Theme information generation method and device, storage medium and electronic equipment

Abstract

The invention discloses a theme information generation method and device, a storage medium and electronic equipment, and relates to the technical field of artificial intelligence. The method comprises the steps that feedback information of a product, first feature information and second feature information of an object used for issuing the feedback information are obtained, the first feature information is used for representing features of the product, and the second feature information is used for representing features of the object; integrating the feedback information, the first feature information and the second feature information into target information; the target information is clustered, a clustering result is obtained, and the clustering result is used for representing the incidence relation between themes to which multiple pieces of sub-feedback information in the feedback information belong; based on the clustering result, theme information of the multiple pieces of sub-feedback information is generated, and the theme information is used for representing the theme to which the demand information of the object for the product belongs. The technical problem of low accuracy of theme information generation is solved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a method, device, storage medium and electronic device for generating subject information. Background Art

[0002] In related technologies, text clustering technology is usually used to summarize feedback information. First, a deep neural network is used to extract the text features of the feedback information, and then a common clustering algorithm (such as K-means) is used to cluster the feedback information to determine the topic information in the feedback information.

[0003] However, the above scheme ignores the characteristic information of the object used to publish the feedback information and the characteristic information of the product. In addition, the above analysis uses non-hierarchical clustering. For a certain feedback information, it is difficult to determine the commonality between multiple topics related to the feedback information, so there is a technical problem of low accuracy in generating topic information.

[0004] To address the above-mentioned problems, no effective solution has been proposed yet. Summary of the invention

[0005] The embodiments of the present application provide a method, device, storage medium and electronic device for generating topic information, so as to at least solve the technical problem of low accuracy in generating topic information.

[0006] According to one aspect of an embodiment of the present application, a method for generating topic information is provided. The method may include: obtaining feedback information of a product, first feature information, and second feature information of an object for publishing feedback information, wherein the first feature information is used to represent the feature of the product, and the second feature information is used to represent the feature of the object; integrating the feedback information, the first feature information, and the second feature information into target information; clustering the target information to obtain clustering results, wherein the clustering results are used to represent the association relationship between the topics to which multiple sub-feedback information in the feedback information belongs; based on the clustering results, generating topic information of multiple sub-feedback information, wherein the topic information is used to represent the topic to which the object's demand information for the product belongs.

[0007] Optionally, the feedback information, the first feature information and the second feature information are integrated into target information, including: integrating the feedback information, the first feature information and the second feature information into initial information; converting the initial information into vector data, and performing dimensionality reduction processing on the vector data to obtain the target information.

[0008] Optionally, the vector data is subjected to dimensionality reduction processing to obtain target information, including: performing mean centering processing on the vector data to obtain processed vector data; determining the covariance matrix of the processed vector data; performing eigenvalue decomposition on the covariance matrix to decompose the covariance matrix into eigenvalues ​​and eigenvectors; taking the eigenvectors corresponding to the top target number of eigenvalues ​​as principal components, and projecting the vector data onto the principal components to obtain target information.

[0009] Optionally, before integrating the feedback information, the first characteristic information and the second characteristic information into the target information, the method further includes: cleaning the feedback information, the first characteristic information and the second characteristic information; after integrating the feedback information, the first characteristic information and the second characteristic information into the initial information, the method further includes: cleaning the initial information.

[0010] Optionally, clustering the target information to obtain a clustering result includes: determining the mutual reachable distances between multiple data points in the target information; determining core points among the multiple data points based on the mutual reachable distances; and clustering the target information based on the core points to obtain a clustering result.

[0011] Optionally, the target information is clustered based on the core points to obtain a clustering result, including: constructing a core graph based on the core points, and generating a minimum spanning tree based on the core graph, wherein the weights of the edges between the nodes of the core graph are the mutually reachable distances between the nodes; sorting the edges of the minimum spanning tree according to the weights of the edges of the minimum spanning tree, and clustering the sorted edges to obtain a clustering result.

[0012] Optionally, based on the clustering results, topic information of multiple sub-feedback information is generated, including: calling a text classification model to predict the feedback information to obtain a prediction result, wherein the prediction result is used to indicate the probability that the feedback information meets a preset topic; based on the prediction result, pruning the clustering results, and generating topic information based on the pruned clustering results.

