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 training data generation system, a summary generation method, a training data generation method, and a program. [Background technology]

[0002] There is known a technique for generating a summary of a text showing 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 a real estate property introduction based on the property's appealing points (features), such as rent and distance from the nearest station.

[0003] In particular, in recent years, technology for generating summaries of commercial product information text using machine learning models has been studied. As an example, Patent Document 2 discloses a machine learning model that generates summaries of user reviews of movies based on aspects such as the story, cast, and direction. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Patent Publication No. 2018-128806 [Patent Document 2] Patent Publication No. 2020-101936 Summary of the Invention [Problem to be solved by the invention]

[0005] However, in the above-mentioned conventional techniques, the accuracy of summaries generated by machine learning models may not be sufficient.

[0006] The present invention has been made in consideration of the above-mentioned problems, and one of its objectives is to provide a summary generation system, a summary generation method, and a program that can improve the accuracy of summaries generated by machine learning models.

[0007] Another object of the present invention is to provide a training data generation system, a training data generation method, and a program that can easily generate training data to be used in generating a machine learning model that generates summaries of 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 indicates information about a product, and includes: a target product information text acquisition means that acquires target product information text, which is the product information text that is the target of the summary; a target attention attribute acquisition means that acquires the content of one or more attention attributes, which are at least a part of the multiple attributes possessed by the target product, which is the product indicated by the target product information text; and 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 summary model and acquires a target summary, which is a summary of the target product information text, output from the summary model.The summary model is generated by learning based on learning data that includes learning product information text, which is the product information text for learning, the content of each of the one or more attention attributes for the learning product, which is the product indicated by the learning product information text, and correct answer data that is a summary of the learning product information text.

[0009] A training data generation system according to one embodiment of the present invention is a training data generation system that receives as input a product information text indicating information about a product and generates training data to be used in generating a summary model that outputs a summary of the product information text.The system includes: a learning product information text acquisition means for acquiring the learning product information text, which is the product information text for learning; a featured attribute acquisition means for acquiring the content of one or more featured attributes, which are at least a part of the multiple attributes possessed by the learning product, which is the product indicated by the learning product information text; a string extraction means for extracting strings from the learning product information text that correspond to the content of each of the one or more featured attributes for the learning product, based on the content of each of the one or more featured attributes for the learning product; and a summary output means for outputting a summary that is 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, it is possible to improve the accuracy of summaries generated by machine learning models.

[0011] Furthermore, according to another aspect of the present invention, it is possible to easily generate training data used to generate a machine learning model that generates a summary of product information text. [Brief explanation of the drawings]

[0012] [Figure 1] FIG. 1 is a diagram illustrating an example of a hardware configuration of a summarization system according to an embodiment of the present invention. [Figure 2A] FIG. 10 is a diagram showing an example of a screen displayed on an e-commerce site. [Figure 2B] FIG. 10 is a diagram showing an example of a screen displayed on an e-commerce site. [Figure 3] FIG. 1 is a diagram illustrating generation of a summary using a summary model according to an embodiment of the present invention. [Figure 4A] FIG. 10 is a diagram illustrating generation of a summary model. [Figure 4B]FIG. 10 is a diagram illustrating generation of a summary model. [Figure 4C] FIG. 10 is a diagram illustrating generation of a summary model. [Figure 5] FIG. 10 is a diagram illustrating generation of learning data. [Figure 6] 1 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] FIG. 2 is a diagram illustrating an example of a product database. [Figure 8] FIG. 2 is a functional block diagram showing an example of functions realized by the summary model generation system according to the embodiment of the present invention. [Figure 9] FIG. 2 is a functional block diagram showing an example of functions realized by the training data generation system according to the embodiment of the present invention. [Figure 10] FIG. 10 is a diagram illustrating an example of an importance database. [Figure 11] FIG. 2 is a flowchart showing an example of processing executed by the summary generation system according to the embodiment of the present invention. [Figure 12] FIG. 2 is a flowchart illustrating an example of processing executed by the summary model generation system according to the embodiment of the present invention. [Figure 13] FIG. 2 is a flowchart illustrating an example of processing executed in the training data generation system according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0013] An embodiment of the present invention will be described below with reference to the drawings. In this embodiment, the summarization system 1 is used to summarize the titles of products displayed on an e-commerce site. Note that products include not only tangible items but also intangible items (e.g., downloadable content such as movies and manga).

[0014] [1. Hardware configuration of the summarization system according to this embodiment] Fig. 1 is a diagram showing an example of the hardware configuration of a summarization system 1 according to an embodiment of the present invention. As shown in Fig. 1, the summarization system 1 includes a summary generation system 10, a summary model generation system 12, and a training data generation system 14. The summary generation system 10, the summary model generation system 12, and the training data generation system 14 are connected to each other via a network N such as the Internet or a LAN. Although the example shows a case in which the summary generation system 10, the summary model generation system 12, and the training data generation system 14 are implemented by one or more different computers, the summary generation system 10, the summary model generation system 12, and the training data generation system 14 may also be implemented by the same one or more computers.

[0015] The summary generation system 10 is a computer system including one or more computers. The summary generation system 10 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-controlled device such as a CPU that operates according to a program stored in the memory unit 10b. The memory unit 10b is, for example, a storage element such as a ROM or RAM, or a hard disk drive. The memory unit 10b stores programs and the like 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 made up of one or more computers. The summary model generation system 12 is a system that generates a summary model M (see FIG. 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, the storage unit 12b, the communication unit 12c, the operation unit 12d, and the display unit 12e may be similar to the physical configurations of the control unit 10a, the storage unit 10b, the communication unit 10c, the operation unit 10d, and the display unit 10e, respectively.

