Method for classifying invoice categories and system therefor

Abstract

The present invention relates to a method and system for classifying the category of a delivered item from an invoice used in a courier service. Specifically, the present invention relates to a method and system that estimates and classifies the category of a delivered product based on information generated during the courier service.

Description

METHOD FOR CLASSIFYING INVOICE CATEGORIES AND SYSTEM THEREFOR

[0001] The present invention relates to a method and system for classifying the category of a delivered item from an invoice used in a courier service. Specifically, the present invention relates to a method and system that estimates and classifies the category of a delivered product using deep learning algorithms based on information generated during the courier service, and further facilitates re-learning of results that require modification among the estimates and classifications made by the deep learning algorithms.

[0002] The entity that operates an e-commerce platform manages a category system to improve the usability of product search / navigation and to establish and evaluate its sales / marketing strategies. Typically, product category information is linked manually during the product registration process by sellers or merchandisers (MDs), but the product category information they determine is often inaccurate or unreliable. There are various reasons for this, among others, including: (1) the sellers may arbitrarily designate categories with high sales potential (such as popular categories or avoiding highly competitive ones) to secure a competitive edge for their products on the e-commerce platform; (2) with the spread of shopping mall solutions that allow products to be registered and sold through a single central system, it may be difficult to fully reflect the different category systems across multiple e-commerce platforms during mass registration and exposure to multiple platforms; and (3) it may be difficult to respond promptly when modifying the existing operating methods during the reorganization of the category system on the e-commerce platform. To this end, e-commerce platforms maintain their own category classification technologies to swiftly reflect their category systems. Although category information is initially collected during the product registration process, additional classification and inspection work is required for a significant number of products at the user service stage. At this point, since it is nearly impossible to classify millions or tens of millions of products manually, deep learning technologies are commonly employed to handle the complexity and diversity of the data.

[0003] Meanwhile, in the 'sales' stage, various information that can help infer the types of products (such as product names, product images, detailed information, etc.) can be utilized. Unlike the e-commerce environment, the logistics data of courier services contains only information necessary for the 'delivery' of products that have already been completed in transactions. This means that the information available for classifying invoice categories is very limited.

[0004] The key information needed to classify products from courier data is the "invoice name" stated on the invoice. Generally, the invoice name is written in a way that allows the recipient to understand only which product has been delivered to them. As a result, it is common for some letters to be obscured, for abstract terms to be used, or even for the information to be omitted altogether. Furthermore, there are limitations in specifying the invoice name even when multiple products are bundled together rather than a single product. Logistics companies recommend entering detailed information about the products for courier services, but since this is not mandatory, there are many restrictions in practice. There are several reasons why courier data may not contain sufficient information for product classification, among others, including: (1) the absence of standard product names due to senders writing arbitrary names, which leads to variations in how product names are recorded even for the same item; (2) the ambiguity in product descriptions resulting from the omission or abstraction of product names when the recipient does not need to check the courier item in advance; and (3) the possibility of product names duplication within each category, as the product name may not necessarily include the category features required to specify the category solely based on the name stated on the invoice.

[0005] The present invention has been proposed in light of these issues, with the aim of providing a method and system for estimating the categories of handled products based on invoices used in courier services and the information that can be obtained in the process where users utilize these courier services.

[0006] Therefore, an object of the present invention is to classify the category of a product as accurately as possible from an invoice from which only little information can be obtained.

[0007] Another object of the present invention is to minimize classification errors while preserving as much usable category information as possible during the classification process.

[0008] Still another object of the present invention is to obtain accumulated data by repeatedly performing the process of estimating product categories from invoice information, and to utilize such data as a basis for improving overall service profitability, enhancing consumer services, and identifying consumer trends.

[0009] Still yet another object of the present invention is to enable the use of data obtained from a large volume of courier shipments to design additional predictive models, such as large-scale product recommendation models.

[0010] Meanwhile, the above-mentioned objects of the present invention are not limited to those mentioned above, and other objects not mentioned will be clearly understood by those skilled in the art from the following description.

