User recommendation method and system based on enterprise product portrait analysis

Abstract

The present application provides a user recommendation method and system based on enterprise product portrait analysis. First, a first enterprise product portrait analysis network is generated based on frequent business feedback texts in a first training feedback text sequence to improve the analysis accuracy of the enterprise product portrait analysis network for frequent business feedback texts. Then, the first enterprise product portrait analysis network is advanced learned based on reference training feedback texts in the first training feedback text sequence, the second training feedback text sequence, and the third training feedback text sequence to improve the analysis accuracy of sparse business feedback texts. While improving the analysis accuracy of the enterprise product portrait analysis network for frequent business feedback texts, the analysis accuracy of the enterprise product portrait analysis network for sparse business feedback texts is also greatly improved, thereby improving the accuracy of enterprise product portrait analysis, and further improving the accuracy of subsequent recommendation of enterprise product portrait groupings to users.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and more specifically, to a user recommendation method and system based on enterprise product portrait analysis. Background Art

[0002] In the era of big data in smart parks, various enterprise products are flooding the Internet, and enterprise product portraits can reflect the user experience feedback of users on the Internet in the process of using related enterprise products. In the actual process of mining enterprise product portraits, the user's attributes, behaviors and expected data are often transformed and linked in the most simple and life-like words. The core of enterprise product portraits is to label enterprise products, that is, to abstract each specific user experience information of enterprise products into labels, and use these labels to concretize enterprise products, so as to provide users with targeted services. In the process of relevant users providing various text feedback on enterprise products, how to accurately recommend enterprise product portrait groups to users is a technical problem that needs to be solved in the relevant technical field. Summary of the invention

[0003] In view of this, the purpose of this application is to provide a user recommendation method and system based on enterprise product portrait analysis.

[0004] According to a first aspect of the present application, a user recommendation method based on enterprise product portrait analysis is provided, which is applied to a server system, and the method includes:

[0005] Acquire a first training feedback text sequence, a second training feedback text sequence, and a third training feedback text sequence for conversation feedback by different users in the process of using an enterprise product, wherein the training feedback texts in the first training feedback text sequence are frequent business feedback texts, the training feedback texts in the second training feedback text sequence are sparse business feedback texts, and the third training feedback text sequence includes frequent business feedback texts and sparse business feedback texts;

[0006] Optimizing the initialization weight parameters of the basic enterprise product portrait analysis network according to the frequent business feedback text in the first training feedback text sequence to generate a first enterprise product portrait analysis network;

[0007] The first training feedback text sequence, the second training feedback text sequence and the third training feedback text sequence are traversed to generate a reference training feedback text sequence required for advanced learning. When the first enterprise product portrait analysis network is advanced-learned based on the reference training feedback text in the reference training feedback text sequence, after removing the learning change rate corresponding to the reference training feedback text belonging to the third training feedback text sequence and including frequent business keywords, the network weight information of the first enterprise product portrait analysis network is optimized based on the learning change rate corresponding to the remaining reference training feedback text in the reference training feedback text sequence, until the traversal is finished, an enterprise product portrait analysis network for performing enterprise product portrait analysis on frequent business feedback texts and sparse business feedback texts is generated, and the enterprise product portrait analysis network is used to identify the target enterprise product portrait of the candidate feedback text of any target user and recommend the target user to the product portrait group where the target enterprise product portrait is located.

[0008] In a possible implementation of the first aspect, the initializing weight parameter optimization of the basic enterprise product portrait analysis network according to the frequent business feedback text in the first training feedback text sequence to generate the first enterprise product portrait analysis network includes:

[0009] Obtaining frequent business feedback text and corresponding enterprise product portrait annotation data from the first training feedback text sequence;

[0010] According to the semantic encoder in the basic enterprise product portrait analysis network, extract the text semantic encoding vector corresponding to each of the frequent business feedback texts;

[0011] According to the fully connected output unit in the basic enterprise product portrait analysis network, the predicted enterprise product portrait data corresponding to each of the frequent business feedback texts is obtained according to the text semantic encoding vector;

[0012] Constructing a first Loss value based on the enterprise product portrait annotation data and the predicted enterprise product portrait data;

[0013] After using the network weight information when the first Loss value converges as the network weight information for the basic enterprise product portrait analysis network optimization, return to the step of obtaining frequent business feedback text from the first training feedback text sequence to continue training until the convergence requirements are met.

[0014] In a possible implementation of the first aspect, generating a reference training feedback text sequence required for advanced learning from the first training feedback text sequence, the second training feedback text sequence, and the third training feedback text sequence includes:

[0015] Get the set extraction parameters;

[0016] The first training feedback text sequence, the second training feedback text sequence and the third training feedback text sequence are extracted respectively according to the set extraction parameters to generate the reference training feedback text sequence required for advanced learning.

[0017] In a possible implementation of the first aspect, when performing advanced learning on the first enterprise product portrait analysis network based on the reference training feedback text in the reference training feedback text sequence, after removing the learning change rate corresponding to the reference training feedback text belonging to the third training feedback text sequence and including frequent business keywords, optimizing the network weight information of the first enterprise product portrait analysis network based on the learning change rate corresponding to the remaining reference training feedback text in the reference training feedback text sequence, including:

[0018] Loading each reference training feedback text in the reference training feedback text sequence into the first enterprise product portrait analysis network to generate predicted enterprise product portrait data corresponding to each reference training feedback text;

[0019] Determine the advanced learning Loss value based on the enterprise product portrait annotation data corresponding to each of the reference training feedback texts and the predicted enterprise product portrait data;

[0020] When the advanced learning Loss value converges, obtaining the learning change rate corresponding to each of the reference training feedback texts;

[0021] From the learning change rates corresponding to the reference training feedback texts, after removing the learning change rates corresponding to the reference training feedback texts that belong to the third training feedback text sequence and include frequent business keywords, the network weight information of the first enterprise product portrait analysis network is optimized according to the learning change rates corresponding to the remaining reference training feedback texts in the reference training feedback text sequence.

[0022] In a possible implementation of the first aspect, loading each reference training feedback text in the reference training feedback text sequence into the first enterprise product portrait analysis network to generate predicted enterprise product portrait data corresponding to each reference training feedback text includes:

[0023] Extracting text semantic encoding vectors corresponding to each of the reference training feedback texts according to the semantic encoder in the first enterprise product portrait analysis network;

[0024] According to the fully connected output unit in the first enterprise product portrait analysis network, the predicted enterprise product portrait data corresponding to each reference training feedback text is obtained according to the text semantic encoding vector.