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

[0019] [2. Overview of the summarization system according to this embodiment] 2A and 2B are diagrams illustrating examples of screens displayed on an e-commerce site. Screen G2A in FIG. 2A is an example of a product detail page displaying product details. As shown in FIG. 2A, screen G2A displays a product title T2. The product title T2 is a string of characters indicating various information about the product, such as "ABC Company Iroha Laundry Detergent, Liquid Type, Refill, Eco Size (500 mL), Set of 6 [Iroha]" shown in FIG. 2A. As an example, the product title T2 displayed on screen G2A includes the product brand name ("ABC Company"), product name ("Iroha"), product type ("Laundry Detergent"), product form ("Liquid Type," "Refill"), product size ("Eco Size (500 mL)"), and product quantity ("Set of 6"). The character strings included in the product title T2 are not limited to the above example. The product title T2 may include any character string, such as a product introduction or promotional statement created by the seller of the product.

[0020] When the "Back" button displayed in the upper left corner of screen G2A is selected, the screen transitions from screen G2A, which is a product details page, to screen G2B, which is a product list page shown in Fig. 2B. On a typical e-commerce site, when product-related keywords (such as product type, product name, or brand name) are entered into a search box or the like, a product list page is displayed showing a list of products that match the keywords. Screen G2B shown in Fig. 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 ("Refill Eco Size (500 mL) Set of 6 [Iroha]"), such as "ABC Company Iroha Laundry Detergent Liquid Type..." shown in Figure 2B.

[0022] However, the product title T2 shown in Figure 2A was created with the assumption that it would be displayed on a large screen such as a desktop device, and is too long (i.e., too many characters) for the small screen of a mobile device such as a smartphone or tablet. As a result, the entire product title T2 cannot fit on the product list page displayed on the screen of a mobile device. As a result, as shown in Figure 2B, instead of the full product title T2, a character string t2 with part of the product title T2 omitted must be displayed on the product list page. This situation reduces the readability of the product list page on a mobile device, and ultimately reduces usability.

[0023] Therefore, in the field of e-commerce, there is a growing need for technology that can create summaries of product titles. If product title summaries can be prepared that are long enough to fit the small screens of mobile devices, the browsability of product list pages on mobile devices can be ensured. Furthermore, if product title summaries succinctly express the information users are looking for, they can be motivated to access the product detail page of each product and consider purchasing that product.

[0024] In light of the above situation, the inventors of the present application have been studying a technology for generating product title summaries using a machine learning model. While manually creating product title summaries is extremely time-consuming, using a machine learning model makes it possible to easily obtain product title summaries. While rule-based text summaries have been known as a method for automating text summaries, summary generation using a machine learning model has the advantage of being able to obtain summaries with more natural expressions than summary generation using such rule-based methods. Therefore, the inventors of the present application have decided to newly study a summaries model that accepts product titles as input and outputs a summary of the product titles.

[0025] However, the accuracy of summaries generated by machine learning models may not be sufficient. For example, if certain constraints are imposed on the summaries generated by machine learning models to obtain summaries that satisfy requests such as including certain information in product title summaries or limiting the number of characters in product title summaries, it is difficult to improve the accuracy of the summaries generated by machine learning models.

[0026] Therefore, in this embodiment, as shown in Fig. 3, a target product title T3 to be summarized and the contents FA3 of each of a plurality of focus attributes FA (see Fig. 7) of the target product indicated by the target product title T3 are input to a summary model M, and a target summary S3, which is a summary of the target product title T3, is obtained, output from the summary model M. Fig. 3 is a diagram illustrating the generation of a summary by the summary model M according to an embodiment of the present invention. The plurality of focus attributes FA are part of a plurality of attributes A (see Fig. 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 summary model M generates a target summary S3, which is a summary of the target product title T3, in addition to the target product title T3, the content FA3 of each of the multiple focus attributes FA for the target product is taken into consideration, thereby improving the accuracy of the generated target summary S3.

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

[0029] [2-1. Summary model generation method] A method for generating the summary model M shown in Fig. 3 will be described below with reference to Figs. 4A to 4C. Figs. 4A to 4C are diagrams for explaining the generation of the summary model M. The summary model M is generated by the summary model generation system 12 shown in Fig. 1. As shown in Figs. 4A and 4B, the summary model generation system 12 according to this embodiment generates the summary model M by re-training a pre-trained large-scale language model m based on training data P so as to receive a product title as input and output a summary of the product title. In other words, the summary model M is a large-scale language model that has been re-trained based on the training data P and is specialized for the task of receiving a product title as input and outputting a summary of the product title.

[0030] As shown in Figure 4A, the learning data P used to generate the summary model M includes a learning product title T4, the contents FA4 of each of multiple focus attributes FA for the learning product indicated by the learning product title T4, correct answer data L4, an instruction statement I4, and a conditional statement C4.

[0031] The learning product title T4 is the title of a product for learning. The learning product title T4 is a character 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 below 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 character string, such as an introduction or appeal statement about the learning product. Note that the learning product title refers to the product title used to generate the summary model M.

[0032] The multiple focus attributes FA are part of the multiple attributes A that a product has. Each attribute A is a feature that multiple products have in common. The multiple attributes A may be, for example, the brand name of the product, the product name, the size of the product, the quantity of the product, etc., but are not limited to this example. The multiple focus attributes FA may be, for example, attributes A that have been artificially selected by a user of the summarization system 1 (in this embodiment, a business operator operating an e-commerce site). As an example, this embodiment will explain a case where the multiple focus attributes FA are the product name, the size of the product, and the quantity of the product from the attributes A listed above.

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

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

[0035] Conditional statement C is text data indicating a condition that the output of the large-scale language model must satisfy. In this embodiment, conditional statement C indicates that the output of the large-scale language model must include character strings corresponding to the contents FA4 of the multiple focus attributes FA of the study product. Conditional statement C also indicates that the output of the large-scale language model must be a character string that is shorter than the study product title T4 and has a predetermined number of characters or less.

