Summary generation system, training data generation system, summary generation method, training data generation method, and program

Abstract

To improve the accuracy of summary generated by a machine learning model.SOLUTION: A target commodity title acquisition unit 102 acquires a target commodity title T3 which is a target object to be summarized. A target focus attribute acquisition unit 104 acquires, for a target commodity which is a commodity indicated by the target commodity title T3, contents FA3 of a plurality of focus attributes FA. A target summary acquisition unit 106 inputs, to a summary model M, the target commodity title T3 and the contents FA3 of the focus attributes FA for the target commodity, to obtain a target summary S3 which is summary of the target commodity title T3 output from the summary model M. The summary model M is generated by learning based on training data P including a learning commodity title T4, contents FA4 of the focus attributes FA for a learning commercial material, and ground-truth data L4.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present invention relates to a summary generation system, a learning data generation system, a summary generation method, a learning data generation method, and a program.

Background Art

[0002] There is a known technique for generating a summary of text indicating information about a commercial product (hereinafter sometimes referred to as "commercial product information text"). As an example, Patent Document 1 discloses a technique for generating a summary of an introduction text of a real estate property based on the appealing points (features) of the property, such as rent and distance from the nearest station.

[0003] Particularly in recent years, techniques for generating a summary of commercial product information text using a machine learning model have been studied. As an example, Patent Document 2 discloses a machine learning model for generating a summary for each viewpoint such as the story, cast, and production of a movie user review.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Patent Document 2

Summary of the Invention

Problems to be Solved by the Invention

[0005] However, in the above prior art, the accuracy of the summary generated by the machine learning model may not be sufficient.

[0006] The present invention has been made in view of the above problems, and one of its objects is to provide a summary generation system, a summary generation method, and a program capable of improving the accuracy of a summary generated by a machine learning model.

[0007] Another objective of the present invention is to provide a learning data generation system, a learning data generation method, and a program that can easily generate learning data used to generate a machine learning model for generating summaries of commercial product information text. [Means for solving the problem]

[0008] A summary generation system according to one embodiment of the present invention is a summary generation system that generates a summary of product information text that shows information about a product, comprising: a target product information text acquisition means for acquiring a target product information text which is the product information text to be summarized; a target attention attribute acquisition means for acquiring the content of one or more attention attributes which are at least a part of a plurality of attributes that the product has, for the target product which is the product indicated by the target product information text; and a target summary acquisition means for inputting the target product information text and the content of the one or more attention attributes for the target product into a summary model and acquiring a target summary which is a summary of the target product information text output from the summary model, wherein the summary model is generated by learning based on learning data which includes a learning product information text which is the product information text for learning; the content of one or more attention attributes for the learning product which is the product indicated by the learning product information text; and ground truth data which is a summary of the learning product information text.

[0009] A learning data generation system according to one embodiment of the present invention is a learning data generation system that receives product information text indicating information about a product as input and generates learning data used to generate a summarization model that outputs a summary of the product information text, and comprises: a learning product information text acquisition means for acquiring a learning product information text which is the product information text for learning; a focus attribute acquisition means for acquiring the content of one or more focus attributes which are at least a part of the multiple attributes that the product has, for the learning product which is the product indicated by the learning product information text; a string extraction means for extracting strings corresponding to the content of one or more focus attributes for the learning product from the learning product information text based on the content of one or more focus attributes for the learning product; and a summary output means for outputting a summary shorter than the learning product information text based on the extracted strings. [Effects of the Invention]

[0010] According to one aspect of the present invention, the accuracy of summaries generated by machine learning models can be improved.

[0011] Furthermore, according to another aspect of the present invention, training data used to generate a machine learning model for generating summaries of product information text can be easily generated. [Brief explanation of the drawing]

[0012] [Figure 1] This figure shows an example of the hardware configuration of the summary system according to an embodiment of the present invention. [Figure 2A] This figure shows an example of a screen displayed on an e-commerce website. [Figure 2B] This figure shows an example of a screen displayed on an e-commerce website. [Figure 3] This figure illustrates the generation of a summary using a summary model according to an embodiment of the present invention. [Figure 4A] This is a diagram illustrating the generation of a summary model. [Figure 4B]This is a diagram for explaining the generation of a summary model. [Figure 4C] This is a diagram for explaining the generation of a summary model. [Figure 5] This is a diagram for explaining the generation of training data. [Figure 6] This is a functional block diagram showing an example of functions realized by a summary generation system according to an embodiment of the present invention. [Figure 7] This is a diagram showing an example of a product database. [Figure 8] This is a functional block diagram showing an example of functions realized by a summary model generation system according to an embodiment of the present invention. [Figure 9] This is a functional block diagram showing an example of functions realized by a training data generation system according to an embodiment of the present invention. [Figure 10] This is a diagram showing an example of an importance database. [Figure 11] This is a flowchart showing an example of a process executed by a summary generation system according to an embodiment of the present invention. [Figure 12] This is a flowchart showing an example of a process executed by a summary model generation system according to an embodiment of the present invention. [Figure 13] This is a flowchart showing an example of a process executed by a training data generation system according to an embodiment of the present invention. [Embodiment for Carrying Out the Invention]

[0013] Hereinafter, an embodiment of the present invention will be described with reference to the drawings. In this embodiment, a case where the summary system 1 is used for summarizing the titles of products displayed on an e-commerce site will be described as an example. Note that products include not only tangible goods but also intangible goods (such as downloadable content such as movies and comics).

[0014] [1. Hardware Configuration of the Summary System According to this Embodiment] Figure 1 shows an example of the hardware configuration of a summarization system 1 according to an embodiment of the present invention. As shown in Figure 1, the summarization system 1 includes a summarization generation system 10, a summarization model generation system 12, and a training data generation system 14. The summarization generation system 10, the summarization model generation system 12, and the training data generation system 14 are interconnected by a network N such as the Internet or a LAN. Although this example illustrates a case where the summarization generation system 10, the summarization model generation system 12, and the training data generation system 14 are each implemented on one or more different computers, the summarization generation system 10, the summarization model generation system 12, and the training data generation system 14 may also be implemented on the same one or more computers.

[0015] The summary generation system 10 is a computer system composed of one or more computers. The summary generation system 10 is a system that generates summaries of product titles. The summary generation system 10 includes a control unit 10a, a storage unit 10b, a communication unit 10c, an operation unit 10d, and a display unit 10e.

[0016] The control unit 10a is a program control device such as a CPU that operates according to a program stored in the memory unit 10b. The memory unit 10b is a memory element such as ROM or RAM or a hard disk drive. The memory unit 10b stores programs that are executed by the control unit 10a. The communication unit 10c is a communication interface such as a network board or a wireless LAN module. The operation unit 10d is an input device such as a keyboard, mouse, or touch panel. The display unit 10e is a display such as a liquid crystal display or an organic EL display.

[0017] The summary model generation system 12 is a computer system composed of one or more computers. The summary model generation system 12 is a system that generates a summary model M (see Figure 3, etc.), which will be described later. The summary model generation system 12 includes a control unit 12a, a storage unit 12b, a communication unit 12c, an operation unit 12d, and a display unit 12e. The physical configurations of the control unit 12a, storage unit 12b, communication unit 12c, operation unit 12d, and display unit 12e may be the same as the physical configurations of the control unit 10a, storage unit 10b, communication unit 10c, operation unit 10d, and display unit 10e, respectively.

[0018] The learning data generation system 14 is a computer system composed of one or more computers. The learning data generation system 14 is a system that generates learning data P used to generate a summary model M. The learning data generation system 14 includes a control unit 14a, a storage unit 14b, a communication unit 14c, an operation unit 14d, and a display unit 14e. The physical configurations of the control unit 14a, storage unit 14b, communication unit 14c, operation unit 14d, and display unit 14e may be the same as the physical configurations of the control unit 10a, storage unit 10b, communication unit 10c, operation unit 10d, and display unit 10e, respectively.