[0025] In a possible implementation of the first aspect, determining the advanced learning Loss value based on the enterprise product portrait annotation data corresponding to each of the reference training feedback texts and the predicted enterprise product portrait data includes:

[0026] Determine a first Loss value based on the enterprise product portrait annotation data corresponding to the reference training feedback text belonging to the first training feedback text sequence in the reference training feedback text sequence and the predicted enterprise product portrait data;

[0027] Determine a second Loss value based on the enterprise product portrait annotation data corresponding to the reference training feedback text belonging to the second training feedback text sequence in the reference training feedback text sequence and the predicted enterprise product portrait data;

[0028] Determine a third Loss value based on the enterprise product portrait annotation data corresponding to the reference training feedback text belonging to the third training feedback text sequence in the reference training feedback text sequence and the predicted enterprise product portrait data;

[0029] Get the set weight distribution information;

[0030] The first Loss value, the second Loss value and the third Loss value are weightedly fused according to the weight allocation information to generate the advanced learning Loss value.

[0031] In a possible implementation of the first aspect, when performing advanced learning on the first enterprise product portrait analysis network based on the reference training feedback text in the reference training feedback text sequence, after removing the learning change rate corresponding to the reference training feedback text belonging to the third training feedback text sequence and including frequent business keywords, optimizing the network weight information of the first enterprise product portrait analysis network based on the learning change rate corresponding to the remaining reference training feedback text in the reference training feedback text sequence, including:

[0032] When, in the reference training feedback text sequence, the reference training feedback text extracted from the third training feedback text sequence includes a sparse business feedback text, optimizing the network weight information of the first enterprise product portrait analysis network according to the learning change rates corresponding to the reference training feedback text belonging to the third training feedback text sequence and including the sparse business keywords, the reference training feedback text belonging to the first training feedback text sequence, and the reference training feedback text belonging to the second training feedback text sequence;

[0033] When the reference training feedback text extracted from the third training feedback text sequence in the reference training feedback text sequence does not include sparse business feedback text, the network weight information of the first enterprise product portrait analysis network is optimized based on the learning change rate corresponding to the reference training feedback text belonging to the first training feedback text sequence and the reference training feedback text belonging to the second training feedback text sequence.

[0034] In a possible implementation of the first aspect, the method further includes:

[0035] Acquire a test feedback text and a test enterprise product portrait corresponding to the test feedback text, wherein the test feedback text includes frequent business keywords or sparse business keywords;

[0036] Obtaining a test text semantic vector corresponding to the test feedback text through the enterprise product portrait analysis network generated by advanced learning;

[0037] Obtaining the annotated text semantic vector corresponding to the product portrait of the test enterprise;

[0038] When the representation vector distance between the test text semantic vector and the annotated text semantic vector is less than a set distance, it is determined that the enterprise product portrait analysis network meets the target deployment conditions.

[0039] In a possible implementation of the first aspect, the method further includes:

[0040] Acquire candidate feedback texts, where the candidate feedback texts include frequent business keywords or sparse business keywords;

[0041] The enterprise product portrait analysis network generated by advanced learning is used to obtain the enterprise product portrait corresponding to the candidate feedback text.

[0042] According to the second aspect of the present application, a server system is provided, which includes a processor and a readable storage medium, wherein the readable storage medium stores a program, which, when executed by the processor, implements the aforementioned user recommendation method based on enterprise product portrait analysis.

[0043] According to the third aspect of the present application, a computer-readable storage medium is provided, in which computer-executable instructions are stored. When the computer-executable instructions are monitored to be executed, the aforementioned user recommendation method based on enterprise product portrait analysis is implemented.

[0044] According to any of the above aspects, in the present application, the weight parameters of the basic enterprise product portrait analysis network are first initialized and optimized based on the frequent business feedback text in the first training feedback text sequence to generate a first enterprise product portrait analysis network, thereby improving the analysis accuracy of the enterprise product portrait analysis network for frequent business feedback text; then, the first enterprise product portrait analysis network is advanced learned based on the reference training feedback text in the first training feedback text sequence, the second training feedback text sequence, and the third training feedback text sequence. The first enterprise product portrait analysis network is trained to distinguish the representation vectors of frequent businesses based on the frequent business feedback text in the first training feedback text sequence, to distinguish the representation vectors of sparse businesses based on the sparse business feedback text in the second training feedback text sequence, and to distinguish the representation vectors of sparse businesses based on the third training feedback text sequence. Frequent business feedback text and sparse business feedback text are trained to distinguish the representation vectors of frequent business and sparse business, and the frequent business feedback text included in the third training feedback text sequence can be used to constrain the first enterprise product portrait analysis network to learn the representation vector of sparse business, thereby expanding the representation vector range of sparse business, enhancing the representation vector distance between different categories of text in sparse business determined by the enterprise product portrait analysis network, and improving the analysis accuracy of sparse business feedback text; in this way, while improving the analysis accuracy of the enterprise product portrait analysis network for frequent business feedback text, it also greatly improves the analysis accuracy of the enterprise product portrait analysis network for sparse business feedback text, thereby improving the accuracy of enterprise product portrait analysis, and then improving the accuracy of subsequent recommendation of enterprise product portrait grouping to users. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0046] Figure 1 A flowchart of a user recommendation method based on enterprise product portrait analysis provided in an embodiment of the present application;

[0047] Figure 2 A schematic diagram of the component structure of a server system provided in an embodiment of the present application for implementing the above-mentioned user recommendation method based on enterprise product portrait analysis is shown. DETAILED DESCRIPTION

[0048] In order to enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without creative work should fall within the scope of protection of the present application.

[0049] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0050] Figure 1 The flowchart of the user recommendation method based on enterprise product portrait analysis provided by the embodiment of the present application is shown. It should be understood that in other embodiments, the order of some steps in the user recommendation method based on enterprise product portrait analysis of this embodiment can be interchanged according to the need for annotation, or some steps can be omitted or deleted. The detailed steps of the user recommendation method based on enterprise product portrait analysis are introduced as follows.

[0051] Step S102, obtaining a first training feedback text sequence, a second training feedback text sequence, and a third training feedback text sequence of conversation feedback by different users in the process of using the enterprise product, wherein the training feedback text in the first training feedback text sequence is a frequent business feedback text, the training feedback text in the second training feedback text sequence is a sparse business feedback text, and the third training feedback text sequence includes frequent business feedback text and sparse business feedback text.