[0011] To address the aforementioned issues, an embodiment of the present invention provides a method for classifying invoice categories, the method comprising a step of inferring the category of a product matching an invoice based on invoice information obtained from the input invoice.

[0012] Moreover, in the method for classifying invoice categories of according to the present invention, the step of inferring the category may be performed using an invoice classification model, and the invoice classification model may be trained through a step of training the invoice classification model using multiple pieces of training invoice information before the step of inferring the category.

[0013] Furthermore, in the method for classifying invoice categories of according to the present invention, the step of training the invoice classification model may include: a general language training step of training a general language model to distinguish the general features of product names described on invoices; a product category training step of training a product classification model by utilizing the general language model to extract common features from two or more pieces of invoice information and then producing the common features as result values; and an invoice category training step of training the invoice classification model to produce a classification result value for any given invoice through a comparison calculation with ground-truth classification information using the product classification model.

[0014] In addition, in the method for classifying invoice categories of according to the present invention, the invoice categories may be classified according to a hierarchical category, and the hierarchical category subdivides the features of a specific product from a higher concept to a lower concept. Moreover, the invoice category may include a plurality of hierarchical values, wherein an unknown value is assigned to hierarchies where an inference calculation is impossible.

[0015] Additionally, in the method for classifying invoice categories of according to the present invention, the step of inferring the category may include: a pre-processing step of receiving any invoice information and pre-processing it to remove duplicate invoice information; and an inference calculation step of receiving the pre-processed invoice information and producing a classification result for each piece of invoice information.

[0016] Moreover, the method for classifying invoice categories of according to the present invention may further include, after the inference calculation step, a post-processing step of storing the classification result in association with the invoice information.

[0017] Furthermore, the method for classifying invoice categories of according to the present invention may further include a step of retraining the invoice classification model using the classification result produced by the step of inferring the category.

[0018] In addition, the method for classifying invoice categories of according to the present invention may further include, before the step of retraining the invoice classification model, a step of executing a quality inspection of the invoice classification model based on the classification result produced by the step of inferring the category, and the step of retraining the invoice classification model may utilize the classification result produced from the quality inspection.

[0019] Another embodiment of the present invention provides a calculation device for classifying invoice categories, the calculation device including a central processing unit and a memory, wherein the central processing unit may execute instructions stored in the memory for performing a method for classifying invoice categories, and wherein the method for classifying invoice categories may include a step of inferring the category of a product matching an invoice based on invoice information obtained from the input invoice.

[0020] Still another embodiment of the present invention provides a method for classifying the category of a handled product, the method comprising a step of inferring the category of the handled product that matches basic data using at least one piece of information about the handled product included in the basic data.

[0021] Moreover, in the method for classifying the category of a handled product, the step of inferring the category of the handled product may be performed using a handled product classification model, and the handled product classification model may be trained through a step of training the handled product classification model using multiple pieces of training basic data before the step of inferring the category of the handled product.

[0022] According to the present invention, it is possible to relatively accurately classify the categories of handled products from invoice information.

[0023] Moreover, according to the present invention, it is possible to utilize the collected data for various purposes.

[0024] As an example, the present invention can be used to optimize the overall logistics services in the logistics sector. By estimating what items are being handled by the courier service from invoice information, it is possible to establish a direction for innovation in various consumer services. Accurately estimating the handled items contributes to providing reliable and timely delivery services (such as fast delivery of fresh food and safe delivery of high-value products), which is expected to create a virtuous cycle effect of improving customer satisfaction and enhancing revenue.

[0025] Furthermore, the present invention also can be used to statistically analyze seller and consumer trends. Logistics data (approximately 150 million entries per month) serves as an indicator to identify nationwide e-commerce consumer trends, providing valuable information for understanding market flows and consumer demands by analyzing online consumption trends. The present invention enables the classification results of courier invoices to be utilized in statistical analysis without additional processing. From the manufacturer's perspective, the statistical data obtained through the present invention can be used to plan products preferred by consumers in the market or to assess the brand's competitiveness against competitors. In addition, in the e-commerce industry, it allows for the detection of market signals based on transaction history and consumer responses, enabling the establishment of product strategies by comparing them with their won transaction history. Furthermore, it is expected to be utilized in research and analysis applications related to industrial changes and shift in consumer psychology changes by examining the consumption flow from manufacturers to consumers over time.