[0019] [2. Overview of the summarization system according to this embodiment] Figures 2A and 2B show examples of screens displayed on an e-commerce site. Screen G2A shown in Figure 2A is an example of a product details page that displays product details. As shown in Figure 2A, the product title T2 is displayed on screen G2A. The product title T2 is a string of characters that shows various information about the product, such as "ABC Company Iroha Laundry Detergent Liquid Type Refill Eco Size (500mL) 6-Pack [Iroha]" shown in Figure 2A. As an example, the product title T2 displayed on screen G2A shows the product brand name ("ABC Company"), product name ("Iroha"), product type ("Laundry Detergent"), product form ("Liquid Type", "Refill"), product size ("Eco Size (500mL)"), and product quantity ("6-Pack"). Note that the string of characters that the product title T2 contains is not limited to the above example, and the product title T2 may contain any string of characters, such as a product description or appeal text devised by the product seller.

[0020] On screen G2A, selecting the "Back" button displayed in the upper left corner transitions the user from screen G2A, which is the product details page, to screen G2B, which is the product list page shown in Figure 2B. In typical e-commerce sites, when keywords related to a product (such as product type, product name, or brand name) are entered into a search box, a product list page is displayed showing a list of products that match those keywords. Screen G2B, shown in Figure 2B, is an example of such a product list page.

[0021] As shown in Figure 2B, screen G2B displays a portion of the product title t2B for each product. The portion of the product title t2B is an abbreviated version of the product title T2 displayed on screen G2A (e.g., "Refill Eco Size (500mL) 6-pack [Iroha]"), as shown in Figure 2B as "ABC Company Iroha Laundry Detergent Liquid Type...".

[0022] Incidentally, the product title T2 shown in Figure 2A was created assuming it would be displayed on a large screen such as a desktop terminal, and is too long (i.e., has too many characters) for the small screens of mobile devices such as smartphones and tablets. As a result, the entire product title T2 does not fit on the product list page displayed on a mobile device screen. Consequently, as shown in Figure 2B, instead of the entire product title T2, a string t2 with part of the product title T2 omitted has to be displayed on the product list page. This situation leads to a decrease in the readability of the product list page on mobile devices, and ultimately to a decrease in usability.

[0023] Therefore, in the field of e-commerce, there is a growing need for technology to create summaries of product titles. If summaries of product titles can be prepared at a length suitable for the small screen of a mobile device, the readability of product list pages on mobile devices can be ensured. Furthermore, if the product title summary concisely expresses the information that the user is looking for, it can motivate the user to access the product details page for each product and consider purchasing that product.

[0024] In light of the above circumstances, the inventors of the present application are considering a technology for generating product title summaries using a machine learning model. Creating product title summaries manually is extremely time-consuming, but using a machine learning model makes it possible to obtain product title summaries easily. As a method for automating text summarization, rule-based methods are conventionally known, but summarization generation using machine learning models has the advantage of producing summaries with more natural expressions compared to such rule-based methods. Therefore, the inventors of the present application decided to consider a new summarization model that accepts a product title as input and outputs a summary of the product title.

[0025] However, the accuracy of summaries generated by machine learning models may not always be sufficient. For example, if certain constraints are imposed on the summaries generated by the machine learning model in order to obtain summaries that satisfy requirements such as including specific information in the product title summary or keeping the product title summary within a certain character limit, it becomes difficult to improve the accuracy of the summaries generated by the machine learning model.

[0026] Therefore, in this embodiment, as shown in Figure 3, the target product title T3, which is the subject of the summary, and the contents FA3 of each of the multiple attention attributes FA (see Figure 7) for the target product indicated by the target product title T3 are input to the summary model M, and a target summary S3, which is a summary of the target product title T3 output from the summary model M, is obtained. Figure 3 is a diagram illustrating the generation of a summary by the summary model M according to an embodiment of the present invention. The multiple attention attributes FA are part of the multiple attributes A (see Figure 7) that the product has. Details of the summary model M will be described later.

[0027] In other words, according to this embodiment, when the summarization model M generates a target summary S3, which is a summary of the target product title T3, the content FA3 of each of the multiple attention attributes FA about the target product is taken into consideration in addition to the target product title T3, thereby improving the accuracy of the generated target summary S3.

[0028] As will be described later, the summarization model M in this embodiment is a large-scale language model. The summarization model M in this embodiment, being a large-scale language model, receives input including the target product title T3 and the content FA3 of each of the multiple attention attributes FA for the target product, as well as an instruction statement I3 that specifies the output content of the large-scale language model and a conditional statement C3 that indicates the conditions that the output of the large-scale language model must satisfy. This further improves the accuracy of the summaries generated by the machine learning model. When the summarization model M is a large-scale language model as in this embodiment, the data input to the summarization model M, including the training data P described later, is sometimes called a prompt.

[0029] [2-1. Method for generating summary models] The method for generating the summary model M shown in Figure 3 will be explained below using Figures 4A to 4C. Figures 4A to 4C are diagrams illustrating the generation of the summary model M. The summary model M is generated by the summary model generation system 12 shown in Figure 1. As shown in Figures 4A and 4B, the summary model generation system 12 according to this embodiment generates the summary model M by retraining a pre-trained large-scale language model m based on training data P so that it accepts a product title as input and outputs a summary of the product title. In other words, the summary model M is a large-scale language model specialized for the task of accepting a product title as input and outputting a summary of the product title, achieved through retraining based on training data P.

[0030] As shown in Figure 4A, the training data P used to generate the summarization model M includes the training product title T4, the content FA4 of each of the multiple attention attributes FA for the training product which is the product indicated by the training product title T4, the ground truth data L4, the instruction statement I4, and the conditional statement C4.

[0031] The learning product title T4 is the product title used for learning. The learning product title T4 is a string that indicates various information about the learning product. The information indicated by the learning product title T4 includes the content A4 of each of the multiple attributes A (see Figure 7) described later about the learning product. In addition to the content A4 of each of the multiple attributes A, the learning product title T4 may also include any string such as an introductory text or appeal text about the learning product. Note that the learning product title refers to the product title used to generate the summary model M.

[0032] Multiple attention attributes FA are a part of multiple attributes A that the product possesses. Each attribute A is a characteristic that multiple products have in common. Multiple attributes A may be, for example, the brand name, product name, product size, product quantity, etc., but are not limited to these examples. Multiple attention attributes FA may be, for example, artificially selected from among the multiple attributes A by the user of the summarization system 1 (in this embodiment, a business operator running an e-commerce site). In this embodiment, as an example, we will explain the case where the multiple attention attributes FA are the product name, product size, and product quantity from among the attributes A listed above.

[0033] The correct answer data L is a summary of the learned product title T4. This correct answer data L was generated by the learning data generation system 14, which will be described later.

[0034] Instruction I is text data that instructs the output content of the large-scale language model. Specifically, instruction I is text data that indicates that the large-scale language model should output a summary of the learned product title T4, based on the learned product title T4, the content of each of the multiple attention attributes FA, and the conditions indicated by conditional statement C. Figure 2A shows an example of the text of instruction I, but the text of instruction I is not limited to this example.

[0035] Conditional statement C is text data that indicates the conditions that the output of the large-scale language model should satisfy. In this embodiment, conditional statement C indicates that the output of the large-scale language model should contain strings corresponding to each of the content FA4 of multiple attention attributes FA about the learned product. Conditional statement C also indicates that the output of the large-scale language model should be a string of a predetermined number of characters or less that is shorter than the learned product title T4.