[0013] Optionally, based on the pruned clustering results, topic information is generated, including: splicing prompt information associated with the product and the pruned clustering results to obtain query information; using the prompt information to guide the large language model to generate reply information matching the query information, wherein the large language model is trained using query information samples and reply information samples; and adjusting the reply information to obtain topic information.

[0014] Optionally, the method also includes: extracting sub-feedback information that meets a preset theme from the feedback information, and annotating the sub-feedback information to obtain annotated data; constructing a training set based on the annotated data; and training an initial text classification model based on the training set to obtain a text classification model.

[0015] According to another aspect of the embodiment of the present application, a device for generating topic information is also provided. The device may include: an acquisition unit, used to acquire feedback information of a product, first feature information, and second feature information of an object for publishing feedback information, wherein the first feature information is used to represent the feature of the product, and the second feature information is used to represent the feature of the object; an integration unit, used to integrate the feedback information, the first feature information, and the second feature information into target information; a clustering unit, used to cluster the target information to obtain a clustering result, wherein the clustering result is used to represent the association relationship between the topics to which multiple sub-feedback information in the feedback information belongs; a generation unit, used to generate topic information of multiple sub-feedback information based on the clustering result, wherein the topic information is used to represent the topic to which the object's demand information for the product belongs.

[0016] According to another aspect of an embodiment of the present application, a non-volatile storage medium is also provided, including: the storage medium includes a stored program, wherein when the program is running, the device where the storage medium is located is controlled to execute any method for generating subject information.

[0017] According to another aspect of an embodiment of the present invention, a processor is further provided. The processor is used to run a program. When the program is run, the method for generating subject information in an embodiment of the present invention is executed.

[0018] According to another aspect of an embodiment of the present application, there is also provided an electronic device, comprising: a processor; and a memory for storing processor executable instructions; wherein the processor is configured to execute instructions to implement any method for generating topic information.

[0019] According to another aspect of the embodiment of the present application, a computer program product is further provided. The computer program product may include a computer program, wherein the computer program is configured to execute any of the above methods for generating topic information when running.

[0020] According to another aspect of the embodiment of the present application, a computer program product is also provided. The computer program product may include a non-volatile computer-readable storage medium, wherein the non-volatile computer-readable storage medium stores a computer program, and the computer program is configured to execute any of the above methods for generating subject information when running.

[0021] According to another aspect of the embodiment of the present application, a computer program is further provided, which is configured to execute any of the above methods for generating topic information when running.

[0022] In an embodiment of the present application, feedback information of a product, first feature information, and second feature information of an object for publishing feedback information are obtained, wherein the first feature information is used to represent the feature of the product, and the second feature information is used to represent the feature of the object; the feedback information, the first feature information, and the second feature information are integrated into target information; the target information is clustered to obtain a clustering result, wherein the clustering result is used to represent the association relationship between the topics to which multiple sub-feedback information in the feedback information belongs; based on the clustering result, topic information of multiple sub-feedback information is generated, wherein the topic information is used to represent the topic to which the object's demand information for the product belongs, thereby achieving a technical effect of improving the accuracy of generated topic information, and further solving the technical problem of low accuracy of generated topic information. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0024] Figure 1 It is a hardware structure block diagram of a computer terminal (or mobile device) for implementing a method for generating topic information according to an embodiment of the present application;

[0025] Figure 2 is a flow chart of a method for generating topic information according to an embodiment of the present application;

[0026] Figure 3 is a schematic diagram of a device for automatically summarizing and summarizing the pain points of users of financial products in batches based on natural language processing according to an embodiment of the present application;

[0027] Figure 4 is a schematic diagram of a data integration module according to an embodiment of the present application;

[0028] Figure 5 is a schematic diagram of a text clustering module according to an embodiment of the present application;

[0029] Figure 6 is a schematic diagram of a pruning unit according to an embodiment of the present application;

[0030] Figure 7 is a schematic diagram of an automatic summarization and induction module according to an embodiment of the present application;

[0031] Figure 8 This is a flow chart of a method for automatically summarizing and summarizing the pain points of financial product users based on natural language processing according to an embodiment of the present application;

[0032] Fig. 9is a structural schematic diagram of a device for generating topic information according to an embodiment of the present application;

[0033] Fig.10 is a schematic block diagram of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0034] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described 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 this field without creative work should fall within the scope of protection of the present application.