[0036] 4B and 4C, a more detailed description will be given of a method for generating the summary model M. Specifically, as shown in Fig. 4B, the summary model generation system 12 retrains the pre-trained large-scale language model m by updating a plurality of update target parameters for the plurality of pre-trained parameters based on the training data P without updating a plurality of pre-trained parameters included in the pre-trained large-scale language model m.

[0037] As shown in the following formula (1), the summary system 1 does not update the pre-trained parameter matrix W, whose elements are multiple pre-trained parameters, but instead updates the difference matrix ΔW, which is the difference between the pre-trained parameter matrix W and the re-trained parameter matrix W', whose elements are multiple re-trained parameters, based on the training data P, thereby re-training the pre-trained large-scale language model m. As an example in this embodiment, the number of rows and the number of columns are all d for the re-trained parameter matrix W', the pre-trained parameter matrix W, and the difference matrix ΔW. Note that the number of rows and the number of columns for the re-trained parameter matrix W', the pre-trained parameter matrix W, and the difference matrix ΔW may differ from one another.

[0038]

number

[0039] The difference matrix ΔW is the product of two low-rank matrices X and Y, as shown in the following equation (2). Each of the two low-rank matrices X and Y has multiple parameters to be updated as elements. The number of rows and the number of columns of the low-rank matrix X are d and r, respectively. The number of rows and the number of columns of the low-rank matrix Y are 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 the number of 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 large-scale pre-trained language model m by directly updating the pre-trained parameter matrix W, the number of parameters to be calculated in the retraining is d 2 In contrast, in summary model generation system 12, which re-trains pre-trained large-scale language model m by updating each of low-rank matrices X and Y, the number of parameters included in each of low-rank matrices X and Y is rd, so the number of parameters to be calculated in re-training is at most 2rd. Here, r is sufficiently smaller than d, and therefore summary model generation system 12 can reduce the computational load required for re-training compared to when re-training is performed by directly updating the pre-trained parameter matrix W.

[0042] The re-learning process executed by the summary model generation system 12 according to this embodiment specifically includes the following processes (1) to (4).

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

[0044] (2) Low-rank matrix update Next, the summary model generation system 12 updates each of the low-rank matrices X and Y calculated in (1) based on the training data P. At this time, the pre-trained parameter matrix W is not updated. That is, in the process of (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 4B, the summary model generation system 12 calculates a re-trained parameter matrix W' having multiple re-trained parameters as elements, based on the two updated low-rank matrices X and Y and the pre-trained parameter matrix W. Specifically, the summary model generation system 12 calculates the re-trained 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) Summary model output Finally, the summary model generation system 12 outputs the large-scale language model including the retrained parameter matrix W′ as the summary model M.

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

[0048] Specifically, the summary 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 so that each has a higher rank as the calculated number of conditional statements C4 increases. In this embodiment, as an example, as shown in FIG. 4C , when the training data P does not include any conditional statements C4, the summary model generation system 12 calculates two low-rank matrices X and Y so that each has a rank of 2. On the other hand, when the training data P includes one or more conditional statements C4, the summary model generation system 12 calculates two low-rank matrices X and Y so that each has a rank of 3.

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

[0050] As described above, the summary model generation system 12 calculates two low-rank matrices X and Y so that each has a higher rank as the number of conditional statements C4 increases or when the conditional statement C4 indicates a predetermined condition. As a result, the more stringent the constraints imposed on the output of the large-scale language model, the more rank the two low-rank matrices X and Y can be calculated. In other words, the more stringent the constraints imposed on the output of the large-scale language model, the more parameters the summary model generation system 12 updates during re-learning. Therefore, the summary model generation system 12 can generate a summary model M that can accurately generate summaries that are difficult to perform.

[0051] [2-2. How to generate training data] A method for generating training data P used to generate a summary model M will be described below with reference to Fig. 5. Fig. 5 is a diagram illustrating the generation of training data P. The training data P is generated by the training data generation system 14 shown in Fig. 1.

[0052] As shown in FIG. 4A, the training data P used to generate the summary model M includes correct answer data L4, i.e., a correct summary determined for the training product title T4. Here, as shown in FIG. 3, the summary model M accepts input of not only the target product title T3 but also the content FA3 of each of the multiple focus attributes FA of the target product, the instruction I3, and the conditional statement C3, and outputs a summary of the content corresponding to the content FA3 of each of the multiple focus attributes FA of the target product, the instruction I3, and the conditional statement C3. In other words, the correct answer data L is not obtained by simply summarizing the training product title T4, but is obtained by summarizing the training product title T4 in consideration of the content FA4 of each of the multiple focus attributes FA of the training product included in the training data P, the instruction I4, and the conditional statement C4. Because preparing such correct answer data L is extremely time-consuming, it is difficult to prepare a large amount of training data P used to train the summary model M.

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

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

[0055] (2) Acquiring the content of each of multiple key attributes for educational materials The learning data generation system 14 acquires the content of each of the multiple focus 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 identification number ID4 (see Figure 7) of the learning product, and acquires the content FA4 of each of the multiple focus attributes FA for the learning product stored in the product database storage unit 140 based on the identification number ID4.

[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 has, which are stored in the importance database storage unit 142, and selects the multiple attributes A whose importance is equal to or greater than a predetermined value as the focus attribute FA, and acquires the content FA4 of each of the multiple focus attributes FA for the learning material.

[0057] (3) Extraction of character strings corresponding to the content of multiple important attributes of educational materials The learning data generation system 14 extracts character strings corresponding to the contents FA4 of each of the plurality of noteworthy attributes FA of the learning product from the learning product title T4 based on the contents FA4 of each of the plurality of noteworthy attributes FA of the learning product.