[0036] Next, the method for generating the summary model M will be explained in more detail using Figures 4B and 4C. Specifically, as shown in Figure 4B, the summary model generation system 12 retrains the pre-trained large-scale language model m by updating multiple parameters to be updated for multiple pre-trained parameters based on the training data P, without updating multiple pre-trained parameters included in the pre-trained large-scale language model m.

[0037] As shown in equation (1) below, the summarization system 1 does not update the pre-trained parameter matrix W, which has multiple pre-trained parameters as elements, but instead updates the difference matrix ΔW relating to the difference between the pre-trained parameter matrix W and the retrained parameter matrix W', which has multiple retrained parameters as elements, based on the training data P, thereby retraining the pre-trained large-scale language model m. In this embodiment, as an example, the number of rows and columns of the retrained parameter matrix W', the pre-trained parameter matrix W, and the difference matrix ΔW are all d. Note that the number of rows and columns of the retrained parameter matrix W', the pre-trained parameter matrix W, and the difference matrix ΔW may be different from each other.

[0038]

number

[0039] The difference matrix ΔW is the product of two low-rank matrices X and Y, as shown in equation (2) below. The two low-rank matrices X and Y each have multiple parameters to be updated as elements. The number of rows and columns of low-rank matrix X is d and r, respectively. Similarly, the number of rows and columns of low-rank matrix Y is r and d, respectively. The low-rank matrices X and Y each have a rank r that is smaller than the number of rows and columns d of the pre-trained parameter matrix W (rank(X)=r, rank(Y)=r). Here, r is sufficiently smaller than d (r< <d)ものとする。

[0040]

number

[0041] When retraining a pre-trained large-scale language model m by directly updating the pre-trained parameter matrix W, the number of parameters to be computed during retraining is d. 2 This results in r. In contrast, in the summarization model generation system 12, which retrains a pre-trained large-scale language model m by updating the low-rank matrices X and Y respectively, the number of parameters included in the low-rank matrices X and Y is rd, so the number of parameters to be computed during retraining is at most 2rd. Here, since r is sufficiently smaller than d, the summarization model generation system 12 can reduce the computational load required for retraining compared to the case where retraining is performed by directly updating the pre-trained parameter matrix W.

[0042] The retraining process performed in the summary model generation system 12 according to this embodiment specifically includes the following processes (1) to (4).

[0043] (1) Calculation of the low-rank matrix As shown in Figure 4B, the summarization model generation system 12 first calculates two low-rank matrices X and Y for the pre-trained parameter matrix W.

[0044] (2) Updating the low-rank matrix Next, the summarization model generation system 12 updates each low-rank matrix X and Y calculated in (1) based on the training data P. At this time, the pre-trained parameter matrix W is not updated. In other words, in process (2), the pre-trained parameters are not updated, and only the multiple parameters to be updated are updated.

[0045] (3) Calculation of the retrained parameter matrix As shown in Figure 4B, the summarization model generation system 12 calculates a retrained parameter matrix W', which has multiple retrained parameters as elements, based on the two updated low-rank matrices X and Y and the pre-trained parameter matrix W. Specifically, the summarization model generation system 12 calculates the retrained parameter matrix W' by adding the product of the two low-rank matrices X and Y (i.e., the difference matrix ΔW) to the pre-trained parameter matrix W (see equation (1)).

[0046] (4) Output of the summary model Finally, the summarization model generation system 12 outputs a large-scale language model containing the retrained parameter matrix W' as the summarization model M.

[0047] Now, in the summary model generation system 12, as shown in Figure 4C, in the process of (1) above, two low-rank matrices X and Y are calculated such that each has a rank corresponding to the content of the training data P.

[0048] Specifically, the summarization model generation system 12 calculates how many conditional statements C4 are included in the training data P, and calculates two low-rank matrices X and Y such that the rank of each matrix increases as the number of calculated conditional statements C4 increases. In this embodiment, as an example, as shown in Figure 4C, if the training data P does not contain any conditional statements C4, the summarization model generation system 12 calculates two low-rank matrices X and Y such that each matrix has a rank of 2. On the other hand, if the training data P contains one or more conditional statements C4, the summarization model generation system 12 calculates two low-rank matrices X and Y such that each matrix has a rank of 3.

[0049] Furthermore, the summarization model generation system 12 determines whether the conditional statement C4 included in the training data P indicates a predetermined condition. If it is determined that the conditional statement C4 indicates a predetermined condition, it calculates two low-rank matrices X and Y such that each has a rank greater than the rank specified for conditions other than the predetermined condition. In this embodiment, an example is given where the predetermined condition is that the output of the large-scale language model should be a string of a predetermined number of characters or less that is shorter than the training product title T4 (character count specification). In this embodiment, as an example, as shown in Figure 4C, if it is determined that the conditional statement C4 indicates a predetermined condition, the summarization model generation system 12 calculates two low-rank matrices X and Y such that each has a rank of 4.

[0050] As explained above, the summarization model generation system 12 calculates two low-rank matrices X and Y such that each has a higher rank as the number of conditional statements C4 increases, or as the conditional statements C4 express predetermined conditions. This allows the system to calculate two low-rank matrices X and Y such that each has a higher rank as the constraints imposed on the output of the large-scale language model become stricter. In other words, the summarization model generation system 12 increases the number of parameters updated during retraining as the constraints imposed on the output of the large-scale language model become stricter. Therefore, the summarization model generation system 12 can generate a summarization model M that can accurately perform high-difficulty summarization.

[0051] [2-2. Method for generating training data] The method for generating the training data P used to create the summary model M will be explained below, using Figure 5 as an example. Figure 5 is a diagram illustrating the generation of the training data P. The training data P is generated by the training data generation system 14 shown in Figure 1.

[0052] As shown in Figure 4A, the training data P used to generate the summarization model M includes the ground truth data L4, i.e., the correct summaries defined for the training product title T4. Here, as shown in Figure 3, the summarization model M accepts input not only of the target product title T3, but also the content FA3, instruction I3, and condition C3 of each of the multiple attention attributes FA for the target product, and outputs summaries corresponding to the content FA3, instruction I3, and condition C3 of each of the multiple attention attributes FA for the target product. In other words, the ground truth data L is not obtained by a simple summarization of the training product title T4, but by summarizing the training product title T4 while considering the content FA4, instruction I4, and condition C4 of each of the multiple attention attributes FA for the training product included in the training data P. Since preparing such ground truth data L is very time-consuming, it is difficult to prepare a large amount of training data P to use for training the summarization model M.

[0053] Therefore, in the learning data generation system 14 according to this embodiment, learning data P is generated as shown in Figure 5. That is, the learning data generation system 14 generates learning data P through the following processes (1) to (5). Note that in Figure 5, "database" is abbreviated as "DB" (the same applies to other figures).

[0054] (1) Acquisition of learning product titles As shown in Figure 5, the learning data generation system 14 first obtains the learning product title T4.

[0055] (2) Obtaining the content of each of the multiple notable attributes of the learning material. The learning data generation system 14 acquires the content of each of the multiple attention attributes FA for the learning product indicated by the learning product title T4. Here, as shown in Figure 5, the learning data generation system 14 acquires the learning product identification number ID4 (see Figure 7), and based on the identification number ID4, acquires the content FA4 of each of the multiple attention attributes FA for the learning product stored in the product database storage unit 140.

[0056] Specifically, the learning data generation system 14 acquires the importance Im (see Figure 10) of each of the multiple attributes A that the product possesses, which are stored in the importance database storage unit 142. Among the multiple attributes A, those whose importance is equal to or greater than a predetermined value are designated as attention attributes FA, and the system acquires the content FA4 of each of the multiple attention attributes FA for the learning product.