[0026] On the other hand, the present invention can also be used to develop additional predictive models. Data obtained through the present invention, such as customer product (group)-specific cargo volume data, can also be used to develop supplementary predictive models, including large-scale product recommendation models. Currently, most product recommendation models embedded in e-commerce platforms rely solely on transaction (click) history from a single platform. However, such models can only analyze data from a single domain, resulting in insights based on restricted information from a narrow perspective, which limits their ability to discover new relationships or patterns. The category classification method and system according to the present invention are designed to accumulate product (group) information from over 1.8 billion invoice records annually. It is expected to contribute to the development of predictive models, such as product recommendations, from a cross-domain perspective by securing high-quality consumer transaction data generated from various companies.

[0027] Meanwhile, the above-mentioned effects of the present invention are not limited to those mentioned above, and other technical effects not mentioned will be clearly understood by those skilled in the art from the following description.

[0028] FIG. 1 conceptually illustrates an invoice category classification method according to the present invention.

[0029] FIG. 2 illustrates the invoice category classification method according to the present invention, divided into two main steps.

[0030] FIG. 3 illustrates the detailed steps of a process for training an invoice classification model, and FIG. 4 illustrates the detailed components of a training unit.

[0031] FIG. 5 illustrates the process of training the product classification model.

[0032] FIG. 6 illustrates the detailed steps of a process for inferring invoice categories, and FIG. 7 illustrates the detailed components of an inference unit.

[0033] FIG. 8 illustrates the modified appearance resulting from the application of the product classification model, trained in the training unit, to the inference unit.

[0034] FIG. 9 illustrates the process of retraining the invoice classification model.

[0035] FIG. 10 illustrates the quality control process by the invoice classification model.

[0036] Details regarding the objects and technical features of the present invention and the resulting effects will be more clearly understood from the following detailed description based on the drawings attached to the specification of the present invention. Preferred embodiments according to the present invention will be described in detail with reference to the attached drawings.

[0037] The embodiments disclosed in this specification should not be construed or used as limiting the scope of the present invention. It is obvious to those skilled in the art that the description, including the embodiments, of this specification has various applications. Therefore, any embodiments described in the detailed description of the present invention are illustrative to better illustrate the present invention and are not intended to limit the scope of the present invention to the embodiments.

[0038] The functional blocks shown in the drawings and described below are only examples of possible implementations. In other implementations, different functional blocks may be used without departing from the spirit and scope of the detailed description. Moreover, although one or more functional blocks of the present invention are shown as individual blocks, one or more of the functional blocks of the present invention may be a combination of various hardware and software components that perform the same function.

[0039] Furthermore, the term "comprising" certain components, which is an "open-ended" term, simply refers to the presence of the corresponding components, and should not be understood as excluding the presence of additional components.

[0040] In addition, if a specific component is referred to as being "connected" or "coupled" to another component, it should be understood that it may be directly connected or coupled to another other component, but there may be other components therebetween.

[0041] FIG. 1 conceptually illustrates an invoice category classification method and overall system according to the present invention.

[0042] Referring to the drawing, it is assumed that the present invention is executed by a calculation device 100, and the system mentioned in this detailed description includes the calculation device 100. As used herein, the term "system" refers to a broader system in which a plurality of terminals are connected by a network, and it is understood that, in some cases, the system may consist only of the calculation device 100.

[0043] Referring again to the drawing, the invoice category classification method according to the present invention is primarily characterized by the calculation device 100 receiving an invoice, performing an inference calculation on it, and then producing a result, specifically a category classification result for the corresponding invoice. At this time, the calculation device 100 receiving an invoice means that it receives data or information that can be obtained from the invoice, which may include, for example, receiving a captured image of the invoice or receiving text (invoice text) extracted from the invoice image. Additionally, the calculation device 100 may receive other supplementary data (additional information) mapped to the invoice, such as the courier (transportation) company identified from the invoice, the sender who sent the courier, the recipient who will receive the courier, the address of the sender or recipient, the place of purchase where the couriered product was bought, or the place of production where the product was manufactured. Such information will be collectively referred to as courier information.