[0052] Among them, the first training feedback text sequence, the second training feedback text sequence and the third training feedback text sequence are all text sequences used to train the enterprise product portrait analysis network. Frequent business feedback text means that the business type corresponding to the keywords in the text is a frequent business (that is, the business whose user usage frequency is greater than the first set frequency), and sparse business feedback text means that the business type corresponding to the keywords in the text is a sparse business (that is, the business whose user usage frequency is lower than the second set frequency). Enterprise product portraits can refer to user experience evaluation information of enterprise products, for example, it can be used to describe the user experience evaluation of enterprise products, including but not limited to user preference fields, aversion fields, excitement fields, etc. for enterprise products. Feedback text can refer to text data generated by users in various online conversations during the process of using enterprise products.

[0053] In this embodiment, by constructing a first training feedback text sequence, a second training feedback text sequence and a third training feedback text sequence, the training feedback text in the first training feedback text sequence is a frequent business feedback text, which can be used for enterprise product portrait analysis network training to distinguish representation vectors for frequent businesses; the training feedback text in the second training feedback text sequence is a sparse business feedback text, which can be used for enterprise product portrait analysis network training to distinguish representation vectors for sparse businesses; the third training feedback text sequence includes frequent business feedback text and sparse business feedback text, which can be used for enterprise product portrait analysis network training to distinguish representation vectors for frequent businesses and sparse businesses, and the frequent business feedback text included in the third training feedback text sequence can be used to constrain the enterprise product portrait analysis network to learn the representation vectors of sparse businesses, so as to derive the representation vector interval of sparse businesses.

[0054] In an alternative implementation, a first training feedback text sequence and a second training feedback text sequence are obtained, and the union of the first training feedback text sequence and the second training feedback text sequence is used as a third training feedback text sequence. Alternatively, the training feedback text in the third training feedback text sequence does not overlap or partially overlaps with the training feedback text in the first training feedback text sequence and the second training feedback text sequence, and the third training feedback text sequence includes frequent business feedback text and sparse business feedback text.

[0055] Step S104, optimizing the initialization weight parameters of the basic enterprise product portrait analysis network according to the frequent business feedback text in the first training feedback text sequence, and generating a first enterprise product portrait analysis network.

[0056] The basic enterprise product image analysis network may be an enterprise product image analysis network that has not been trained with the training feedback text in the first training feedback text sequence, the second training feedback text sequence, or the third training feedback text sequence. The first enterprise product image analysis network may be an enterprise product image analysis network that has been trained with the training feedback text in the first training feedback text sequence.

[0057] In an alternative implementation, a first training feedback text sequence, a second training feedback text sequence, and a third training feedback text sequence are obtained, and the weight parameters of the basic enterprise product portrait analysis network are optimized based on the frequent business feedback text in the first training feedback text sequence to generate the first enterprise product portrait analysis network. Since the first training feedback text sequence only includes frequent business feedback text, the weight parameters of the basic enterprise product portrait analysis network are optimized based on the frequent business feedback text in the first training feedback text sequence, so that the enterprise product portrait analysis network is trained to distinguish the significant content of frequent business, thereby ensuring the analysis accuracy of the enterprise product portrait analysis network for frequent business feedback text.

[0058] In an alternative implementation, the initialization weight parameters of the basic enterprise product portrait analysis network are optimized, and the network weight information of the semantic encoder and the fully connected output unit is optimized.

[0059] In an alternative implementation, more than one frequent business feedback text is obtained from the first training feedback text sequence, and more than one frequent business feedback text is input into the basic enterprise product portrait analysis network. The text semantic encoding vector corresponding to each frequent business feedback text is extracted through the semantic encoder in the basic enterprise product portrait analysis network, and the predicted enterprise product portrait data corresponding to each frequent business feedback text is obtained according to the text semantic encoding vector through the fully connected output unit in the basic enterprise product portrait analysis network. A first Loss value is constructed based on the predicted enterprise product portrait data corresponding to each frequent business feedback text and the enterprise product portrait annotation data, and the learning change rate of this round of learning is determined according to the first Loss value. After optimizing the network weight information of the semantic encoder and the fully connected output unit according to the learning change rate, the training is returned to the step of obtaining more than one frequent business feedback text from the first training feedback text sequence to continue until the convergence requirement is met.

[0060] Step S106, traverse the first training feedback text sequence, the second training feedback text sequence and the third training feedback text sequence to generate a reference training feedback text sequence required for advanced learning. When performing advanced learning on the first enterprise product portrait analysis network based on the reference training feedback text in the reference training feedback text sequence, after removing the learning change rate corresponding to the reference training feedback text belonging to the third training feedback text sequence and including frequent business keywords, optimize the network weight information of the first enterprise product portrait analysis network based on the learning change rate corresponding to the remaining reference training feedback text in the reference training feedback text sequence, until the traversal is completed, and an enterprise product portrait analysis network for performing enterprise product portrait analysis on frequent business feedback texts and sparse business feedback texts is generated.

[0061] In an alternative implementation, a first training feedback text sequence, a second training feedback text sequence, and a third training feedback text sequence are obtained, and the weight parameters of the basic enterprise product portrait analysis network are optimized based on the frequent business feedback text in the first training feedback text sequence to generate the first enterprise product portrait analysis network, and then the reference training feedback text sequence is obtained from the first training feedback text sequence, the second training feedback text sequence, and the third training feedback text sequence, and the first enterprise product portrait analysis network is advancedly learned based on the reference training feedback text in the reference training feedback text sequence. The first enterprise product portrait analysis network trains a representation vector for distinguishing frequent businesses based on the frequent business feedback text in the first training feedback text sequence, trains a representation vector for distinguishing sparse businesses based on the sparse business feedback text in the second training feedback text sequence, and trains a representation vector for distinguishing frequent businesses from sparse businesses based on the frequent business feedback text and sparse business feedback text in the third training feedback text sequence, and the frequent business feedback text included in the third training feedback text sequence can be used to constrain the learning of the representation vector of the sparse business by the first enterprise product portrait analysis network to derive the representation vector interval of the sparse business.

[0062] In an alternative implementation, advanced learning is performed on the first enterprise product portrait analysis network, and the network weight information of the semantic encoder and the fully connected output unit is optimized.