[0058] (4) Summary output Then, as shown in FIG. 5, the learning data generation system 14 outputs correct answer data L4, which is a summary that is shorter than the learning product title T4, based on the character string extracted in (3).

[0059] (5) Output of training data Finally, the learning data generation system 14 outputs learning data P including the learning product title T4, the content FA4 of each of the multiple focus attributes FA for the learning product, and the correct answer data L4. Note that, 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 the present embodiment described above, character strings corresponding to the content FA4 of each of the plurality of focus attributes FA of the learning product are extracted from the learning product title T4, and correct answer data L4 is output based on the character strings. In other words, according to the learning data generation system 14, 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 plurality of focus attributes FA of the learning product, can be easily obtained, and therefore learning data P can be easily generated.

[0061] The summarization system 1 will be described in detail below with reference to FIGS.

[0062] [3. Functions realized by the summarization system according to this embodiment] The following describes the functions realized by the summarization system 1 according to this embodiment.

[0063] [3-1. Functions realized by the summary generation system] Fig. 6 is a functional block diagram showing an example of functions realized by the summary generation system 10 according to an embodiment of the present invention. As shown in Fig. 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 a product database DB1, as shown in FIG. 7. FIG. 7 is a diagram illustrating an example of the product database DB1. The product database storage unit 100 stores an identification number ID (identification information) of each of a plurality of products and the contents of each of a plurality of focus attributes FA for the product in association with each other. Specifically, as shown in FIG. 7, the product database DB1 stores a plurality of product data in each record. Each product data has the contents of a plurality of items, including an identification number ID ("product ID") and a plurality of attributes A ("brand," "product name," "size," "quantity," etc.), for the product corresponding to the product data. As described above, in this embodiment, the case where the attributes are "product name," "size," and "quantity" is exemplified. Note that in this embodiment, as an example, the first record of the product database DB1 shown in FIG. 7 is used for processing by the summary generation system 10, and the second record is used for processing by the training data generation system 14, which will be described later. That is, in this embodiment, as an example, it is assumed that the product ID "12345" is the identification number ID4 of the target product, and the 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 target of the summary (see FIG. 3). The target product title acquisition unit 102 may acquire the target product title T3, for example, by 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 acquire the target product title T3 input 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 attributes FA for the target product, which is the product indicated by the target product title T3 (see FIG. 3). Specifically, the target attention attribute acquisition unit 104 acquires the identification number ID3 of the target product (see FIG. 7), and acquires the content FA3 of each of the multiple attention attributes FA for the target product stored in the product database DB1 based on the identification number ID3. Note that the target attention attribute acquisition unit 104 may acquire the content FA3 of each of the multiple attention attributes FA for the target product by referring to the storage contents of an external device or information storage medium.

[0067] The target summary acquisition unit 106 inputs the target product title T3 and the content FA3 of each of a plurality of focus attributes FA of the target product into the summary model M, and acquires a target summary S3, which is a summary of the target product title T3, output from the summary model M (see FIG. 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 FIG. 6. The summary model M is generated by the summary model generation system 12, which will be described later. Therefore, details of the summary model M will be described later. The summary model M may be stored in an external device or information storage medium. The summary model M may also be generated by an external system, etc.

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

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

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

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

[0072] Specifically, the summary model generation unit 124 re-trains the pre-trained large-scale language model m by updating a plurality of update target parameters for a plurality of pre-trained parameters based on the training data P, without updating a plurality of pre-trained parameters included in the pre-trained large-scale language model m. That is, as shown in FIG. 8 , the summary model generation unit 124 includes a low-rank matrix calculation unit 2020, a low-rank matrix update unit 2021, a re-trained parameter matrix calculation unit 2022, and a summary model output unit 2023.

[0073] The low-rank matrix calculation unit 2020 calculates two low-rank matrices X and Y, each of which has a rank smaller than the number of rows and columns of the pre-trained parameter matrix W, for the pre-trained parameter matrix W, which is a matrix having a plurality of pre-trained parameters as elements, and which have a plurality of parameters to be updated as elements (see FIG. 4B ). In this embodiment, in which 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 a query weight matrix, a key weight matrix, a value weight matrix, and a weight matrix of a fully connected layer used in an attention mechanism. As shown in FIG. 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 but 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 but 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 FIG. 4A) indicating 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 so that the greater the number of calculated conditional statements C4, the greater the ranks of each of the matrices X and Y (see FIG. 4C).

[0075] Furthermore, the low-rank matrix calculation unit 2020 determines whether or not the conditional statement C4 indicates a predetermined condition, and if it is determined that the conditional statement C4 indicates the predetermined condition, calculates two low-rank matrices X and Y so that each has a rank greater than the rank determined for conditions other than the predetermined condition (see FIG. 4C).

[0076] The low-rank matrix update unit 2021 updates each of the low-rank matrices X and Y based on the training data P without updating the pre-trained parameter matrix W (see FIG. 4B ). Specifically, the low-rank matrix update unit 2021 updates each of the low-rank matrices X and Y based on the output from a large-scale language model including the pre-trained parameter matrix W and the low-rank matrices X and Y when the training data P is input to the large-scale language model, and the supervised data L included in the training data P. More specifically, the low-rank matrix update unit 2021 updates each of the low-rank matrices X and Y so as to reduce the loss based on the output from the large-scale language model including the pre-trained parameter matrix W and the low-rank matrices X and Y when the training data P is input to the large-scale language model, and the supervised data L included in the training data P. As a method for adjusting parameters based on the loss, a known method such as backpropagation or gradient descent may be used.