[0044] The result produced by the calculation device 100 can be represented as a hierarchical category as shown in the drawing. In the drawing, a result such as [clothing]-[pants]-[jeans]-[slim fits]-[unknown] is illustrated. A category that classifies a single product by subdividing it from a higher concept to a lower concept, as in this example, will be called a hierarchical category. In other words, the "hierarchical category" system mentioned in this detailed description is understood as a classification system that subdivides product categories from higher concepts to lower concepts, such as [major category-medium category-minor category-subcategory-...]. It is characterized by the fact that products sharing the same higher category exhibit greater classification similarity compared to those that do not.

[0045] Meanwhile, the result produced by the calculation device 100 may explicitly indicate that classification is impossible beyond a certain classification hierarchy. As can be seen from the drawing, when assuming that an invoice is received as input to infer the category of that invoice, the depth of the hierarchy that can be inferred may vary for each invoice. When the calculation device 100 can no longer infer a result, it indicates the value at that hierarchy as "unknown", indicating that the category inference beyond that hierarchy is impossible. However, the categories up to the higher levels of the hierarchy can still be specified through inference. Unlike the classification result shown in FIG. 1, it is understood that for any given invoice, a classification result such as [clothing]-[unknown]-[unknown]-[unknown] may also be produced.

[0046] As described above, the basic concept of the invoice category classification method according to the present invention has been discussed with reference to FIG. 1.

[0047] For reference, in term of hardware, it is assumed that the calculation device 100 includes a central processing unit and memory. In this case, the central processing unit can also be referred to as a controller, microcontroller, microprocessor, microcomputer, etc. Moreover, the central processing unit can be implemented in hardware, firmware, software, or a combination thereof. When it is implemented using hardware, it can take the form of an application specific integrated circuit (ASIC), a digital signal processor (DSP), a digital signal processing device (DSPD), a programmable logic device (PLD), or a field programmable gate array (FPGA), and when it is implemented using firmware or software, the firmware or software can be configured to include modules, procedures, or functions that perform the above-mentioned functions or operations. Furthermore, the memory can be implemented using a memory such as a Read Only Memory (ROM), Random Access Memory (RAM), Erasable Programmable Read-Only Memory (EPROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), flash memory, Static RAM (SRAM), Hard Disk Drive (HDD), or Solid State Drive (SSD).

[0048] In some cases, the calculation device 100 may be implemented as a server, and in this case, the server may be a device that stores and executes a program, specifically a set of instructions, to actually implement the invoice category classification method according to the present invention. The server may take the form of at least one server PC managed by a specific user or may be in the form of a cloud server provided by another company, such as a cloud server that a user can access after registering as a member. Moreover, the calculation device can be implemented in the form of a cloud system without a dedicated server, specifically as a serverless cloud system where multiple distributed calculation devices are individually managed but perform calculations together as a single system.

[0049] Furthermore, in some cases, the invoice category classification method according to the present invention may be executed on a single calculation system consisting of multiple calculation devices rather than just a single calculation device 100. Within this calculation system, multiple calculation devices may be configured to perform different calculations required to execute the prediction method.

[0050] FIG. 2 illustrates the invoice category classification method according to the present invention, divided into two main steps. Referring to FIG. 2, the invoice category classification method may include a step (S100) of training an invoice classification model and a step (S300) of inferring an invoice category.

[0051] The inference of the invoice category is performed by the invoice classification model, which will be described later. This invoice classification model may be an algorithm that can enhance its performance through training. The accuracy of the final result of the inference calculation can be improved based on the degree of training of the invoice classification model. Therefore, the process of training the invoice classification model (algorithm) is crucial to the invoice category classification method according to the present invention. Accordingly, the process of training the invoice classification model will now be described in detail.

[0052] FIGS. 3 and 4 illustrate the process of training the invoice classification model. FIG. 3 illustrate the detailed steps of a process (S100) for training the invoice classification model, and FIG. 4 illustrates the detailed components of a training unit 110 that may exist within the calculation device 100.