[0063] In an alternative implementation, the first training feedback text sequence, the second training feedback text sequence and the third training feedback text sequence are extracted to obtain a reference training feedback text sequence, the reference training feedback text in the reference training feedback text sequence is input into the first enterprise product portrait analysis network, and the text semantic encoding vector corresponding to each reference training feedback text is extracted based on the semantic encoder in the first enterprise product portrait analysis network, and the predicted enterprise product portrait data corresponding to each reference training feedback text is obtained based on the text semantic encoding vector through the fully connected output unit in the first enterprise product portrait analysis network. In an alternative implementation, an advanced learning Loss value can be constructed based on the predicted enterprise product portrait data and enterprise product portrait annotation data corresponding to the reference training feedback text, and the learning change rate of this round of learning is determined based on the advanced learning Loss value, and the learning change rate corresponding to the reference training feedback text belonging to the third training feedback text sequence and including frequent business keywords is removed from the learning change rate, and the network weight information of the first enterprise product portrait analysis network is optimized based on the learning change rate corresponding to the remaining reference training feedback text in the reference training feedback text sequence, and then the step of extracting the first training feedback text sequence, the second training feedback text sequence, and the third training feedback text sequence is returned to continue training until the convergence requirements are met. Since the learning change rate corresponding to the reference training feedback text of the third training feedback text sequence and including frequent business keywords is removed, the reference training feedback text of the third training feedback text sequence and including frequent business keywords is only used to derive the representation vector interval of sparse business, and does not affect the optimization of network weight information.

[0064] In an alternative implementation, the learning change rate of this round of learning is determined based on the advanced learning Loss value, and the learning change rate corresponding to the reference training feedback text of the third training feedback text sequence and including frequent business keywords in the learning change rate is hidden, so that the learning change rate corresponding to the reference training feedback text of the third training feedback text sequence and including frequent business keywords does not participate in the optimization of network weight information.

[0065] Therefore, while improving the analysis accuracy of the enterprise product portrait analysis network for frequent business feedback texts, it can also greatly improve the analysis accuracy of the enterprise product portrait analysis network for sparse business feedback texts.

[0066] Based on the above steps, the weight parameters of the basic enterprise product portrait analysis network are first initialized and optimized according to the frequent business feedback text in the first training feedback text sequence to generate the first enterprise product portrait analysis network, thereby improving the analysis accuracy of the enterprise product portrait analysis network for frequent business feedback text; then, the first enterprise product portrait analysis network is advancedly learned according to the reference training feedback text in the first training feedback text sequence, the second training feedback text sequence, and the third training feedback text sequence. The first enterprise product portrait analysis network is trained to distinguish the representation vector of frequent business according to the frequent business feedback text in the first training feedback text sequence, and is trained to distinguish the representation vector of sparse business according to the sparse business feedback text in the second training feedback text sequence. Vector, based on the frequent business feedback text and sparse business feedback text in the third training feedback text sequence, training is used to distinguish between frequent business and sparse business characterization vectors, and the frequent business feedback text included in the third training feedback text sequence can be used to constrain the first enterprise product portrait analysis network to learn the characterization vector of sparse business, so as to derive the characterization vector interval of sparse business, reduce the characterization vector distance between different categories of text in the sparse business calculated by the enterprise product portrait analysis network, and improve the analysis accuracy of sparse business feedback text; in this way, while improving the analysis accuracy of the enterprise product portrait analysis network for frequent business feedback text, it also greatly improves the analysis accuracy of the enterprise product portrait analysis network for sparse business feedback text.

[0067] In an alternative implementation, initialization weight parameter optimization is performed on the basic enterprise product portrait analysis network according to the frequent business feedback text in the first training feedback text sequence to generate the first enterprise product portrait analysis network, including:

[0068] Step S202: Obtain frequent business feedback text and corresponding enterprise product portrait annotation data from the first training feedback text sequence.

[0069] In an alternative implementation, more than one frequent business feedback text is obtained from the first training feedback text sequence, and the more than one frequent business feedback text is taken as a group to perform an initialization weight parameter optimization on the basic enterprise product portrait analysis network.

[0070] Step S204: input the frequent business feedback text obtained from the first training feedback text sequence into the basic enterprise product portrait analysis network to generate predicted enterprise product portrait data.

[0071] In an alternative implementation, the text semantic encoding vector corresponding to each frequent business feedback text is extracted through the semantic encoder in the basic enterprise product portrait analysis network; and the predicted enterprise product portrait data corresponding to each frequent business feedback text is obtained based on the text semantic encoding vector through the fully connected output unit in the basic enterprise product portrait analysis network.

[0072] In an alternative implementation, the text semantic encoding vector corresponding to each frequent business feedback text is extracted through the semantic encoder in the basic enterprise product portrait analysis network, and each text semantic encoding vector is passed to the fully connected output unit in the basic enterprise product portrait analysis network. The fully connected output unit performs regression analysis on each text semantic encoding vector to generate predicted enterprise product portrait data corresponding to each frequent business feedback text.

[0073] Step S206, constructing a first Loss value based on the enterprise product portrait annotation data and the predicted enterprise product portrait data.

[0074] In an alternative implementation, the first Loss value is determined based on the difference between the enterprise product portrait annotation data corresponding to each frequent business feedback text and the predicted enterprise product portrait data.

[0075] Step S208, after using the network weight information when the first Loss value converges as the network weight information for basic enterprise product portrait analysis network optimization, return to the step of obtaining frequent business feedback text from the first training feedback text sequence to continue training until the convergence requirements are met.

[0076] In an alternative implementation, based on the direction of minimizing the first Loss value, the learning change rate corresponding to the current round of learning is obtained according to the learning change rate reduction algorithm, and the network weight information of the semantic encoder and the fully connected output unit is optimized according to the learning change rate.

[0077] In this embodiment, the initialization weight parameters of the basic enterprise product portrait analysis network are optimized based on the frequent business feedback text in the first training feedback text sequence, so that the enterprise product portrait analysis network training is used to distinguish the significant features of frequent business, thereby ensuring the analysis accuracy of the enterprise product portrait analysis network on frequent business feedback text.

[0078] In an alternative implementation, a reference training feedback text sequence required for advanced learning is generated from a first training feedback text sequence, a second training feedback text sequence, and a third training feedback text sequence, including: obtaining set extraction parameters; extracting the first training feedback text sequence, the second training feedback text sequence, and the third training feedback text sequence according to the set extraction parameters, respectively, to generate a reference training feedback text sequence required for advanced learning.