[0035] 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.

[0036] It should be noted that the collected information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for display, data for analysis, etc.) involved in this application are information and data authorized by the user or fully authorized by all parties, and the collection, storage, use, processing, transmission, provision, disclosure and application of relevant data are in compliance with relevant laws, regulations and standards, necessary confidentiality measures are taken, and public order and good customs are not violated, and corresponding operation entrances are provided for users to choose to authorize or refuse. For example, an interface is set up between this system and relevant users or institutions to provide users with corresponding operation entrances for users to choose to agree or refuse the results of automated decision-making; if the user chooses to refuse, the expert decision-making process will be entered.

[0037] According to an embodiment of the present application, a method embodiment of a method for generating subject information is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0038] The method embodiments provided in the embodiments of the present application can be executed in a mobile terminal, a computer terminal or a similar computing device. Figure 1 A hardware structure block diagram of a computer terminal (or mobile device) for implementing a method for generating topic information is shown. Figure 1 As shown, the computer terminal 10 (or mobile device) may include one or more (102a, 102b, ..., 102n are used to illustrate) 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 for communication functions. In addition, it may also include: a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of the BUS bus), a network interface, a power supply and / or a camera. It can be understood by those skilled in the art that Figure 1 The structure shown is only for illustration and does not limit the structure of the above electronic device. Figure 1 More or fewer components as shown, or with Figure 1 Different configurations are shown.

[0039] It should be noted that the one or more processors 102 and / or other data processing circuits described above may generally be referred to herein as "data processing circuits". The data processing circuits may be embodied in whole or in part as software, hardware, firmware, or any other combination thereof. In addition, the data processing circuit may be a single independent processing module, or may be incorporated in whole or in part into any of the other components in the computer terminal 10 (or mobile device). As described in the embodiments of the present application, the data processing circuit acts as a processor control (e.g., selection of a variable resistor terminal path connected to an interface).

[0040] The memory 104 can be used to store software programs and modules of application software, such as program instructions / data storage devices corresponding to the method for generating the subject information in the embodiment 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, the method for generating the subject information described above is realized. 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 examples, the memory 104 may further include a memory remotely arranged relative to the processor 102, and these remote memories may be connected to the computer terminal 10 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.

[0041] 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 computer terminal 10. In one example, the transmission device 106 includes a network adapter (Network Interface Controller, NIC), 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 (RF) module, which is used to communicate with the Internet wirelessly.

[0042] The display may be, for example, a touch screen liquid crystal display (LCD) that enables a user to interact with a user interface of the computer terminal 10 (or mobile device).

[0043] Under the above operating environment, this application provides Figure 2 The method for generating the topic information shown. Figure 2 is a flow chart of a method for generating topic information according to an embodiment of the present application, such as Figure 2 As shown, the method may include the following steps:

[0044] Step S202, obtaining feedback information of the product, first characteristic information, and second characteristic information of the object for publishing the feedback information, wherein the first characteristic information is used to represent the characteristics of the product, and the second characteristic information is used to represent the characteristics of the object.

[0045] In the technical solution provided in the above step S202, feedback information of the product, first characteristic information, and second characteristic information of the object for publishing the feedback information can be obtained. Among them, the product can be a financial product, and the object can be a user. The feedback information can be the text content of the above product that the user feedbacks on the feedback page, and can also be called feedback data or feedback text. The first characteristic information can be product characteristic data, and the first characteristic information can include the type of product, the name of the product, the issuing company, etc. The second characteristic information can be user characteristic data, and the second characteristic information can include the user's financial management preference, occupation, gender, age, etc.

[0046] Optionally, the feedback data of user evaluations can be integrated by using crawler technology to crawl the text content of user feedback on the feedback page, and collect feedback texts from the same user on different platforms but for the same service or product. The above feedback texts can be generated from different feedback channels such as questionnaires, online comments and messages, and no specific restrictions are made here on the way feedback data is generated. When feedback is given through a telephone message, the feedback content can be converted into text using speech-to-text technology. No specific restrictions are made here on the way feedback data is fed back, including but not limited to the form of textual feedback. Among them, the content of the feedback text is unstructured text data, and the feedback data can be generated using D 反馈文本 To express.