[0057] (3) Extraction of strings corresponding to the content of each of the multiple attention attributes of the learning material. The learning data generation system 14 extracts strings corresponding to the content FA4 of each of the multiple attention attributes FA for the learning product from the learning product title T4, based on the content FA4 of each of the multiple attention attributes FA for the learning product.

[0058] (4) Output of summary Then, as shown in Figure 5, the training data generation system 14 outputs ground truth data L4, which is a shorter summary than the training product title T4, based on the string extracted in (3).

[0059] (5) Output of training data Finally, the learning data generation system 14 outputs learning data P which includes the learning product title T4, the content FA4 of each of the multiple attention attributes FA for the learning product, and the correct answer data L4. As shown in Figure 5, the learning data P output by the learning data generation system 14 further includes an instruction statement I4 and a conditional statement C4.

[0060] According to the learning data generation system 14 of this embodiment described above, strings corresponding to the content FA4 of each of the multiple attention attributes FA for the learning product are extracted from the learning product title T4, and correct answer data L4 is output based on these strings. In other words, the learning data generation system 14 makes it easy to obtain correct answer data L4, which is a summary of the learning product title T4 that takes into account the content FA4 of each of the multiple attention attributes FA for the learning product, and thus easy to generate learning data P.

[0061] The details of the summarization system 1 will be explained below using Figures 6 to 13.

[0062] [3. Functions implemented in the summarization system according to this embodiment] The following describes the functions implemented in the summary system 1 according to this embodiment.

[0063] [3-1. Functions implemented by the summary generation system] Figure 6 is a functional block diagram showing an example of the functions realized by the summary generation system 10 according to an embodiment of the present invention. As shown in Figure 6, the summary generation system 10 functionally includes a product database storage unit 100, a target product title acquisition unit 102, a target attention attribute acquisition unit 104, a target summary acquisition unit 106, and a model storage unit 108. The target product title acquisition unit 102, the target attention attribute acquisition unit 104, and the target summary acquisition unit 106 are mainly implemented by the control unit 10a. The product database storage unit 100 and the model storage unit 108 are mainly implemented by the storage unit 10b.

[0064] The product database storage unit 100 stores the product database DB1 as illustrated in Figure 7. Figure 7 is a diagram showing an example of the product database DB1. The product database storage unit 100 stores the identification number ID (identification information) of each of multiple products and the contents of each of the multiple attention attributes FA for that product in association with each other. Specifically, as shown in Figure 7, the product database DB1 stores multiple product data in each record. Each product data has the contents of each of the multiple items, including the identification number ID ("product ID") and multiple attributes A ("brand", "product name", "size", "quantity", etc.) for the product corresponding to the product data. As mentioned above, in this embodiment, the cases of "product name", "size", and "quantity" are illustrated. In this embodiment, as an example, the first record of the product database DB1 shown in Figure 7 is used for processing by the summary generation system 10, and the second record is used for processing by the learning data generation system 14, which will be described later. In other words, in this embodiment, as an example, let's assume that product ID "12345" is the identification number ID4 of the target product, and product ID "12346" is the identification number ID3 of the learning product.

[0065] The target product title acquisition unit 102 acquires the target product title T3 that is the subject of the summary (see Figure 3). The target product title acquisition unit 102 may acquire the target product title T3 by, for example, extracting the target product title T3 displayed on the e-commerce site. If the e-commerce site provides an API for acquiring the product title of each product, the target product title acquisition unit 102 may acquire the target product title T3 from the e-commerce site via the API. The target product title acquisition unit 102 may also acquire the target product title T3 entered by the user via the operation unit 10d.

[0066] The target attention attribute acquisition unit 104 acquires the content FA3 of each of the multiple attention attribute FAs for the target product, which is the product indicated by the target product title T3 (see Figure 3). Specifically, the target attention attribute acquisition unit 104 acquires the identification number ID3 of the target product (see Figure 7), and based on the identification number ID3, acquires the content FA3 of each of the multiple attention attribute FAs for the target product, which are stored in the product database DB1. The target attention attribute acquisition unit 104 may also acquire the content FA3 of each of the multiple attention attribute FAs for the target product by referring to the stored contents of an external device or information storage medium.

[0067] The target summary acquisition unit 106 inputs the target product title T3 and the contents FA3 of each of the multiple attention attributes FA for the target product into the summary model M, and acquires the target summary S3, which is a summary of the target product title T3 output from the summary model M (see Figure 3). In this embodiment, the target summary acquisition unit 106 acquires the target summary S3 based on the summary model M stored in the model storage unit 108 shown in Figure 6. The summary model M is generated by the summary model generation system 12, which will be described later. Therefore, the details of the summary model M will be described later. Note that the summary model M may be stored in an external device or information storage medium. Also, the summary model M may be generated by an external system, etc.

[0068] In this embodiment, the target summary acquisition unit 106 further inputs an instruction statement I3 that instructs the summary model M on the output content of the summary model M, and a condition statement C3 that indicates the conditions that the output of the summary model M should satisfy (see Figure 3). In this embodiment, the condition statement C3 indicates that the output of the large-scale language model should include strings corresponding to the content FA3 of each of the multiple attention attributes FA for the target product. The condition statement C3 also indicates that the output of the large-scale language model should be a string of a predetermined number of characters or less that is shorter than the target product title T3. In other words, in this embodiment, as an example, there are two condition statements C3. Note that the condition statement C3 may indicate any other conditions besides those described above.

[0069] [3-2. Functions implemented by the summary model generation system] Figure 8 is a functional block diagram showing an example of the functions realized in the summary model generation system 12 according to an embodiment of the present invention. As shown in Figure 8, the summary model generation system 12 functionally includes a learning database storage unit 120, a learning data acquisition unit 122, a summary model generation unit 124, and a model storage unit 126. The learning data acquisition unit 122 and the summary model generation unit 124 are mainly implemented by the control unit 12a. The learning database storage unit 120 and the model storage unit 126 are mainly implemented by the storage unit 12b.

[0070] The learning data acquisition unit 122 acquires learning data P (see Figure 4A). In this embodiment, the learning data acquisition unit 122 acquires learning data P by referring to the contents of the learning database storage unit 120. The learning data acquisition unit 122 may also acquire learning data P by referring to the contents of an external device or information storage medium. In this embodiment, the learning data P is generated by the learning data generation system 14, which will be described later. Therefore, the details of the learning data P will be described later. The learning data P may also be generated by another system or manually.

[0071] The summarization model generation unit 124 generates a summarization model M that accepts a product title as input and outputs a summary of the product title by learning based on the training data P (see Figure 4A). In this embodiment, the summarization model generation unit 124 generates the summarization model M by retraining a pre-trained large-scale language model m based on the training data P (see Figure 4B). In this embodiment, as an example, the case where the pre-trained large-scale language model m is LLaMA2 (Large Language Model Meta AI 2) is described. Note that the pre-trained large-scale language model m may be other large-scale language models such as GPT (Generative Pre-trained Transformer) or BERT (Bidirectional Encoder Representations from Transformers).

[0072] Specifically, the summarization model generation unit 124 retrains the pre-trained large-scale language model m by updating multiple parameters to be updated for multiple pre-trained parameters based on the training data P, without updating multiple pre-trained parameters included in the pre-trained large-scale language model m. That is, as shown in Figure 8, the summarization model generation unit 124 includes a low-rank matrix calculation unit 2020, a low-rank matrix update unit 2021, a retrained parameter matrix calculation unit 2022, and a summarization model output unit 2023.