[0079] In an alternative implementation, the first training feedback text sequence, the second training feedback text sequence, and the third training feedback text sequence are extracted according to the set extraction parameters, the frequent business feedback text extracted from the first training feedback text sequence is used as the first subsequence, the sparse business feedback text extracted from the second training feedback text sequence is used as the second subsequence, the frequent business feedback text and / or the sparse business feedback text extracted from the third training feedback text sequence is used as the third subsequence, and the first subsequence, the second subsequence, and the third subsequence are merged into a reference training feedback text sequence. The reference training feedback text in the reference training feedback text sequence is input into the first enterprise product portrait analysis network to perform multi-task advanced learning on the first enterprise product portrait analysis network.

[0080] In an alternative implementation, when performing advanced learning on the first enterprise product portrait analysis network based on the reference training feedback text in the reference training feedback text sequence, after removing the learning change rate corresponding to the reference training feedback text belonging to the third training feedback text sequence and including the frequent business keywords, the network weight information of the first enterprise product portrait analysis network is optimized based on the learning change rate corresponding to the remaining reference training feedback text in the reference training feedback text sequence, including:

[0081] Step S302, performing advanced learning on the first enterprise product portrait analysis network according to each reference training feedback text in the reference training feedback text sequence, and obtaining the learning change rate corresponding to each reference training feedback text during the advanced learning process.

[0082] In an alternative implementation, advanced learning is performed on the first enterprise product portrait analysis network based on each reference training feedback text in a reference training feedback text sequence, and the learning change rate corresponding to each reference training feedback text is obtained during the advanced learning process, including: inputting each reference training feedback text in the reference training feedback text sequence into the first enterprise product portrait analysis network to generate predicted enterprise product portrait data corresponding to each reference training feedback text; determining an advanced learning Loss value based on the enterprise product portrait annotation data and predicted enterprise product portrait data corresponding to each reference training feedback text; and obtaining the learning change rate corresponding to each reference training feedback text when the advanced learning Loss value converges.

[0083] In an alternative implementation, each reference training feedback text in a reference training feedback text sequence is input into a first enterprise product portrait analysis network to generate predicted enterprise product portrait data corresponding to each reference training feedback text, including: extracting a text semantic encoding vector corresponding to each reference training feedback text through a semantic encoder in the first enterprise product portrait analysis network; and obtaining predicted enterprise product portrait data corresponding to each reference training feedback text based on the text semantic encoding vector through a fully connected output unit in the first enterprise product portrait analysis network.

[0084] In an alternative implementation, the advanced learning Loss value is determined based on the difference between the enterprise product portrait annotation data corresponding to each reference training feedback text and the predicted enterprise product portrait data.

[0085] In an alternative implementation, based on the enterprise product portrait annotation data and predicted enterprise product portrait data corresponding to each reference training feedback text, an advanced learning Loss value is determined, including: determining a first Loss value based on the enterprise product portrait annotation data and predicted enterprise product portrait data corresponding to the reference training feedback text belonging to the first training feedback text sequence in the reference training feedback text sequence; determining a second Loss value based on the enterprise product portrait annotation data and predicted enterprise product portrait data corresponding to the reference training feedback text belonging to the second training feedback text sequence in the reference training feedback text sequence; determining a third Loss value based on the enterprise product portrait annotation data and predicted enterprise product portrait data corresponding to the reference training feedback text belonging to the third training feedback text sequence in the reference training feedback text sequence; obtaining set weight distribution information; and weightedly fusing the first Loss value, the second Loss value, and the third Loss value according to the weight distribution information to generate an advanced learning Loss value.

[0086] In an alternative implementation, weight allocation information may be set respectively for the reference training feedback text (first subsequence) belonging to the first training feedback text sequence in the reference training feedback text sequence, the reference training feedback text (second subsequence) belonging to the second training feedback text sequence in the reference training feedback text sequence, and the reference training feedback text (third subsequence) belonging to the third training feedback text sequence in the reference training feedback text sequence.

[0087] In an alternative implementation, based on the direction of minimizing the advanced learning Loss value, the learning change rate corresponding to the current round of learning is obtained according to the learning change rate reduction algorithm, and the learning change rate corresponding to each reference training feedback text is determined from the learning change rate.

[0088] Therefore, the learning change rate corresponding to each reference training feedback text can be determined from the learning change rate corresponding to the current round of learning.

[0089] Step S304, after removing the learning change rate corresponding to the reference training feedback text that belongs to the third training feedback text sequence and includes frequent business keywords from the learning change rate corresponding to each reference training feedback text, optimize the network weight information of the first enterprise product portrait analysis network according to the learning change rate corresponding to the remaining reference training feedback text in the reference training feedback text sequence.

[0090] In an alternative implementation, it is assumed that reference training feedback text 4 and reference training feedback text 5 are reference training feedback texts that belong to the third training feedback text sequence in the reference training feedback text sequence and include frequent business keywords. The text semantic encoding vector corresponding to each reference training feedback text can be extracted, and the predicted enterprise product portrait data corresponding to each reference training feedback text can be obtained based on the text semantic encoding vector. Based on the enterprise product portrait annotation data and predicted enterprise product portrait data corresponding to the reference training feedback text of the first subsequence, the first Loss value is determined, based on the enterprise product portrait annotation data and predicted enterprise product portrait data corresponding to the reference training feedback text of the second subsequence, the second Loss value is determined, based on the enterprise product portrait annotation data and predicted enterprise product portrait data corresponding to the reference training feedback text of the third subsequence, the third Loss value is determined, and the first Loss value, the second Loss value and the third Loss value are weighted and fused according to the weight allocation information to generate an advanced learning Loss value. According to the direction of minimizing the advanced learning Loss value, the learning change rate corresponding to this round of learning is obtained according to the learning change rate reduction algorithm, and the learning change rate corresponding to each reference training feedback text is determined from the learning change rate. From the learning change rates corresponding to each reference training feedback text, after removing the learning change rates corresponding to reference training feedback text 4 and reference training feedback text 5, the network weight information of the first enterprise product portrait analysis network is optimized according to the learning change rates corresponding to reference training feedback text 1, reference training feedback text 2, reference training feedback text 3 and reference training feedback text 6.

[0091] In this embodiment, the first enterprise product portrait analysis network is advanced-learned based on the reference training feedback texts in the first training feedback text sequence, the second training feedback text sequence, and the third training feedback text sequence. The first enterprise product portrait analysis network is trained with the frequent business feedback texts in the first training feedback text sequence to distinguish the representation vectors of frequent businesses, and is trained with the sparse business feedback texts in the second training feedback text sequence to distinguish the representation vectors of sparse businesses. The representation vectors are trained with the frequent business feedback texts and the sparse business feedback texts in the third training feedback text sequence to distinguish the representation vectors of frequent businesses and sparse businesses. The frequent business feedback texts included in the third training feedback text sequence can be used to constrain the learning of the representation vectors of sparse businesses by the first enterprise product portrait analysis network, so as to derive the representation vector range of sparse businesses, expand the representation vector distance between different categories of texts in sparse businesses calculated by the enterprise product portrait analysis network, and improve the analysis accuracy of sparse business feedback texts.