[0077] The re-trained parameter matrix calculation unit 2022 calculates a re-trained parameter matrix W' having multiple re-trained 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 FIG. 4B). Specifically, the re-trained parameter matrix calculation unit 2022 calculates the re-trained 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 summary model output unit 2023 outputs the large-scale language model including the re-trained parameter matrix W' as the summary model M (see FIG. 4B). Specifically, the summary model output unit 2023 outputs, as the summary model M, the summary model obtained by replacing the pre-trained parameter matrix W with the re-trained parameter matrix W' in the pre-trained large-scale language model m.

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

[0080] Furthermore, LoRA, a well-known technology, can be used for the above-described low-rank matrix calculation unit 2020, low-rank matrix update unit 2021, retrained parameter matrix calculation unit 2022, and summary model output unit 2023. 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 be, for example, QLoRA (Quantized LoRA) or LoHA (LoRA with Hadamard Product Representation), which are known as improved versions of LoRA.

[0081] [3-3. Functions realized by the learning data generation system] FIG. 9 is a functional block diagram illustrating an example of functions implemented by the training data generation system 14 according to an embodiment of the present invention. As shown in FIG. 9 , the training data generation system 14 functionally includes a product database storage unit 140, a training product title acquisition unit 141, an importance database storage unit 142, an importance acquisition unit 143, a key attribute acquisition unit 144, a character string extraction unit 145, a summary output unit 146, a training data output unit 147, and a training database storage unit 148. The training product title acquisition unit 141, the importance acquisition unit 143, the key attribute acquisition unit 144, the character string extraction unit 145, the summary output unit 146, and the training data output unit 147 are primarily implemented by the control unit 14a. The product database storage unit 140, the importance database storage unit 142, and the training database storage unit 148 are primarily implemented by the storage unit 14b. Details of the product database storage unit 140 are similar to those of the product database storage unit 100 of the summary generation system 10, and therefore will not be described here.

[0082] The learning product title acquisition unit 141 acquires a learning product title T4, which is a product title for learning (see FIG. 5). The learning product title acquisition unit 141 may acquire the learning product title T4, for example, by 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 acquire the learning product title T4 input by the user via the operation unit 14d.

[0083] The importance database storage unit 142 stores the importance database DB2 shown in FIG. 10. FIG. 10 is a diagram illustrating an example of the importance database DB2. As shown in FIG. 10, the importance database DB2 stores a plurality of attributes A in association with the importance Im of the attribute. FIG. 7 illustrates, as an example, a case where the importance Im of the attribute "brand name" is "B," and the importance Im of the attributes "product name," "size," and "quantity" is "A." The importance Im of each attribute A may be manually set, for example, by the operator of an e-commerce site, who 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 highest, "B" being the second highest, and "C" being the third highest. The importance Im is not limited to a discrete value as shown in FIG. 10, but may also be a continuous value. The notation format of the importance Im is not limited to the example shown in FIG. 10.

[0084] The importance obtaining unit 143 obtains the importance Im of each of the multiple attributes A. In this embodiment, the importance obtaining unit 143 refers to the contents stored in the importance database storage unit 142, and obtains the importance Im of each of the multiple attributes A. Note that the importance obtaining unit 143 may obtain the importance Im of each of the multiple attributes A by referring to the contents stored in an external device or information storage medium.

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

[0086] The focused attribute acquisition unit 144 sets the multiple attributes A whose importance Im is equal to or greater than a predetermined value as the multiple focused attributes FA, and acquires the contents of each of the multiple focused attributes FA for the learning product indicated by the learning product title T4. In this embodiment, as an example, the predetermined value is "A" (see FIG. 10). In this embodiment, the focused attribute acquisition unit 144 sets the multiple attributes A whose importance Im is equal to or greater than "A" as the multiple focused attributes FA, and acquires the contents of each of the "product name," "size," and "quantity" for the learning product.

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

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

[0089] The learning data output unit 147 outputs learning data P including the learning product title T4, the content FA4 of each of the multiple focus 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 training data output unit 147 outputs training data P that further includes an instruction statement I4 that indicates the output content of the large-scale language model and a conditional statement C4 that indicates a condition that the output of the large-scale language model must satisfy. In this embodiment, the conditional statement C4 indicates that the output of the large-scale language model should include a character string corresponding to the content FA4 of each of multiple focus attributes FA of the training product. The conditional statement C4 also indicates that the output of the large-scale language model should be a character string that is shorter than the training product title T4 and has a predetermined number of characters or less. That is, as an example, in this embodiment, the number of conditional statements C4 is two. Note that the conditional statement C4 may indicate any condition other than those described above.

[0091] 4. Processing Executed by the Summarization System According to the Present Embodiment The following describes the processing executed by the summarization system 1 according to this embodiment with reference to Figures 11 to 13. The processing in Figures 11 to 13 is executed by the control units 10a, 12a, and 14a executing programs stored in the storage units 10b, 12b, and 14b, respectively.

[0092] [4-1. Processing performed by the summary generation system] 11 is a flow diagram showing an example of processing executed by the summary generation system 10 according to an embodiment of the present invention. As shown in FIG. 11, the summary generation system 10 first acquires a target product title T3 (S1100). Next, the summary generation system 10 acquires the content FA3 of each of a plurality of focus attributes FA of the target product (S1101). Finally, the summary generation system 10 inputs the target product title T3, the content FA3 of each of a plurality of focus attributes FA of the target product, a directive statement I3 that specifies the output content of the summary model M, and a conditional statement C3 that indicates the condition that the output of the summary model M must satisfy into the summary model M, acquires a target summary S3 that is a summary of the target product title T3 output from the summary model M (S1102), and terminates this processing.

[0093] The order of the process for acquiring the target product title T3 (S1100) and the process for acquiring the content FA3 of each of the multiple target attributes FA for the target product (S1101) may be reversed from the order shown in FIG.