[0053] Referring to FIG. 4, the training unit 110 may broadly include a general language training unit 1101, a product category training unit 1102, and an invoice category training unit 1103, and each of these units can be implemented to execute the detailed steps listed in FIG. 3.

[0054] First, the general language training unit 1101 is a detailed component designed to learn data that includes general product names, word meanings, etc., from sources such as news articles, encyclopedias, and blogs, in order to distinguish the general linguistic features of product names described on invoices. It can be implemented to learn that a specific word refers to a specific product, and this general language training unit 1101 may be based on pre-trained algorithms that have been trained on large-scale language data, such as BERT, RoBERTa, and GPT.

[0055] Next, the product category training unit 1102 is a detailed components that collects and learns data containing product names and category classification information from online malls or public datasets. The product category training unit 1102 can learn which product category a specific product name belongs to or which product category includes a product containing a specific brand name. This is because the product category training unit 1102 is designed to learn not only the categories of products but also the hierarchical structure of those categories. For example, the product category training unit 1102 can learn that the product name "cardigan" belongs to the higher category "outerwear", which in turn belongs to the higher category "men's clothing". Alternatively, it can be trained that products containing the brand name "Lev*s" belong to the categories "jeans"-"pants"-"clothing". Furthermore, specific product names (model names) of "Lev*s" jeans, collected through searches, can also be trained in the same manner to determine which specific product category they belong to.

[0056] The product category training unit 1102 can be trained through the steps illustrated in FIG. 5, for example.

[0057] Referring to FIG. 5, in order to reduce classification errors and learn each category hierarchy from the upper levels, the product category training unit 1102 can be repeatedly trained by receiving two different invoice datasets as input and presenting the common parts of the classification categories from each invoice as the ground truth. In this case, the input data (invoice data) corresponds to training data created for the training of the product category training unit 1102 and may include sufficient information to clearly identify the categories of products. Moreover, each pair of input invoice data can be processed to match at the level of the major category, medium category, minor category, or detailed subcategory, as needed. In this learning process, the product category training unit 1102 can be trained to assign the value "unknown" for hierarchies where the common parts are not identified, allowing invoices with ambiguous classifications to be marked as "unknown" starting from a specific classification depth (hierarchy). FIG. 5 illustrate the process where Invoice A, from which information up to [clothing]-[pants]-[jeans] can be obtained, and Invoice B, from which information up to [clothing]-[skirts]-[long skirts] can be obtained, are input. The process includes extracting features for each invoice (S501), aggregating common features among the extracted features (S502), and classifying the categories based on the aggregated common features to produce the common category for the two invoices as a result value (S503). This sequence represents a single cycle in the learning process. In the example of FIG. 5, the only common feature that can be aggregated from Invoice A and Invoice B is [clothing], and thus it can be seen that the final result value produced is [clothing]-[unknown]-[unknown]. The embodiment illustrated in FIG. 5 is intended to explain an example in which a product classification model learns the higher-level category of [clothing]. It depicts how the product category training unit 1102 begins to progressively learn the product category hierarchy of a specific product, starting from the higher category, based on any given invoice data during the initial training stage. The product category training unit 1102 can conduct repeated training based on a large volume of collected training data, allowing it to learn the product category classification for specific products hierarchically. In this process, by adding classification hierarchies, promoting or demoting ranks, or excluding classification hierarchies, a single normalized "hierarchical category classification" can ultimately be defined. Continuing the description of FIG. 5, as the training process, such as that shown in FIG. 5 is repeated, the product category training unit 1102 can be trained to develop a system that can classify product categories, such as [clothing]-[pants]-[jeans]-..., when any two invoice datasets are input.

[0058] For reference, the model that has been trained by the product category training unit 1102 can be applied to the inference unit 120, which will be described later. The inference unit must be implemented to perform calculations to infer the invoice category from a single invoice, in other words, comparison of two invoices is not necessary. Therefore, the structure of the classification model can be modified to classify the category from a single invoice. This will be further discussed in the description of FIG. 8.