[0092] In an alternative embodiment, the method further includes: fixing the network weight information of the semantic encoder of the first enterprise product portrait analysis network, traversing to obtain the reference training feedback text sequence required for advanced learning from the first training feedback text sequence, the second training feedback text sequence and the third training feedback text sequence, and performing advanced learning on the fully connected output unit of the first enterprise product portrait analysis network according to the reference training feedback text in the reference training feedback text sequence required for advanced learning. During the advanced learning process, after removing the learning change rate corresponding to the reference training feedback text belonging to the third training feedback text sequence and including frequent business keywords, the learning change rate of the fully connected output unit is optimized according to the learning change rate corresponding to the remaining reference training feedback text in the reference training feedback text sequence. network weight information; until the traversal is finished, the network weight information of the fully connected output unit of the first enterprise product portrait analysis network is fixed, and the reference training feedback text sequence required for advanced learning is obtained from the first training feedback text sequence, the second training feedback text sequence and the third training feedback text sequence. The semantic encoder of the first enterprise product portrait analysis network is subjected to advanced learning based on the reference training feedback text in the reference training feedback text sequence required for advanced learning. In the process of advanced learning, after removing the learning change rate corresponding to the reference training feedback text belonging to the third training feedback text sequence and including frequent business keywords, the network weight information of the semantic encoder is optimized based on the learning change rate corresponding to the remaining reference training feedback text in the reference training feedback text sequence.

[0093] In an alternative implementation, when performing advanced learning on the first enterprise product portrait analysis network based on the reference training feedback text in the reference training feedback text sequence, after removing the learning change rate corresponding to the reference training feedback text belonging to the third training feedback text sequence and including frequent business keywords, the network weight information of the first enterprise product portrait analysis network is optimized based on the learning change rate corresponding to the remaining reference training feedback text in the reference training feedback text sequence, including: when in the reference training feedback text sequence, the reference training feedback text extracted from the third training feedback text sequence includes sparse business feedback text, then the network weight information of the first enterprise product portrait analysis network is optimized based on the learning change rate corresponding to the reference training feedback text belonging to the third training feedback text sequence and including sparse business keywords, the reference training feedback text belonging to the first training feedback text sequence, and the reference training feedback text belonging to the second training feedback text sequence; when in the reference training feedback text sequence, the reference training feedback text extracted from the third training feedback text sequence does not include sparse business feedback text, then the network weight information of the first enterprise product portrait analysis network is optimized based on the learning change rate corresponding to the reference training feedback text belonging to the first training feedback text sequence and the reference training feedback text belonging to the second training feedback text sequence.

[0094] In an alternative implementation, when extracting the first training feedback text sequence, the second training feedback text sequence, and the third training feedback text sequence, since the third training feedback text sequence includes frequent business feedback text and sparse business feedback text, the reference training feedback text extracted from the third training feedback text sequence has the following three situations: A. Only frequent business feedback text is extracted from the third training feedback text sequence. B. Only sparse business feedback text is extracted from the third training feedback text sequence. C. Frequent business feedback text and sparse business feedback text are extracted from the third training feedback text sequence. Different network weight information optimization methods are adopted for different extraction situations. For situation A, the network weight information of the first enterprise product portrait analysis network is optimized based on the learning change rate corresponding to the reference training feedback text belonging to the first training feedback text sequence and the reference training feedback text belonging to the second training feedback text sequence in the reference training feedback text sequence. For cases B and C, the network weight information of the first enterprise product portrait analysis network is optimized based on the learning change rate corresponding to the reference training feedback text belonging to the third training feedback text sequence and including sparse business keywords, the reference training feedback text belonging to the first training feedback text sequence, and the reference training feedback text belonging to the second training feedback text sequence in the reference training feedback text sequence. In this way, the reference training feedback text of the third training feedback text sequence and including frequent business keywords is only used to derive the representation vector interval of the sparse business, and does not affect the optimization of the network weight information.

[0095] In an alternative embodiment, the method also includes: obtaining a test feedback text and a test enterprise product portrait corresponding to the test feedback text, the test feedback text including frequent business keywords or sparse business keywords; obtaining a test text semantic vector corresponding to the test feedback text through an enterprise product portrait analysis network generated by advanced learning; obtaining an annotated text semantic vector corresponding to the test enterprise product portrait; when the representation vector distance between the test text semantic vector and the annotated text semantic vector is less than a set distance, determining that the enterprise product portrait analysis network meets the target deployment conditions.

[0096] A further embodiment is described below, comprising the following steps:

[0097] Step S402: Obtain a first training feedback text sequence, a second training feedback text sequence, and a third training feedback text sequence, wherein the training feedback text in the first training feedback text sequence is a frequent business feedback text, the training feedback text in the second training feedback text sequence is a sparse business feedback text, and the third training feedback text sequence includes frequent business feedback text and sparse business feedback text.

[0098] Step S404, obtaining frequent business feedback texts and corresponding enterprise product portrait annotation data from the first training feedback text sequence, extracting the text semantic encoding vectors corresponding to each frequent business feedback text through the semantic encoder in the basic enterprise product portrait analysis network, and obtaining the predicted enterprise product portrait data corresponding to each frequent business feedback text based on the text semantic encoding vector through the fully connected output unit in the basic enterprise product portrait analysis network.

[0099] Step S406, based on the difference between the enterprise product portrait annotation data and the predicted enterprise product portrait data, determine the first Loss value, use the network weight information when the first Loss value converges as the network weight information for optimizing the basic enterprise product portrait analysis network, and then return to the step of obtaining frequent business feedback text from the first training feedback text sequence to continue training until the convergence requirements are met, thereby generating the first enterprise product portrait analysis network.

[0100] Step S408 , extracting the first training feedback text sequence, the second training feedback text sequence and the third training feedback text sequence respectively according to the set extraction parameters to generate a reference training feedback text sequence required for advanced learning.

[0101] Step S410, input each reference training feedback text in the reference training feedback text sequence into the first enterprise product portrait analysis network, extract the text semantic coding vector corresponding to each reference training feedback text through the semantic encoder in the first enterprise product portrait analysis network, and obtain the predicted enterprise product portrait data corresponding to each reference training feedback text based on the text semantic coding vector through the fully connected output unit in the first enterprise product portrait analysis network.