[0073] The low-rank matrix calculation unit 2020 calculates two low-rank matrices X and Y, each having multiple parameters to be updated as elements, and having ranks smaller than the number of rows and columns of the pre-trained parameter matrix W, which is a matrix with multiple pre-trained parameters as elements (see Figure 4B). In this embodiment, where the Transformer-based LLaMA2 is used as the pre-trained large-scale language model m, the pre-trained parameter matrix W may be, for example, at least one of the query weight matrix, key weight matrix and value weight matrix used in the attention mechanism, and the weight matrix of the fully connected layer. As shown in Figure 4B, the low-rank matrix calculation unit 2020 calculates a low-rank matrix X having the same number of rows as the pre-trained parameter matrix W and a rank smaller than the number of rows and columns of the pre-trained parameter matrix W, and a low-rank matrix Y having the same number of columns as the pre-trained parameter matrix W and a rank smaller than the number of rows and columns of the pre-trained parameter matrix W.

[0074] Specifically, the low-rank matrix calculation unit 2020 calculates how many conditional statements C4 (see Figure 4A) that indicate the conditions that the output of the large-scale language model must satisfy are included in the training data P, and calculates two low-rank matrices X and Y such that the rank of each matrix increases with the number of calculated conditional statements C4 (see Figure 4C).

[0075] Furthermore, the low-rank matrix calculation unit 2020 determines whether the condition statement C4 indicates a predetermined condition, and if it determines that the condition statement C4 indicates a predetermined condition, it calculates two low-rank matrices X and Y such that each has a rank greater than the rank specified for conditions other than the predetermined condition (see Figure 4C).

[0076] The low-rank matrix update unit 2021 updates each low-rank matrix X,Y based on the training data P without updating the pre-trained parameter matrix W (see Figure 4B). Specifically, the low-rank matrix update unit 2021 updates each low-rank matrix X,Y based on the output from the large-scale language model when the training data P is input to the large-scale language model which includes the pre-trained parameter matrix W and the low-rank matrices X,Y, and the ground truth data L included in the training data P. More specifically, the low-rank matrix update unit 2021 updates each low-rank matrix X,Y in such a way that the loss based on the output from the large-scale language model when the training data P is input to the large-scale language model which includes the pre-trained parameter matrix W and the low-rank matrices X,Y, and the ground truth data L included in the training data P is minimized. Known methods such as backpropagation and gradient descent may be used to adjust the parameters based on the loss.

[0077] The retrained parameter matrix calculation unit 2022 calculates a retrained parameter matrix W', which has multiple retrained parameters as elements, based on the two low-rank matrices X and Y updated by the low-rank matrix update unit 2021 and the pre-trained parameter matrix W (see Figure 4B). Specifically, the retrained parameter matrix calculation unit 2022 calculates the retrained parameter matrix W' by adding the product of the two low-rank matrices X and Y to the pre-trained parameter matrix W.

[0078] The summarization model output unit 2023 outputs a large-scale language model containing the retrained parameter matrix W' as the summarization model M (see Figure 4B). Specifically, the summarization model output unit 2023 outputs the summarization model M by replacing the pretrained parameter matrix W with the retrained parameter matrix W' in the pretrained large-scale language model m.

[0079] In this embodiment, the pre-trained large-scale language model m and the generated summarization model M are stored in the model storage units 126 and 108, but these models may be stored in an external device or information storage medium.

[0080] Furthermore, the low-rank matrix calculation unit 2020, the low-rank matrix update unit 2021, the retrained parameter matrix calculation unit 2022, and the summary model output unit 2023 described above can utilize LoRA, which is a known technology. That is, LoRA may be used as a method for retraining a pre-trained large-scale language model m. Note that the retraining method is not limited to LoRA, and may also be QLoRA (Quantized LoRA) or LoHA (LoRA with Hadamerd Product Representation), which are known as improved versions of LoRA.

[0081] [3-3. Functions implemented in the training data generation system] Figure 9 is a functional block diagram showing an example of the functions realized in the learning data generation system 14 according to an embodiment of the present invention. As shown in Figure 9, the learning data generation system 14 functionally includes a product database storage unit 140, a learning product title acquisition unit 141, an importance database storage unit 142, an importance acquisition unit 143, a focus attribute acquisition unit 144, a string extraction unit 145, a summary output unit 146, a learning data output unit 147, and a learning database storage unit 148. The learning product title acquisition unit 141, the importance acquisition unit 143, the focus attribute acquisition unit 144, the string extraction unit 145, the summary output unit 146, and the learning data output unit 147 are mainly implemented by the control unit 14a. The product database storage unit 140, the importance database storage unit 142, and the learning database storage unit 148 are mainly implemented by the storage unit 14b. Details of the product database storage unit 140 are the same as those of the product database storage unit 100 of the summary generation system 10, so a description is omitted.

[0082] The learning product title acquisition unit 141 acquires the learning product title T4, which is the title of the product to be studied (see Figure 5). The learning product title acquisition unit 141 may acquire the learning product title T4 by, for example, extracting the learning product title T4 displayed on an e-commerce site. If the e-commerce site provides an API for acquiring the product title of each product, the learning product title acquisition unit 141 may acquire the learning product title T4 from the e-commerce site via the API. The learning product title acquisition unit 141 may also acquire the learning product title T4 entered by the user via the operation unit 14d.

[0083] The importance database storage unit 142 stores the importance database DB2 shown in Figure 10. Figure 10 is a diagram showing an example of the importance database DB2. As shown in Figure 10, the importance database DB2 stores each of multiple attributes A in association with the importance Im of that attribute. Figure 7 shows, as an example, the case where the importance Im of the attribute "brand name" is "B", and the importance Im of the attributes "product name", "size", and "quantity" are "A". The importance Im of each attribute A may be set artificially, for example, by the operator of the e-commerce site, which is a user of the summarization system 1. In this embodiment, the importance Im is one of "A", "B", and "C", with "A" being the largest, "B" being the second largest, and "C" being the third largest. The importance Im is not limited to discrete values ​​as shown in Figure 10, but may also be continuous values. Furthermore, the notation format of the importance Im is not limited to the example shown in Figure 10.

[0084] The importance acquisition unit 143 acquires the importance Im for each of the multiple attributes A. In this embodiment, the importance acquisition unit 143 acquires the importance Im for each of the multiple attributes A by referring to the contents of the importance database storage unit 142. Alternatively, the importance acquisition unit 143 may acquire the importance Im for each of the multiple attributes A by referring to the contents of an external device or information storage medium.

[0085] The attention attribute acquisition unit 144 acquires the content FA4 of each of the multiple attention attributes FA for the learning product indicated by the learning product title T4. Specifically, the attention attribute acquisition unit 144 acquires the learning product identification number ID4 (see Figure 7) and, based on the identification number ID4, acquires the content FA4 of each of the multiple attention attributes FA for the learning product stored in the product database storage unit 140. The attention attribute acquisition unit 144 may also acquire the content FA4 of each of the multiple attention attributes FA for the learning product by referring to the stored contents of an external device or information storage medium.

[0086] The attention attribute acquisition unit 144 selects multiple attention attribute FAs from among multiple attribute A whose importance Im is equal to or greater than a predetermined value, and acquires the content of each of the multiple attention attribute FAs for the learning product which is the product indicated by the learning product title T4. In this embodiment, as an example, the predetermined value is assumed to be "A" (see Figure 10). In this embodiment, the attention attribute acquisition unit 144 selects multiple attention attribute FAs from among multiple attribute A whose importance Im is "A" or greater for "product name", "size", and "quantity", and acquires the content of each of the "product name", "size", and "quantity" for the learning product.