[0059] Referring again to FIGS. 3 and 4, the training unit 110 may further include the invoice category training unit 1103. The invoice category training unit 1103 is a detailed component that learns ground-truth classification information, including accurate classification information, such as invoice category classification information that has been directly classified by a person. The invoice category training unit 1103 can also be trained in the same manner as the product category training unit 1102 described above. In other words, the invoice category training unit 1103 can also define a hierarchical category classification system by inputting two (or more) invoice datasets and aggregating the common features, as illustrated in FIG. 5. However, while the invoice data used as input in the product category training unit 1102 is in a completed classification state, the invoice data used as input in the invoice category training unit 1103 can be designed to include only the information that can be obtained from an actual invoice, meaning that it contains only the details necessary for classification based on the actual invoice. For example, the invoice data input into the invoice category training unit 1102 can be designed to include only the information at the level of [pants]-[jeans] or [clothing]. This approach ensures that the invoice classification model is trained to adapt to the environment in which it will be used in practice.

[0060] Meanwhile, the invoice category training unit 1103 may include additional information (such as courier information) in the training process, alongside the product name information stated on the invoice, to enhance the performance of the classification model. This additional information may include the contracted company of the courier service associated with the invoice, the sender, the recipient, the courier's dimensions (width, length, height), etc. The invoice category training unit 1103 can be trained to utilize keywords or sentences derived from this additional information to estimate the category to which the product stated on invoice belongs.

[0061] In FIG. 4, it is illustrated how each of the detailed components of the training unit 110 receives specific information as input and produces corresponding models as output. The general language training unit 1101 receives general language information and creates a general language model, the product category training unit 1102 creates a product classification model using the general language model when product category information is input, and the invoice category training unit 1103 creates an invoice classification model using the product classification model when courier information, including invoice information, is input.

[0062] As described above, the process of training the invoice classification model has been discussed with reference to FIGS. 3 and 4.

[0063] FIGS. 6 and 7 illustrate the process of inferring invoice categories. FIG. 6 illustrates the detailed steps of a process for inferring invoice categories, and FIG. 7 illustrates the detailed components of an inference unit.

[0064] Referring to FIG. 7, the inference unit 120 may include a pre-processing unit 1201, an inference calculation unit 1202, and a post-processing unit 1203. Each of the detailed components can be implemented to execute each of the detailed steps listed in FIG. 6.

[0065] First, the pre-processing unit 1201 is a detailed component that processes data in order to reduce the calculation load of the inference calculation unit 1202 as much as possible. The pre-processing unit 1201 can transform data with identical configurations among the combined information of invoice and courier information, which are the subjects of category classification, into a single input. This prevents the inference calculation for the same input from being repeated. For example, if the combined information of the invoice name and courier information is the same, the pre-processing unit 1201 can assign an identifier to this combined information and store it separately to filter out redundant invoices or courier information as much as possible. As a result, only the combined information with duplicates removed is passed to the inference calculation unit 1202, thereby reducing the calculation load of the inference calculation unit 1202.

[0066] Next, the inference calculation unit 1202 is a detailed component that performs the actual category classification calculation for invoices. The inference calculation unit 1202 utilizes the classification model that was trained by the training unit 110 to produce a classification result, and the classification model used at this time may take a form similar to that shown in FIG. 8. While explaining FIG. 5 earlier, it was noted that the structure of the classification model trained by the training unit 110 differs from that of the classification model used during inference calculations by the inference unit 120. The classification model used in the inference calculations by the inference unit 120 must be implemented to accept a single invoice as input to produce a classification result, and therefore, it can be implemented in the form as shown in FIG. 8. Referring to FIG. 8, it can be seen that the process of extracting the features from invoice B is omitted, and instead, Feature A extracted from Invoice A is input during the aggregation of common features. In this way, the classification model utilized by the inference unit 120 can be implemented to replace the input corresponding to Feature B with Feature A in the classification model described with reference to FIG. 5 (in the actual implementation of the classification model, the value mapped to Feature A is processed in double) to ensure that category classification is performed for a single invoice input. For reference, the classification model in the inference unit 120 can produce a classification result based on information obtained from an invoice. For example, if the information [skirts] is obtained from the invoice, it could produce a classification result such as [clothing]-[women's clothing]-[skirts]-[split skirts]-... based on this information and additional information (courier information). Alternatively, if the information [Lev*s] is obtained from the invoice, it can yield a result value such as [clothing]-[pants]-[jeans]-[unknown]-... based on this. In the latter case, it was possible to infer that the item classified from the information [Lev*s] falls under [jeans], but a more specific classification could not be inferred from the additional information and other details. Therefore, the result indicates that the hierarchy below [jeans] is marked as [unknown].

