolidate financial statements of subsidiaries and affiliated companies in multiple overseas countries when preparing consolidated financial statements. However, when using the information processing method according to the present embodiment described above, even if the financial statements of such subsidiaries and affiliated companies in multiple overseas countries are expressed in the languages ​​of each country, the same trained model can be used to estimate consolidated account items for individual account items expressed in each local language.

[0042] For example, even when estimating consolidated account items for individual account items expressed in Vietnamese (an example of a second language) different from English (an example of a first language), the corresponding consolidated account items can be estimated using the translation results of a target string expressed in Vietnamese that indicates an individual account item, machine-translated into English, and the trained model. For example, when an English string obtained by machine-translating a target string expressed in Vietnamese that indicates an account item in Vietnamese is input to the trained model, the trained model can output (estimate) one account item (Japanese) that likely corresponds to the target string from among multiple Japanese account items assigned as correct labels during the training phase. Furthermore, even when estimating consolidated account items for individual account items expressed in Thai (an example of a third language different from the first and second languages), the corresponding consolidated account items can be estimated using the translation results of a target string expressed in Thai that indicates an individual account item, machine-translated into English, and the trained model.

[0043] Furthermore, even in the machine learning stage, character strings relating to account items expressed in English that serve as the source of learning data can be machine-translated into English character strings of account items used in financial statements published in languages ​​other than English. This makes it easier to collect learning data for machine learning, and by increasing the amount of learning data, the accuracy of the trained model can be improved.

[0044] In addition, when using this information processing method, the phase of adapting the language used to express individual account items to the language used for machine learning and the phase of estimating consolidated account items are separated. Therefore, as described above, the same trained model is used to convert individual account items into consolidated account items for all group company individual financial statements. Therefore, the results of machine learning used when importing one company's individual financial statements and converting account items can be used when converting individual account items in other group company individual financial statements into consolidated account items, regardless of the language or accounting system used. Note that machine learning by the machine learning execution unit 18 is performed at least once between the import of the previous individual financial statements and the import of the next individual financial statements.

[0045] Furthermore, by selecting English as the target language for machine translation, machine translation can be performed with higher accuracy than when the target language is a language other than English, such as Japanese. This is because translation into English is considered to have the highest accuracy. In particular, when the source language is a minor language, the accuracy of machine translation is high when the target language is English. As a result, the translation accuracy of individual account items and the translation accuracy of character strings related to account items that are expressed in English and form the source of learning data are both high. Furthermore, because English originally has clear boundaries between words, word segmentation is easier than in languages ​​such as Japanese, and this, combined with the fact that English has clear boundaries between words, it is possible to estimate consolidated account items corresponding to individual account items with high accuracy.

[0046] In the above-described embodiment, English has been used as an example of the first language, but the first language may be a language other than English. The first language is preferably a language that the translation engine of the translation server 3 can translate with high accuracy, but may also be, for example, Japanese or German.

[0047] Furthermore, in the above explanation, the consolidated account items stored in the reference data are expressed in Japanese as shown in Figure 2, but any language can be used, and for example, these consolidated account items can be the target language of machine translation (first language, English in this embodiment). In this case, the consolidated account items output from the trained model will also be English in the above example, but if the correspondence between these English consolidated account items and the consolidated account items in the language that is ultimately desired to be displayed on the display screen of the item conversion master is stored in the consolidated account item master, conversion can be performed when the screen is displayed.

[0048] Furthermore, while Vietnamese (an example of a second language) and Thai (an example of a third language) have been given as examples of local languages, individual account items written in any local language can be classified into consolidated account items as long as the language can be machine translated by the translation engine of the translation server 3.

[0049] Furthermore, in the above-described embodiment, classification of individual account items expressed in a local language into consolidated account items has been described as an example, but the present invention is not limited to this, and can also be used, for example, to journalize transaction data written in a local language into a desired group of account items based on the description in the summary column of the transaction data.

[0050] Furthermore, the above-mentioned accounting processing server 2 and translation server 3 are not limited to being a single computer, but may be configured as a combination of multiple computers. Also, the network 6 connecting the accounting processing server 2, translation server, and terminal device 5 does not necessarily have to be the Internet; at least a portion of it may be connected to an intranet. [Explanation of symbols]

[0051] S2: Obtain a character string expressed in a second language that indicates an account item in the second language; S7: Obtain a translation result; S8: Estimate an account item.

Claims

1. An information processing method for causing a computer to estimate an account title using a trained model obtained by machine learning a character string expressed in a first language and an account title assigned as a correct label to the character string, comprising: acquiring a target character string that is expressed in a second language different from the first language and indicates an account item in the second language; obtaining a translation result obtained by machine translating the target character string into the first language; Using the translation result and the trained model, an account item assigned as the correct label corresponding to the target character string is estimated. An information processing method comprising:

2. the first language is English; 2. The information processing method according to claim 1.

3. The account items assigned as the correct labels are account items expressed in Japanese.

2. The information processing method according to claim 1.

4. The translation result is obtained by the computer communicating with a translation server.

2. The information processing method according to claim 1.

5. The character string expressed in the first language is a character string obtained by machine-translating an account item expressed in a language different from the first language into the first language.

2. The information processing method according to claim 1.

6. acquiring another target character string that is expressed in a third language different from the first language and the second language and indicates an account item in the third language; obtaining a translation result obtained by machine translating the other target character string into the first language; using a translation result obtained by machine translating the other target character string into the first language and the trained model, to estimate an account item assigned as the correct label corresponding to the other target character string; 2. The information processing method according to claim 1.

7. The account items assigned as the correct labels estimated using the trained model are consolidated account items used in consolidated financial statements.

6. The information processing method according to claim 1, wherein:

Images