[0087] The string extraction unit 145 extracts strings corresponding to the content FA4 of each of the multiple focus attributes FA for the learning product from the learning product title T4, based on the content FA4 of each of the multiple focus attributes FA for the learning product. For example, the string extraction unit 145 may decompose the learning product title T4 into one or more tokens using a known lexical analyzer and obtain tokens whose similarity to the content FA4 of each of the multiple focus attributes FA for the learning product is equal to or greater than a predetermined value, thereby obtaining strings corresponding to the content FA4 of each of the multiple focus attributes FA for the learning product. Here, a known similarity calculation method may be used to calculate the similarity between the content FA4 of each of the multiple focus attributes FA for the learning product and each token.

[0088] The summary output unit 146 outputs correct answer data L4, which is a shorter summary than the learning product title T4, based on the strings corresponding to the content FA4 of each of the multiple attention attribute FAs extracted by the string extraction unit 145. For example, the summary output unit 146 may output correct answer data L4 by arranging the extracted strings corresponding to each token according to the order in which the tokens are arranged in the learning product title T4. Alternatively, the summary output unit 146 may output correct answer data L4 by arranging the strings corresponding to the content FA4 of each of the multiple attention attribute FAs for the learning product according to the importance Im of each attention attribute FA.

[0089] The learning data output unit 147 outputs learning data P which includes the learning product title T4, the content FA4 of each of the multiple attention attributes FA for the learning product, and the correct answer data L4 output by the summary output unit 146. The learning data P output by the learning data output unit 147 is stored in the learning database storage units 148 and 120. The learning data P may also be stored in an external device or information storage medium.

[0090] In this embodiment, the learning data output unit 147 outputs learning data P which further includes an instruction statement I4 that specifies the output content of the large-scale language model and a condition statement C4 that indicates the conditions that the output of the large-scale language model should satisfy. In this embodiment, the condition statement C4 indicates that the output of the large-scale language model should include a string corresponding to the content FA4 of each of the multiple attention attributes FA for the learning product. The condition statement C4 also indicates that the output of the large-scale language model should be a string of a predetermined number of characters or less that is shorter than the learning product title T4. In other words, in this embodiment, as an example, there are two condition statements C4. Note that the condition statement C4 may indicate any other conditions besides those described above.

[0091] [4. Processes executed by the summary system according to this embodiment] The following describes the processes performed by the summary system 1 according to this embodiment, using Figures 11 to 13. The processes shown in Figures 11 to 13 are performed by the control units 10a, 12a, and 14a executing programs stored in the storage units 10b, 12b, and 14b, respectively.

[0092] [4-1. Processes performed by the summary generation system] Figure 11 is a flowchart showing an example of processing performed by a summary generation system 10 according to an embodiment of the present invention. As shown in Figure 11, the summary generation system 10 first obtains the target product title T3 (S1100). Next, the summary generation system 10 obtains the content FA3 of each of the multiple attention attributes FA for the target product (S1101). Finally, the summary generation system 10 inputs the target product title T3, the content FA3 of each of the multiple attention attributes FA for the target product, an instruction statement I3 that specifies the output content of the summary model M, and a condition statement C3 that indicates the conditions that the output of the summary model M should satisfy into the summary model M, obtains a target summary S3 which is a summary of the target product title T3 output from the summary model M (S1102), and terminates this process.

[0093] Note that the order of the process for obtaining the target product title T3 (S1100) and the process for obtaining the content FA3 of each of the multiple attention attributes FA for the target product (S1101) may be reversed from the order shown in Figure 11.

[0094] [4-2. Processes performed by the summary model generation system] Figure 12 is a flowchart showing an example of the processing performed by the summarization model generation system 12 according to an embodiment of the present invention. As shown in Figure 12, the summarization model generation system 12 first acquires training data P (S1200) and calculates how many conditional statements C4 are included in the training data P (S1201). In this embodiment, for the sake of explanation, as an example, let's assume that the number of conditional statements C4 calculated in S1201 is 2. Also, in this embodiment, if the number of conditional statements C4 is 2 and the conditional statements C4 do not represent a predetermined condition described later, let's assume that the ranks of the low-rank matrices X and Y described later are set to 3.

[0095] In the following S1202, the summarization model generation system 12 determines whether each conditional statement C4 included in the training data P represents a predetermined condition. If it is determined that a conditional statement C4 represents a predetermined condition (S1202; Y), the summarization model generation system 12 calculates two low-rank matrices X and Y from the pre-trained parameter matrix W, each having a rank of 3 (S1203). If it is determined that a conditional statement C4 does not represent a predetermined condition (S1202; N), the summarization model generation system 12 calculates two low-rank matrices X and Y from the pre-trained parameter matrix W, each having a rank of 4 (S1204).

[0096] Next, the summarization model generation system 12 updates each low-rank matrix X,Y based on the training data P without updating the pre-trained parameter matrix W (S1205). If the summarization model generation system 12 determines that it has not finished updating each low-rank matrix X,Y (S1206; N), the summarization model generation system 12 repeats the update process for each low-rank matrix X,Y (S1105). In this embodiment, the summarization model generation system 12 determines whether to finish updating each low-rank matrix X,Y based on whether the loss has fallen below a predetermined value. Alternatively, the summarization model generation system 12 may determine whether to finish updating each low-rank matrix X,Y based on, for example, whether the process in S1205 has been repeated a predetermined number of times.

[0097] When the summarization model generation system 12 determines that it has finished updating the low-rank matrices X and Y (S1206; Y), the summarization model generation system 12 calculates a retrained parameter matrix W', which has multiple retrained parameters as elements, based on the two updated low-rank matrices X and Y and the pre-trained parameter matrix W (S1207). Finally, the summarization model generation system 12 outputs the large-scale language model containing the retrained parameter matrix W' as the summarization model M (S1208), and terminates this process.

[0098] [4-3. Processing performed in the training data generation system] Figure 13 is a flowchart showing an example of processing performed by the learning data generation system 14 according to an embodiment of the present invention. As shown in Figure 13, the learning data generation system 14 first obtains the learning product title T4 (S1300). Next, the learning data generation system 14 obtains the importance Im of each of the multiple attributes A (S1301). The learning data generation system 14 selects the attributes A whose importance Im is greater than or equal to a predetermined value as multiple attention attributes FA, and obtains the content of each of the multiple attention attributes FA for the learning product which is the product indicated by the learning product title T4 (S1302). Based on the content FA4 of each of the multiple attention attributes FA for the learning product, the learning data generation system 14 extracts strings corresponding to the content FA4 of each of the multiple attention attributes FA for the learning product from the learning product title T4 (S1303). Then, the learning data generation system 14 outputs the correct answer data L4 based on the strings corresponding to the content FA4 of each of the multiple attention attributes FA extracted in S1303 (S1304). Finally, the training data generation system 14 outputs the training data P (S1005) and terminates the process.

[0099] Note that the order of the process for obtaining the learning product title T4 (S1300) and the process for obtaining the content FA4 for each of the multiple attention attributes FA of the learning product (S1301 and S1302) may be reversed from the order shown in Figure 10.

[0100] According to one aspect of the summarization system 1 of this embodiment described above, the accuracy of summaries generated by machine learning models can be improved.

[0101] Furthermore, according to another aspect of the summarization system 1 according to this embodiment described above, training data used to generate a machine learning model for generating summaries of product titles can be easily generated.

[0102] [6. Variant] This disclosure is not limited to the embodiments described above. It may be modified as appropriate without departing from the spirit of this disclosure.

[0103] For example, in this embodiment, the summarization system 1 is used to summarize the titles of products displayed on an e-commerce site, that is, to summarize text that provides information about products. However, the subjects of summarization by the summarization system 1 are not limited to text that provides information about products. That is, the summarization system 1 may also be used to summarize text that provides information about services. For example, the summarization system 1 may be used to summarize the introductory text of a travel plan displayed on a travel booking site. In short, the summarization system 1 may target product information text that provides information about products or services, i.e., merchandise.