[0067] Referring again to FIGS. 6 and 7, the inference unit 120 may include a post-processing unit 1203. The post-processing unit 1203 is a detailed component that links the classification result produced by the inference calculation unit 1202 to the combined information (which includes the invoice information (preferably the invoice name) and the courier information that have been matched and stored).

[0068] FIG. 7 illustrates how each detailed component of the inference unit 120 receives specific information as input and produces corresponding outputs. The pre-processing unit 1201 receives the invoice to be classified along with the courier information stored in the database and generates combined information. The inference calculation unit 1202 receives this combined information from the pre-processing unit 1201 and utilizes the invoice classification model trained by the training unit 110 to produce an invoice category classification result, and finally the post-processing unit 1203 links the invoice category classification result to the combined information for storage.

[0069] As described above, the calculation process for inferring the invoice categories has been discussed with reference to FIGS. 6 and 7.

[0070] Meanwhile, in the description of FIGS. 3 and 4 above, the process by which the training unit 110 trains the invoice classification model based on multiple invoice information has been outlined. However, the training unit 110 can also be implemented to perform retraining based on the classification result produced by the inference unit 120.

[0071] FIG. 9 conceptually illustrates an embodiment in which the training unit 110 performs retraining. The training unit 110 can receive the classification result produced by the inference unit 120 as input in the form of invoice information and execute retraining. By applying the retrained classification model to the inference unit 120, invoice category inference can be performed in a more advanced state of performance. Moreover, the classification result produced by the inference unit 120 in this process can also be used for the purpose of inspecting the quality of the invoice classification model (for quality inspection), which will be described with reference to FIG. 10.

[0072] Referring again to FIG. 9, the training unit 110 can be implemented to perform retraining for a certain classification result that requires modification from those produced by the inference unit 120. This retraining may also be carried out by reintroducing the result from the quality inspection process for the classification result.

[0073] FIG. 10 illustrates the quality inspection process by the invoice classification model. The quality inspection process can be carried out by extracting invoice samples and comparing the classification result classified by the previous invoice classification model with the ground-truth classification result (the results classified by a person) to determine if they match. Referring to FIG. 10, the quality inspection process can begin with a step (S10001) of first extracting invoice samples. This extraction can be conducted either by prioritizing invoices with high logistics volume or by considering the sales scale of the logistics, thereby focusing on extracting invoices with larger sales volumes. In addition, in this stage, it is also possible to allow a user (person) to add random invoice samples for use in quality inspection.

[0074] After the step of extracting invoice samples, a step (S10003) of extracting the classification confidence level of the invoice category can be executed. The classification confidence level refers to the level of certainty predicted by the classification model regarding the accuracy of a particular classification value during the classification process. For example, for an invoice sample containing the information [back-bending cotton split long skirts], it can be observed that the confidence levels can be extracted in the following order: [childbirth / childcare]-[maternity wear]-[skirts], [clothing]-[skirts]-[long skirts], etc. (the actual confidence values are not indicated in the drawing). For reference, the classification confidence levels can be in the form of probability values where the sum of the classification confidence levels for all categories equals 1.0. In cases where prediction is impossible (e.g., [unknown]-[unknown]-[unknown]-[unknown]), this value can be calculated as the probability value excluding the confidence levels of the top nine categories.

[0075] After the step of extracting the classification confidence levels, a step (S10005) of inspecting the invoice category classification can be executed. This step can be carried out by receiving input from a user (inspector) and can be implemented in a manner that allows the inspector to select the most appropriate category from their perspective.