[0094] [4-2. Processing performed by the summary model generation system] 12 is a flow diagram showing an example of processing executed by the summary model generation system 12 according to an embodiment of the present invention. As shown in FIG. 12, the summary model generation system 12 first acquires training data P (S1200) and calculates the number of conditional statements C4 included in the training data P (S1201). In this embodiment, for ease of explanation, it is assumed, as an example, that the number of conditional statements C4 calculated in S1201 is two. Furthermore, in this embodiment, if the number of conditional statements C4 is two and the conditional statements C4 do not indicate a predetermined condition (described later), it is assumed that the ranks of the low-rank matrices X and Y (described later) are each set to three.

[0095] In the following S1202, the summary model generation system 12 determines whether each 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 (S1202; Y), the summary model generation system 12 calculates two low-rank matrices X and Y, each of which has a rank of 3, for the pre-trained parameter matrix W (S1203). If it is determined that the conditional statement C4 does not indicate a predetermined condition (S1202; N), the summary model generation system 12 calculates two low-rank matrices X and Y, each of which has a rank of 4, for the pre-trained parameter matrix W (S1204).

[0096] Next, the summary model generation system 12 updates each of the low-rank matrices X and Y based on the training data P without updating the pre-trained parameter matrix W (S1205). If the summary model generation system 12 determines not to end the update of each of the low-rank matrices X and Y (S1206; N), the summary model generation system 12 repeats the update process (S1105) for each of the low-rank matrices X and Y. In this embodiment, the summary model generation system 12 determines whether to end the update of each of the low-rank matrices X and Y based on whether the loss has become equal to or less than a predetermined value. Note that the summary model generation system 12 may also determine whether to end the update of each of the low-rank matrices X and Y based on, for example, whether the process of S1205 has been repeated a predetermined number of times.

[0097] When the summary model generation system 12 determines to finish updating the low-rank matrices X and Y (S1206; Y), the summary model generation system 12 calculates a re-trained parameter matrix W' having multiple re-trained parameters as elements based on the two updated low-rank matrices X and Y and the pre-trained parameter matrix W (S1207). Finally, the summary model generation system 12 outputs a large-scale language model including the re-trained parameter matrix W' as the summary model M (S1208), and ends this process.

[0098] [4-3. Processing performed by the training data generation system] FIG. 13 is a flow diagram illustrating an example of processing executed by the learning data generation system 14 according to an embodiment of the present invention. As shown in FIG. 13, the learning data generation system 14 first acquires a learning product title T4 (S1300). Next, the learning data generation system 14 acquires the importance Im of each of the multiple attributes A (S1301). The learning data generation system 14 sets the multiple attributes A with importance Im equal to or greater than a predetermined value as multiple focused attributes FA, and acquires the details of each of the multiple focused attributes FA for the learning product indicated by the learning product title T4 (S1302). Based on the details FA4 of each of the multiple focused attributes FA for the learning product, the learning data generation system 14 extracts character strings corresponding to the details FA4 of each of the multiple focused attributes FA for the learning product from the learning product title T4 (S1303). Then, the learning data generation system 14 outputs correct answer data L4 based on the character strings corresponding to the details FA4 of each of the multiple focused attributes FA extracted in S1303 (S1304). Finally, the learning data generation system 14 outputs the learning data P (S1005), and the process ends.

[0099] The order of the process for acquiring the learning product title T4 (S1300) and the process for acquiring the contents FA4 of each of the multiple noteworthy attributes FA for the learning product (S1301 and S1302) may be reversed from the order shown in FIG.

[0100] According to one aspect of the summarization system 1 according to the present embodiment described above, it is possible to improve the accuracy of summaries generated by machine learning models.

[0101] Furthermore, according to another aspect of the summarization system 1 according to the present embodiment described above, it is possible to easily generate training data used to generate a machine learning model that generates summaries of product titles.

[0102] [6. Modifications] The present disclosure is not limited to the above-described embodiments, and can be modified as appropriate without departing from the spirit of the present disclosure.

[0103] For example, in this embodiment, the example is given of the case where the summarization system 1 is used to summarize the titles of products displayed on an e-commerce site, i.e., to summarize text showing information about the products. However, the target of summarization by the summarization system 1 is not limited to text showing information about products. In other words, the summarization system 1 may also be used to summarize text showing information about services. The summarization system 1 may also be used, for example, to summarize the introduction to a travel plan displayed on a travel booking site. In short, the summarization system 1 may summarize product or service, i.e., product information text showing information about the product.

[0104] Although the present embodiment has been described with reference to a case where there are multiple focus attributes, there may be only one focus attribute, or all of the multiple attributes of the product may be set as focus attributes.

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

[0106] Furthermore, the method for retraining a large-scale language model is not limited to LoRA or similar methods, but may be, for example, Prefix Tuning or Adapter, or full fine tuning in which all pre-trained parameters are retrained.

[0107] The learning data generation system 14 does not necessarily have to include the learning data output unit 147. Moreover, the learning data generation system 14 does not necessarily have to include the importance acquisition unit 143. The learning data generation system 14 may include at least the learning product title acquisition unit 141, the important attribute acquisition unit 144, the character string extraction unit 145, and the summary output unit 146.