[0104] Furthermore, although this embodiment illustrates a case where there are multiple attributes of interest, there may be only one attribute of interest. Alternatively, all of the multiple attributes that the product possesses may be designated as attributes of interest.

[0105] Furthermore, the summarization model M is not limited to large-scale language models, but may be any known model trained based on machine learning techniques, such as recurrent neural networks, long-term memory networks, or Transformers.

[0106] Furthermore, the method for retraining large-scale language models is not limited to LoRA or similar methods; for example, it could be prefix tuning or an adapter, or it could be full fine-tuning, which involves retraining all pre-trained parameters.

[0107] The learning data generation system 14 does not necessarily have to include a learning data output unit 147. Furthermore, the learning data generation system 14 does not necessarily have to include an importance acquisition unit 143. The learning data generation system 14 only needs to include at least a learning product title acquisition unit 141, a focus attribute acquisition unit 144, a string extraction unit 145, and a summary output unit 146.

[0108] [7. Addendum] (1) A summary generation system that generates a summary of product information text that shows information about a product, A means for acquiring a target product information text, which is the product information text that is the subject of the summary, A target product information text indicates a target product, and the target product information text indicates a target product product, and the target product product information means for acquiring the content of one or more of the multiple attributes of the product product, which are at least some of the attributes of the product product, A target summary acquisition means inputs the target product information text and the content of each of the one or more attention attributes for the target product into a summarization model and acquires a target summary which is a summary of the target product information text output from the summarization model. It has, The summary model is generated by learning based on learning data that includes the learning product information text, which is the learning product information text for learning purposes; the content of each of the one or more attention attributes of the learning product indicated in the learning product information text; and the ground truth data, which is a summary of the learning product information text. Summary generation system. (2) Product data storage means that stores the identification information of one or more products and the content of each of the one or more notable attributes for said product in association with each other. It further possesses, The aforementioned target attention attribute acquisition means acquires identification information of the target product and, based on the identification information, acquires the content of each of the one or more attention attributes for the target product stored by the product data storage means. (1) The summary generation system described above. (3) The aforementioned training data is To obtain the aforementioned learning material information text, Regarding the aforementioned learning materials, to obtain the content of each of the one or more notable attributes mentioned above, Based on the content of each of the one or more notable attributes of the learning material, extract strings corresponding to each of the one or more notable attributes of the learning material from the learning material information text. Based on the extracted string, output a summary shorter than the learning material information text. It was generated by (1) or (2) the summary generation system. (4) The aforementioned training data is To obtain the importance of each of the aforementioned multiple attributes, Among the multiple attributes, those whose importance is equal to or greater than a predetermined value are designated as one or more attention attributes, and the content of each of these one or more attention attributes is obtained for the learning material. It was generated by (3) The summary generation system described above. (5) The aforementioned training data is The summary is output by arranging the extracted string so that it is less than or equal to a predetermined number of characters, which is shorter than the learning material information text. It was generated by (3) or (4) the summary generation system. (6) The aforementioned summarization model is generated by retraining a pre-trained large-scale language model based on the aforementioned training data. A summary generation system as described in any of (1) to (6). (7) The training data further includes instruction sentences that indicate the output content of the large-scale language model. (6) The summary generation system described in (8). The training data further includes conditional statements indicating the conditions that the output of the large-scale language model must satisfy. (6) or (7) The summary generation system described above. (9) The conditional statement indicates that the output of the large-scale language model should include a string corresponding to each of the one or more attributes of interest. (8) The summary generation system described above. (10) The conditional statement indicates that the output of the large-scale language model should be a string of a predetermined number of characters or less, which is shorter than the learning material information text. (8) or (9) The summary generation system described above. (11) The summarization model is generated by updating multiple parameters to be updated for the multiple pre-trained parameters, based on the training data, without updating the multiple pre-trained parameters included in the pre-trained large-scale language model. A summary generation system as described in any of (8) to (10). (12) The aforementioned summary model is, With respect to the pre-trained parameter matrix, which is a matrix whose elements are the aforementioned multiple pre-trained parameters, two low-rank matrices are calculated, each having the aforementioned multiple parameters to be updated as elements, and each having a rank smaller than the number of rows and columns of the pre-trained parameter matrix. Without updating the pre-trained parameter matrix, update each of the low-rank matrices based on the training data. Based on the two updated low-rank matrices and the pre-trained parameter matrix, a retrained parameter matrix is ​​calculated, with multiple retrained parameters as elements. Outputting the large-scale language model, including the retrained parameter matrix, as the summary model. It was generated by (11) The summary generation system described above. (13) The aforementioned summary model is, Determine whether the conditional statement indicates a predetermined condition, and if it is determined that the conditional statement indicates a predetermined condition, calculate the two low-rank matrices such that each has a rank greater than the rank specified for conditions other than the predetermined condition. It was generated by (12) The summary generation system described above. (14) The aforementioned summary model is, The number of conditional statements in the training data that indicate the conditions that the output of the large-scale language model must satisfy is calculated, and the two low-rank matrices are calculated such that the rank of each matrix increases with the number of calculated conditional statements. It was generated by The summary generation system described in (12) or (13). [Explanation of Symbols]

[0109] 1 Summarization system, N network, 10 Summarization generation system, 12 Summarization model generation system, 14 Training data generation system, 10a,12a,14a Control unit, 10b,12b,14b Storage unit, 10c,12c,14c Communication unit, 10d,12d,14d Operation unit, 10e,12e,14e Display unit, 100,140 Product database storage unit, 102 Target product title acquisition unit, 104 Target attention attribute acquisition unit, 106 Target summary acquisition unit, 108,126 Model storage unit, 120,148 Training database storage unit, 122 Training data acquisition unit, 124 Summarization model generation unit, 1240 Low rank matrix calculation unit, 1242 Low rank matrix update unit, 1244 Parameter matrix calculation unit, 1246 Summarization model output unit, 141 Training product title acquisition unit, 142 Importance database storage unit, 143 Importance acquisition unit, 144 Attribute acquisition unit, 145 String extraction unit, 146 Summary output unit, 147 Training data output unit, M Summary model, P Training data, T3 Target product title, FA3 Content of each of the multiple attributes of interest for the target product, I3,I4 Instructions, C3,C4 Conditional statements, S3 Target summary, T4 Training product title, FA4 Content of each of the multiple attributes of interest for the training product, L4 Ground truth data, A Attribute, FA Attribute of interest, L Ground truth data, m Pre-trained large-scale language model, W Pre-trained parameter matrix, W' Retrained parameter matrix, X,Y Low-rank matrices, DB1 Product database, DB2 Importance database, Im Importance.

Claims

1. A summary generation system that generates a summary of product information text that shows information about a product, A product data storage means that stores the identification information of one or more of the aforementioned products and the content of one or more notable attributes, which are at least a part of the multiple attributes that the product has, in association with each other. A means for acquiring a target product information text, which is the product information text that is the subject of the summary, A target product information text indicates a target product, and a target product information acquisition means acquires the identification information of the target product and, based on the identification information, acquires the content of each of the one or more product information attributes stored in the product data storage means. A target summary acquisition means inputs the target product information text and the content of each of the one or more attention attributes for the target product into a summarization model, and acquires a target summary which is a summary of the target product information text output from the summarization model. It has, The summary model is generated by learning based on learning data that includes the learning product information text, which is the product information text for learning, the content of each of the one or more attention attributes of the learning product indicated in the learning product information text, and the ground truth data, which is a summary of the learning product information text. Summary generation system.

2. The aforementioned training data is To obtain the aforementioned learning material information text, To obtain the content of each of the one or more notable attributes mentioned above for the aforementioned learning materials, Based on the content of each of the one or more notable attributes of the learning material, extract strings corresponding to each of the one or more notable attributes of the learning material from the learning material information text. Based on the extracted string, output a summary shorter than the learning material information text. It was generated by The summary generation system according to claim 1.