[0076] Meanwhile, through these processes, a classification result can be produced (S10007) for the invoice sample, and the classification result produced in this way can be input again as retraining data to the training unit 110.

[0077] Meanwhile, the main steps of the invoice category classification method according to the present invention are not only used for classifying invoice categories in courier services, but can also be applied in various industrial fields that require category classification for handled products, such as the food distribution sector or the e-commerce field, including online market services.

[0078] For example, in the service areas that distribute food products or in the e-commerce sector, there are many instances where the product classifications registered by companies are not accurate. For reasons similar to those mentioned in the background of the invention, it would often be difficult to classify the categories of handled products. However, by utilizing the category classification method through the classification model according to the present invention, as well as the method for training such a classification model, the classification of products registered by various companies can be performed rapidly and in real time. In addition, if the result of classifying products using the classification model requires some modifications, the classification result can be input back into the training unit after undergoing the process of inspecting for the classification result, as described with reference to FIGS. 9 and 10, thereby enhancing the performance of the algorithm, that is, the classification model.

[0079] As described above, the method and system for classifying invoice categories according to the present invention have been discussed. Meanwhile, the present invention is not limited to the specific embodiments and applications described above, and various modifications can be made by those skilled in the art without departing from the gist of the present invention as claimed in the claims. These modified implementations should not be understood as being separate from the technical spirit or scope of the present invention.

Claims

1.A method for classifying invoice categories, the method comprising:a step of inferring the category of a product matching an invoice based on invoice information obtained from the input invoice.2.The method for classifying invoice categories of claim 1, wherein the step of inferring the category is performed using an invoice classification model, andwherein the invoice classification model is trained through a step of training the invoice classification model using multiple pieces of training invoice information before the step of inferring the category.3.The method for classifying invoice categories of claim 2, wherein the step of training the invoice classification model comprises:a general language training step of training a general language model to distinguish the general features of product names described on invoices;a product category training step of training a product classification model by utilizing the general language model to extract common features from two or more pieces of invoice information and then producing the common features as result values; andan invoice category training step of training the invoice classification model to produce a classification result value for any given invoice through a comparison calculation with ground-truth classification information using the product classification model.4.The method for classifying invoice categories of claim 1, wherein the invoice categories are classified according to a hierarchical category, andwherein the hierarchical category subdivides the features of a specific product from a higher concept to a lower concept.5.The method for classifying invoice categories of claim 4, wherein the invoice category includes a plurality of hierarchical values, and wherein an unknown value is assigned to hierarchies where an inference calculation is impossible.6.The method for classifying invoice categories of claim 1, wherein the step of inferring the category comprises:a pre-processing step of receiving any invoice information and pre-processing it to remove duplicate invoice information; andan inference calculation step of receiving the pre-processed invoice information and producing a classification result for each piece of invoice information.7.The method for classifying invoice categories of claim 6, further comprising, after the inference calculation step,a post-processing step of storing the classification result in association with the invoice information.8.The method for classifying invoice categories of claim 2, further comprising:a step of retraining the invoice classification model using the classification result produced by the step of inferring the category.9.The method for classifying invoice categories of claim 8, further comprising, before the step of retraining the invoice classification model,a step of executing a quality inspection of the invoice classification model based on the classification result produced by the step of inferring the category,wherein the step of retraining the invoice classification model utilizes the classification result produced from the quality inspection.10.A calculation device for classifying invoice categories, the calculation device comprising a central processing unit and a memory,wherein the central processing unit executes instructions stored in the memory for performing a method for classifying invoice categories,wherein the method for classifying invoice categories comprises:a step of inferring the category of a product matching an invoice based on invoice information obtained from the input invoice.11.A method for classifying the category of a handled product, the method comprising:a step of inferring the category of the handled product that matches basic data using at least one piece of information about the handled product included in the basic data.12.The method for classifying the category of a handled product of claim 11, wherein the step of inferring the category of the handled product is performed using a handled product classification model, andwherein the handled product classification model is trained through a step of training the handled product classification model using multiple pieces of training basic data before the step of inferring the category of the handled product.

Images