[0108] [7. Notes] (1) A summary generation system for generating a summary of product information text that indicates 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 that is the target of the summary; 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 possessed by a target product, which is the product indicated by the target product information text; a target summary acquisition means for inputting the target product information text and the contents of each of the one or more focus attributes of 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; and The summary model is generated by learning based on learning data including a learning product information text, which is the product information text for learning, the contents of each of the one or more focus attributes of the learning product, which is the product indicated by the learning product information text, and correct answer data, which is a summary of the learning product information text. Summary generation system. (2) a product data storage means for storing identification information of each of one or more products and the content of each of the one or more noteworthy attributes of the product in association with each other; and The target attention attribute acquisition means acquires identification information of the target product, and acquires the content of each of the one or more attention attributes of the target product stored by the product data storage means based on the identification information. A summary generation system according to (1). (3) The learning data is Obtaining the learning material information text; Acquiring the content of each of the one or more focus attributes for the learning material; Extracting character strings corresponding to each of the one or more noteworthy attributes of the learning material from the learning material information text based on the contents of each of the one or more noteworthy attributes of the learning material; outputting a summary shorter than the learning material information text based on the extracted character string; is generated by A summary generation system according to (1) or (2). (4) The learning data is obtaining an importance level for each of the plurality of attributes; Among the plurality of attributes, those having a degree of importance equal to or greater than a predetermined value are designated as the one or more focused attributes, and content of each of the one or more focused attributes is acquired for the learning material; is generated by (3) A summary generation system according to (3). (5) The learning data is outputting the summary by arranging the extracted character string so that the number of characters is equal to or less than a predetermined number of characters that is shorter than the learning material information text; is generated by A summary generation system according to (3) or (4). (6) The summary model is generated by retraining a pre-trained large-scale language model based on the training data. A summary generation system according to any one of (1) to (6). (7) the training data further includes an instruction statement that instructs the output content of the large-scale language model; (8) The summary generation system according to (6). the training data further includes a conditional statement indicating a condition that an output of the large-scale language model should satisfy; A summary generation system according to (6) or (7). (9) the conditional statement indicates that the output of the large-scale language model should include a character string corresponding to each of the one or more focus attributes; (8) A summary generation system according to (8). (10) the conditional statement indicates that the output of the large-scale language model should be a character string that is shorter than the training product information text and has a predetermined number of characters or less. A summary generation system according to (8) or (9). (11) the summary model is generated by updating a plurality of update target parameters for the plurality of pre-trained parameters based on the training data, without updating a plurality of pre-trained parameters included in the pre-trained large-scale language model. A summary generation system according to any one of (8) to (10). (12) The summary model is For a pre-trained parameter matrix, which is a matrix having the plurality of pre-trained parameters as elements, calculating two low-rank matrices, each having the plurality of update target parameters as elements and each having a rank smaller than the number of rows and the number of columns of the pre-trained parameter matrix; updating each of the low-rank matrices based on the training data without updating the pre-trained parameter matrix; calculating a re-trained parameter matrix having a plurality of re-trained parameters as elements based on the two updated low-rank matrices and the pre-trained parameter matrix; outputting the large-scale language model including the retrained parameter matrix as the summary model; is generated by (11) A summary generation system according to (11). (13) The summary model is determining whether the conditional statement indicates a predetermined condition, and if it is determined that the conditional statement indicates the predetermined condition, calculating the two low-rank matrices so that each of the low-rank matrices has a rank greater than a rank determined for a condition other than the predetermined condition; is generated by (12) A summary generation system according to (12). (14) The summary model is calculating the number of conditional statements that indicate conditions that should be satisfied by the output of the large-scale language model, which are included in the training data, and calculating the two low-rank matrices so that the greater the number of calculated conditional statements, the greater the rank of each of the two low-rank matrices; is generated by A summary generation system according to (12) or (13). [Explanation of symbols]

[0109] 1 Summarization system, N network, 10 Summary generation system, 12 Summary model generation system, 14 Training data generation system, 10a, 12a, 14a Control unit, 10b, 12b, 14b Memory unit, 10c, 12c, 14c Communication unit, 10d, 12d, 14d Operation unit, 10e, 12e, 14e Display unit, 100, 140 Product database memory unit, 102 Target product title acquisition unit, 104 Target attention attribute acquisition unit, 106 Target summary acquisition unit, 108, 126 Model memory unit, 120, 148 Training database memory unit, 122 Training data acquisition unit, 124 Summary model generation unit, 1240 Low rank matrix calculation unit, 1242 Low rank matrix update unit, 1244 Parameter matrix calculation unit, 1246 Summary model output unit, 141 Training product title acquisition unit, 142 Importance database storage unit, 143 importance acquisition unit, 144 attention 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 contents of each of multiple attention attributes for the target product, I3, I4 directive statements, C3, C4 conditional statements, S3 target summary, T4 training product title, FA4 contents of each of multiple attention attributes for the training product, L4 correct answer data, A attribute, FA attention attribute, L correct answer data, m pre-trained large-scale language model, W pre-trained parameter matrix, W' re-trained parameter matrix, X, Y low-rank matrices, DB1 product database, DB2 importance database, Im importance.

Claims

1. A summary generation system for generating a summary of product information text that indicates 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 that is the target of the summary; 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 possessed by a target product indicated by the target product information text; a target summary acquisition means for inputting the target product information text and the contents of each of the one or more focus attributes of 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; and The summary model is generated by learning based on learning data including a learning product information text, which is the product information text for learning, the contents of each of the one or more focus attributes of the learning product, which is the product indicated by the learning product information text, and correct answer data, which is a summary of the learning product information text. Summary generation system.

2. a product data storage means for storing identification information of each of the one or more products and the content of each of the one or more noteworthy attributes of the product in association with each other; and The target attention attribute acquisition means acquires identification information of the target product, and acquires the content of each of the one or more attention attributes of the target product stored by the product data storage means based on the identification information. The summary generation system of claim 1 .

3. The learning data is Obtaining the learning material information text; Acquiring the content of each of the one or more focus attributes for the learning material; Extracting character strings corresponding to each of the one or more noteworthy attributes of the learning material from the learning material information text based on the contents of each of the one or more noteworthy attributes of the learning material; outputting a summary shorter than the learning material information text based on the extracted character string; is generated by The summary generation system according to claim 1 or 2.