3. The aforementioned training data is To obtain the importance of each of the aforementioned multiple attributes, Among the multiple attributes, those whose importance is equal to or greater than a predetermined value are designated as one or more attention attributes, and the content of each of these one or more attention attributes is obtained for the learning material. It was generated by The summary generation system according to claim 2.

4. The aforementioned training data is The summary is output by arranging the extracted string so that it is less than or equal to a predetermined number of characters, which is shorter than the learning material information text. It was generated by The summary generation system according to claim 2.

5. The aforementioned summarization model is generated by retraining a pre-trained large-scale language model based on the aforementioned training data. The summary generation system according to claim 1 or 2.

6. The training data further includes instruction sentences that indicate the output content of the large-scale language model. The summary generation system according to claim 5.

7. The training data further includes conditional statements indicating the conditions that the output of the large-scale language model must satisfy. The summary generation system according to claim 5.

8. The conditional statement indicates that the output of the large-scale language model should include strings corresponding to the content of each of the one or more attention attributes of the learning material. The summary generation system according to claim 7.

9. The conditional statement indicates that the output of the large-scale language model should be a string of a predetermined number of characters or less, which is shorter than the learning material information text. The summary generation system according to claim 7.

10. The summarization model is generated by updating multiple parameters to be updated for the multiple pre-trained parameters, based on the training data, without updating the multiple pre-trained parameters included in the pre-trained large-scale language model. The summary generation system according to claim 7.

11. The aforementioned summary model is, With respect to the pre-trained parameter matrix, which is a matrix whose elements are the aforementioned multiple pre-trained parameters, two low-rank matrices are calculated, each having the aforementioned multiple parameters to be updated as elements, and each having a rank smaller than the number of rows and columns of the pre-trained parameter matrix. Without updating the pre-trained parameter matrix, update each of the low-rank matrices based on the training data. Based on the two updated low-rank matrices and the pre-trained parameter matrix, a retrained parameter matrix is ​​calculated, with multiple retrained parameters as elements. Outputting the large-scale language model, including the retrained parameter matrix, as the summary model. It was generated by The summary generation system according to claim 10.

12. The aforementioned summary model is, Determine whether the conditional statement indicates a predetermined condition, and if it is determined that the conditional statement indicates a predetermined condition, calculate the two low-rank matrices such that each has a rank greater than the rank specified for conditions other than the predetermined condition. It was generated by The summary generation system according to claim 11.

13. The aforementioned summary model is, The number of conditional statements in the training data that indicate the conditions that the output of the large-scale language model must satisfy is calculated, and the two low-rank matrices are calculated such that the rank of each matrix increases with the number of calculated conditional statements. It was generated by The summary generation system according to claim 11.

14. A training data generation system that receives product information text, which contains information about a product, as input and generates training data used to generate a summarization model that outputs a summary of said product information text, A means for acquiring learning material information text, which is a learning material information text for learning purposes, A product data storage means that stores the identification information of one or more of the aforementioned products and the content of one or more notable attributes, which are at least a part of the multiple attributes that the product has, in association with each other. Regarding the learning material that is the learning material indicated in the learning material information text, the attention attribute acquisition means acquires identification information of the learning material and, based on the identification information, acquires the content of each of the one or more attention attributes stored in the product data storage means. A string extraction means for extracting strings from the learning material information text that correspond to the content of each of the one or more attention attributes of the learning material, based on the content of each of the one or more attention attributes of the learning material. A summary output means that outputs a summary shorter than the learning material information text based on the extracted string, A learning data output means that outputs learning data including the learning material information text, the content of each of the one or more attention attributes, and the correct answer data which is the summary output by the summary output means, A learning data generation system having the following features.

15. A method for generating a summary of product information text that shows information about a product, The means for obtaining the target product information text includes a step of obtaining the target product information text which is the product information text that is the subject of the summary, Target Attention Attribute Acquisition Step: The target attention attribute acquisition means acquires identification information for the target product which is the product indicated in the target product information text, and based on the identification information, acquires the content of each of the one or more attention attributes stored in the product data storage means, which stores the content of each of the one or more attention attributes which is at least a part of the multiple attributes that the product has, in association with each of the identification information of one or more of the products. A target summary acquisition step involves the target summary acquisition means inputting the target product information text and the content of each of the one or more attention attributes for the target product into a summary model, and acquiring a target summary which is a summary of the target product information text output from the summary model. It has, The summary model is generated by learning based on learning data that includes the learning product information text, which is the product information text for learning, the content of each of the one or more attention attributes of the learning product indicated in the learning product information text, and the ground truth data, which is a summary of the learning product information text. Summary generation method.

16. A method for generating training data used to generate a summarization model that accepts product information text, which contains information about a product, as input and outputs a summary of said product information text, The means for acquiring learning material information text includes a learning material information text acquisition step, which acquires the learning material information text that is the learning material information text, The attention attribute acquisition means acquires identification information for learning materials that are the learning materials indicated in the learning material information text, and based on the identification information, acquires the content of each of the one or more attention attributes stored in the product data storage means, which associates the content of each of the identification information of one or more of the learning materials with the content of each of the one or more attention attributes that are at least a part of the multiple attributes that the learning materials have. The string extraction means includes a string extraction step of extracting strings from the learning material information text that correspond to the content of each of the one or more attention attributes of the learning material, based on the content of each of the one or more attention attributes of the learning material, The summary output means includes a summary output step that outputs a summary shorter than the learning material information text based on the extracted string, The learning data output means includes a learning data output step which outputs learning data including the learning material information text, the content of each of the one or more attention attributes, and the correct answer data which is the summary output by the summary output step, A method for generating training data, comprising the characteristics of a training data generation method.

17. A program for causing a computer to function in order to generate a summary of product information text that shows information about a product, Product data storage means that stores the identification information of one or more of the aforementioned products and the content of one or more of the notable attributes, which are at least a part of the multiple attributes that the product has, in association with each other. A means for acquiring the target product information text, which is the product information text that is the subject of the summary, A target product information text indicates a target product, and a target product information acquisition means acquires the identification information of the target product and, based on the identification information, acquires the content of each of the one or more product information attributes stored in the product data storage means. A target summary acquisition means that inputs the target product information text and the content of each of the one or more attention attributes for the target product into a summarization model, and acquires a target summary which is a summary of the target product information text output from the summarization model. The computer is made to function as follows: The summary model is generated by learning based on learning data that includes the learning product information text, which is the product information text for learning, the content of each of the one or more attention attributes of the learning product indicated in the learning product information text, and the ground truth data, which is a summary of the learning product information text. program.

18. A program for causing a computer to function in order to generate training data used to generate a summarization model that takes product information text, which contains information about a product, as input and outputs a summary of said product information text, Learning material information text acquisition means for acquiring the learning material information text, which is the learning material information text for learning purposes. Product data storage means that stores the identification information of one or more of the aforementioned products and the content of one or more of the notable attributes, which are at least a part of the multiple attributes that the product has, in association with each other. Regarding the learning material that is the learning material indicated in the learning material information text, an attribute acquisition means acquires identification information of the learning material and, based on the identification information, acquires the content of each of the one or more attribute of interest stored in the product data storage means. A string extraction means for extracting strings from the learning material information text that correspond to the content of each of the one or more attention attributes of the learning material, based on the content of each of the one or more attention attributes of the learning material. A summary output means that outputs a summary shorter than the learning material information text based on the extracted string, A learning data output means that outputs learning data including the learning material information text, the content of each of the one or more attention attributes, and the correct answer data which is the summary output by the summary output means. A program to cause the aforementioned computer to function.

Images