[0102] Step S412: determine a first Loss value based on the enterprise product portrait annotation data and predicted enterprise product portrait data corresponding to the reference training feedback text belonging to the first training feedback text sequence in the reference training feedback text sequence; determine a second Loss value based on the enterprise product portrait annotation data and predicted enterprise product portrait data corresponding to the reference training feedback text belonging to the second training feedback text sequence in the reference training feedback text sequence; determine a third Loss value based on the enterprise product portrait annotation data and predicted enterprise product portrait data corresponding to the reference training feedback text belonging to the third training feedback text sequence in the reference training feedback text sequence; obtain set weight distribution information; perform weighted fusion on the first Loss value, the second Loss value and the third Loss value according to the weight distribution information to generate an advanced learning Loss value.

[0103] Step S414, when the advanced learning Loss value converges, obtain the learning change rate corresponding to each reference training feedback text, and remove the learning change rate corresponding to the reference training feedback text that belongs to the third training feedback text sequence and includes frequent business keywords from the learning change rate corresponding to each reference training feedback text, and optimize the network weight information of the first enterprise product portrait analysis network according to the learning change rate corresponding to the remaining reference training feedback text in the reference training feedback text sequence, and then return to the step of extracting the first training feedback text sequence, the second training feedback text sequence, and the third training feedback text sequence according to the set extraction parameters to continue training until the convergence requirements are met, and an enterprise product portrait analysis network for performing enterprise product portrait analysis on frequent business feedback texts and sparse business feedback texts is generated.

[0104] When the enterprise product portrait analysis network attempts to distinguish between the training feedback texts of frequent and sparse businesses, under the constraints of the training feedback texts of frequent businesses, the enterprise product portrait analysis network can learn significant features for distinguishing between sparse businesses. Therefore, the weight parameters of the basic enterprise product portrait analysis network are initialized and optimized according to the frequent business feedback text in the first training feedback text sequence, so that the enterprise product portrait analysis network is trained to distinguish the significant features of frequent business, thereby ensuring the analysis accuracy of the enterprise product portrait analysis network for frequent business feedback text; then, the first enterprise product portrait analysis network is advancedly learned according to the first training feedback text sequence, the second training feedback text sequence and the third training feedback text sequence. The first enterprise product portrait analysis network is trained to distinguish the representation vector of frequent business according to the frequent business feedback text of the first training feedback text sequence, and is trained to distinguish the representation vector of sparse business according to the sparse business feedback text of the second training feedback text sequence. The representation vector of frequent business and sparse business is trained according to the frequent business feedback text and sparse business feedback text of the third training feedback text sequence, and the frequent business feedback text included in the third training feedback text sequence can be used to constrain the first enterprise product portrait analysis network to learn the significant features for distinguishing sparse business, thereby improving the analysis accuracy of the enterprise product portrait analysis network for sparse business feedback text. The above-mentioned processing method of the enterprise product portrait analysis network not only improves the analysis accuracy of the enterprise product portrait analysis network for frequent business feedback texts, but also greatly improves the analysis accuracy of the enterprise product portrait analysis network for sparse business feedback texts.

[0105] A further embodiment is described below, comprising the following steps:

[0106] Step S502: Acquire candidate feedback texts of target users, where the candidate feedback texts include frequent business keywords or sparse business keywords.

[0107] Step S504, obtaining the target enterprise product portrait corresponding to the candidate feedback text through the enterprise product portrait analysis network generated by advanced learning, and recommending the target user to the product portrait group where the target enterprise product portrait is located.

[0108] Further, Figure 2 FIG. 1 is a schematic diagram showing a hardware structure of a server system 100 for implementing the method provided in an embodiment of the present application. Figure 2As shown, the server system 100 may include one or more processors 102 (the processor 102 may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA), a memory 104 for storing data, and a transmission device 106 and a controller 108 for communication functions. It can be understood by those skilled in the art that Figure 2 The structure shown is only for illustration and does not limit the structure of the server system 100. For example, the server system 100 may also include Figure 2 More or fewer components as shown, or with Figure 2 Different configurations shown.

[0109] The memory 104 can be used to store software programs and modules of application software, such as program instructions corresponding to the above-mentioned method embodiments in the embodiments of the present application. The processor 102 executes various functional applications and data processing by running the software programs and modules stored in the memory 104, that is, realizing the above-mentioned user recommendation method based on enterprise product portrait analysis. The memory 104 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 104 may further include a memory remotely arranged relative to the processor 102, and these remote memories may be connected to the server system 100 via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0110] The transmission device 106 is used to receive or send data via a network. The specific example of the above network may include a wireless network provided by a communication provider of the server system 100. In one example, the transmission device 106 includes a network adapter, which can be connected to other network devices through a base station so as to communicate with the Internet. In one example, the transmission device 106 can be a radio frequency module, which is used to communicate with the Internet wirelessly.

[0111] It should be noted that the above-mentioned sequence of the embodiments of the present application is only for description and does not represent the advantages and disadvantages of the embodiments. And the above-mentioned specific embodiments of the embodiments of the present application are described. Other embodiments are within the scope of the attached claims. In some cases, the actions or steps recorded in the claims can be performed according to an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0112] Each embodiment in the embodiments of the present application is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the above different embodiments, since they are basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0113] A person skilled in the art will understand that all or part of the steps to implement the above embodiments may be accomplished by hardware or by instructing related hardware through a program. The above program may be stored in a computer-readable storage medium, and the above storage medium may be a read-only memory, a disk or an optical disk, etc.