4. The learning data is obtaining an importance level for each of the plurality of attributes; acquiring content of each of the one or more attention attributes for the learning material, the one or more attention attributes being those of the plurality of attributes whose importance is equal to or greater than a predetermined value; is generated by The summary generation system according to claim 3 .

5. The learning data is outputting the summary by arranging the extracted character string so that the number of characters is equal to or less than a predetermined number of characters that is shorter than the learning material information text; is generated by The summary generation system according to claim 3 .

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

7. the training data further includes an instruction statement that instructs the output content of the large-scale language model; The summary generation system of claim 6.

8. the training data further includes a conditional statement indicating a condition that an output of the large-scale language model should satisfy; The summary generation system of claim 6.

9. the conditional statement indicates that the output of the large-scale language model should include a character string corresponding to the content of each of the one or more focus attributes of the training product. The summary generation system of claim 8.

10. the conditional statement indicates that the output of the large-scale language model should be a character string that is shorter than the training product information text and has a predetermined number of characters or less. The summary generation system of claim 8.

11. the summary model is generated by updating a plurality of update target parameters for the plurality of pre-trained parameters based on the training data, without updating a plurality of pre-trained parameters included in the pre-trained large-scale language model. The summary generation system of claim 8.

12. The summary model is For a pre-trained parameter matrix that is a matrix having the plurality of pre-trained parameters as elements, calculating two low-rank matrices each having the plurality of update target parameters as elements and each having a rank smaller than the number of rows and the number of columns of the pre-trained parameter matrix; updating each of the low-rank matrices based on the training data without updating the pre-trained parameter matrix; calculating a re-trained parameter matrix having a plurality of re-trained parameters as elements based on the two updated low-rank matrices and the pre-trained parameter matrix; outputting the large-scale language model including the retrained parameter matrix as the summary model; is generated by The summary generation system of claim 11.

13. The summary model is determining whether the conditional statement indicates a predetermined condition, and if it is determined that the conditional statement indicates the predetermined condition, calculating the two low-rank matrices so that each of the low-rank matrices has a rank greater than a rank determined for a condition other than the predetermined condition; is generated by The summary generation system of claim 12.

14. The summary model is calculating the number of conditional statements that indicate conditions that should be satisfied by the output of the large-scale language model, which are included in the training data, and calculating the two low-rank matrices so that the greater the number of calculated conditional statements, the greater the rank of each of the two low-rank matrices; is generated by The summary generation system of claim 12.

15. A training data generation system for generating training data used to generate a summary model that receives product information text indicating information about a product as input and outputs a summary of the product information text, the system comprising: a learning material information text acquisition means for acquiring a learning material information text, which is the learning material information text; A focus attribute acquisition means for acquiring the content of one or more focus attributes, which are at least a part of the attributes of the learning material indicated by the learning material information text; A character string extraction means for extracting character strings corresponding to the content of each of the one or more attention attributes of the learning material from the learning material information text based on the content of each of the one or more attention attributes of the learning material; a summary output means for outputting a summary that is shorter than the learning material information text based on the extracted character string; A training data generation system having the above.

16. A summary generation method for generating a summary of product information text that indicates information about a product, comprising: a target product information text acquisition step of acquiring a target product information text, which is the product information text that is the target of the summary; a target attention attribute acquisition step of acquiring, for a target product which is the product indicated by the target product information text, the contents of one or more attention attributes which are at least a part of a plurality of attributes possessed by the product; a target summary acquisition step of inputting the target product information text and the contents of each of the one or more focus attributes of 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; and The summary model is generated by learning based on learning data including a learning product information text, which is the product information text for learning, the contents of each of the one or more focus attributes of the learning product, which is the product indicated by the learning product information text, and correct answer data, which is a summary of the learning product information text. Summary generation method.

17. A training data generation method for generating training data to be used in generating a summary model that receives product information text indicating information about a product as input and outputs a summary of the product information text, the method comprising: A learning material information text acquisition step of acquiring a learning material information text, which is the learning material information text; A focus attribute acquisition step of acquiring the content of one or more focus attributes, which are at least a part of the attributes of the learning material indicated by the learning material information text; A character string extraction step of extracting character strings corresponding to the content of each of the one or more attention attributes of the learning material from the learning material information text based on the content of each of the one or more attention attributes of the learning material; a summary output step of outputting a summary that is shorter than the learning material information text based on the extracted character string; A training data generation method comprising:

18. A program for causing a computer to function to generate a summary of product information text that indicates information about a product, a target product information text acquisition means for acquiring a target product information text, which is the product information text that is the target of the summary; a target attention attribute acquisition means for acquiring, for a target product which is the product indicated by the target product information text, the contents of one or more attention attributes which are at least a part of a plurality of attributes possessed by the product; a target summary acquisition means for inputting the target product information text and the contents of each of the one or more focus attributes of 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; causing the computer to function as The summary model is generated by learning based on learning data including a learning product information text, which is the product information text for learning, the contents of each of the one or more focus attributes of the learning product, which is the product indicated by the learning product information text, and correct answer data, which is a summary of the learning product information text. program.

19. A program for causing a computer to function to generate training data used to generate a summary model that receives product information text indicating information about a product as input and outputs a summary of the product information text, the program comprising: a learning material information text acquisition means for acquiring a learning material information text, which is the learning material information text; An attention attribute acquisition means for acquiring the content of one or more attention attributes, which are at least a part of the multiple attributes possessed by the learning material indicated by the learning material information text; a character string extraction means for extracting character strings corresponding to the content of each of the one or more attention attributes of the learning material from the learning material information text based on the content of each of the one or more attention attributes of the learning material; a summary output means for outputting a summary shorter than the learning material information text based on the extracted character string; A program for causing the computer to function as a

Images