Claims

1. A user recommendation method based on enterprise product portrait analysis, characterized in that: The method comprises: Acquire a first training feedback text sequence, a second training feedback text sequence, and a third training feedback text sequence for conversation feedback by different users in the process of using the enterprise product, wherein the training feedback text in the first training feedback text sequence is a frequent service feedback text, the training feedback text in the second training feedback text sequence is a sparse service feedback text, and the third training feedback text sequence includes a frequent service feedback text and a sparse service feedback text, wherein the frequent service feedback text means that the service type corresponding to the keywords in the text is a frequent service, and the frequent service is a service whose user usage frequency is greater than a first set frequency, and the sparse service feedback text means that the service type corresponding to the keywords in the text is a sparse service, and the sparse service is a service whose user usage frequency is lower than a second set frequency; Optimizing the initialization weight parameters of the basic enterprise product portrait analysis network according to the frequent business feedback text in the first training feedback text sequence to generate a first enterprise product portrait analysis network; Traversing the first training feedback text sequence, the second training feedback text sequence and the third training feedback text sequence to generate a reference training feedback text sequence required for advanced learning, loading each reference training feedback text in the reference training feedback text sequence into the first enterprise product portrait analysis network, and generating predicted enterprise product portrait data corresponding to each reference training feedback text; Determine the advanced learning Loss value based on the enterprise product portrait annotation data corresponding to each of the reference training feedback texts and the predicted enterprise product portrait data; When the advanced learning Loss value converges, obtaining the learning change rate corresponding to each of the reference training feedback texts; From the learning change rates corresponding to the reference training feedback texts, after removing the learning change rates corresponding to the reference training feedback texts that belong to the third training feedback text sequence and include frequent business keywords, the network weight information of the first enterprise product portrait analysis network is optimized according to the learning change rates corresponding to the remaining reference training feedback texts in the reference training feedback text sequence, until the traversal is completed, and an enterprise product portrait analysis network for performing enterprise product portrait analysis on frequent business feedback texts and sparse business feedback texts is generated, and the enterprise product portrait analysis network is used to identify the target enterprise product portrait of the candidate feedback text of any target user and recommend the target user to the product portrait group where the target enterprise product portrait is located.

2. The user recommendation method based on enterprise product portrait analysis according to claim 1 is characterized in that: The initializing weight parameter optimization of the basic enterprise product portrait analysis network according to the frequent business feedback text in the first training feedback text sequence to generate the first enterprise product portrait analysis network includes: Obtaining frequent business feedback text and corresponding enterprise product portrait annotation data from the first training feedback text sequence; According to the semantic encoder in the basic enterprise product portrait analysis network, extract the text semantic encoding vector corresponding to each of the frequent business feedback texts; According to the fully connected output unit in the basic enterprise product portrait analysis network, the predicted enterprise product portrait data corresponding to each of the frequent business feedback texts is obtained according to the text semantic encoding vector; Constructing a first Loss value based on the enterprise product portrait annotation data and the predicted enterprise product portrait data; After using the network weight information when the first Loss value converges as the network weight information for the basic enterprise product portrait analysis network optimization, return to the step of obtaining frequent business feedback text from the first training feedback text sequence to continue training until the convergence requirements are met.

3. The user recommendation method based on enterprise product portrait analysis according to claim 1 is characterized in that: The step of generating a reference training feedback text sequence required for advanced learning from the first training feedback text sequence, the second training feedback text sequence, and the third training feedback text sequence includes: Get the set extraction parameters; The first training feedback text sequence, the second training feedback text sequence and the third training feedback text sequence are extracted respectively according to the set extraction parameters to generate the reference training feedback text sequence required for advanced learning.

4. The user recommendation method based on enterprise product portrait analysis according to claim 1 is characterized in that: The step of loading each reference training feedback text in the reference training feedback text sequence into the first enterprise product portrait analysis network to generate predicted enterprise product portrait data corresponding to each reference training feedback text includes: Extracting text semantic encoding vectors corresponding to each of the reference training feedback texts according to the semantic encoder in the first enterprise product portrait analysis network; According to the fully connected output unit in the first enterprise product portrait analysis network, the predicted enterprise product portrait data corresponding to each reference training feedback text is obtained according to the text semantic encoding vector.

5. The user recommendation method based on enterprise product portrait analysis according to claim 1 is characterized in that: The step of determining the advanced learning Loss value based on the enterprise product portrait annotation data corresponding to each of the reference training feedback texts and the predicted enterprise product portrait data includes: Determine a first Loss value based on the enterprise product portrait annotation data corresponding to the reference training feedback text belonging to the first training feedback text sequence in the reference training feedback text sequence and the predicted enterprise product portrait data; Determine a second Loss value based on the enterprise product portrait annotation data corresponding to the reference training feedback text belonging to the second training feedback text sequence in the reference training feedback text sequence and the predicted enterprise product portrait data; Determine a third Loss value based on the enterprise product portrait annotation data corresponding to the reference training feedback text belonging to the third training feedback text sequence in the reference training feedback text sequence and the predicted enterprise product portrait data; Get the set weight distribution information; The first Loss value, the second Loss value and the third Loss value are weightedly fused according to the weight allocation information to generate the advanced learning Loss value.

6. The user recommendation method based on enterprise product portrait analysis according to claim 1 is characterized in that: The method further comprises: When, in the reference training feedback text sequence, the reference training feedback text extracted from the third training feedback text sequence includes a sparse business feedback text, optimizing the network weight information of the first enterprise product portrait analysis network according to the learning change rates corresponding to the reference training feedback text belonging to the third training feedback text sequence and including the sparse business keywords, the reference training feedback text belonging to the first training feedback text sequence, and the reference training feedback text belonging to the second training feedback text sequence; When the reference training feedback text extracted from the third training feedback text sequence in the reference training feedback text sequence does not include sparse business feedback text, the network weight information of the first enterprise product portrait analysis network is optimized based on the learning change rate corresponding to the reference training feedback text belonging to the first training feedback text sequence and the reference training feedback text belonging to the second training feedback text sequence.

7. The user recommendation method based on enterprise product portrait analysis according to claim 1 is characterized in that: The method further comprises: Acquire a test feedback text and a test enterprise product portrait corresponding to the test feedback text, wherein the test feedback text includes frequent business keywords or sparse business keywords; Obtaining a test text semantic vector corresponding to the test feedback text through the enterprise product portrait analysis network generated by advanced learning; Obtaining the annotated text semantic vector corresponding to the product portrait of the test enterprise; When the representation vector distance between the test text semantic vector and the annotated text semantic vector is less than a set distance, it is determined that the enterprise product portrait analysis network meets the target deployment conditions.

8. The user recommendation method based on enterprise product portrait analysis according to claim 1 is characterized in that: The method further comprises: Acquire candidate feedback texts of target users, where the candidate feedback texts include frequent business keywords or sparse business keywords; The enterprise product portrait analysis network generated by advanced learning is used to obtain the target enterprise product portrait corresponding to the candidate feedback text, and the target user is recommended to the product portrait group where the target enterprise product portrait is located.

9. A server system, characterized in that: The server system includes a processor and a readable storage medium, wherein the readable storage medium stores a program, and when the program is executed by the processor, the user recommendation method based on enterprise product portrait analysis described in any one of claims 1 to 8 is implemented.

Images