[0047] Optionally, the user's relevant characteristics are filtered out from the online database according to the user identifier (ID), such as financial management preferences, occupation, gender, age and other structured data. The user characteristic data can be used in D 用户 In addition, relevant features of the product being evaluated can be extracted from the database, such as the type of financial product, the name of the financial product, the issuing company, etc. The product feature data can be represented by D 产品 Here, structured data related to users or products is collected. It should be noted that in order to meet the different analysis needs of business personnel, business personnel can be allowed to select the required user feature data and product feature data.

[0048] Step S204: Integrate the feedback information, the first feature information, and the second feature information into target information.

[0049] In the technical solution provided in the above step S204, after obtaining the feedback information, the first characteristic information and the second characteristic information, the feedback information, the first characteristic information and the second characteristic information can be integrated into the target information. Optionally, the text feedback from the same user for the same product but from different sources can be spliced ​​together by text splicing, and then the above cleaned structured data (i.e., user characteristic data and product characteristic data) can be spliced ​​together in the form of key-value pairs {K, V} and placed at the beginning of a data. K can be used to represent key data, and V can be used to represent value data.

[0050] Optionally, firstly, key-value confirmation is performed on the structured data, for example, the user name is the key, and the specific name is the value. Further key-value concatenation operation is performed on the structured data, for example, the "user name" string is concatenated with the actual name, and the final result can be: "The user purchased a certain type of a certain product on a certain date in a certain year". Finally, the structured data and the unstructured data are concatenated, and the formal description of the complete string can be expressed by the following formula:

[0051]

[0052] Among them, D can be used to represent target information, n can be used to represent the number of user feature data, m can be used to represent the number of product feature data, and p can be used to represent the number of feedback data. After the above operations, the structured data related to feedback is integrated into the text data, which makes it easier for the subsequent deep neural network to learn data features that were not integrated by previous technologies. Structured data can assist in subsequent clustering, making the text clustering results more accurate.

[0053] Step S206: clustering the target information to obtain a clustering result, wherein the clustering result is used to represent the association relationship between the topics to which the plurality of sub-feedback information in the feedback information belong.

[0054] In the technical solution provided in the above step S206, after the feedback information, the first characteristic information and the second characteristic information are integrated into the target information, the obtained target information can be clustered to obtain a clustering result, wherein the clustering result can be used to represent the association relationship between the topics to which multiple sub-feedback information in the feedback information belong.

[0055] Optionally, after the feedback information, the first feature information, and the second feature information are integrated into the target information, the mutual reachable distances between the multiple data points in the target information can be determined, and then the core points can be determined based on the mutual reachable distances. Further based on the core points, a minimum spanning tree can be constructed to perform clustering using the minimum spanning tree to obtain a clustering result.

[0056] Step S208: generating topic information of a plurality of sub-feedback information based on the clustering result, wherein the topic information is used to indicate the topic to which the object's demand information for the product belongs.

[0057] In the technical solution provided in the above step S208, after clustering the target information and obtaining the clustering results, the subject information of multiple sub-feedback information can be generated based on the obtained clustering results. The subject information can be used to indicate the subject to which the object's demand information for the product belongs, and the subject information can also be called a subject summary.

[0058] Optionally, after clustering the target information and obtaining the clustering results, the obtained clustering results can be pruned to remove irrelevant results. Further based on the pruned clustering results, the large language model can be guided to generate topic information.

[0059] In an embodiment of the present application, feedback information of a product, first feature information, and second feature information of an object for publishing feedback information are obtained, wherein the first feature information is used to represent the feature of the product, and the second feature information is used to represent the feature of the object; the feedback information, the first feature information, and the second feature information are integrated into target information; the target information is clustered to obtain a clustering result, wherein the clustering result is used to represent the association relationship between the topics to which multiple sub-feedback information in the feedback information belongs; based on the clustering result, topic information of multiple sub-feedback information is generated, wherein the topic information is used to represent the topic to which the object's demand information for the product belongs, thereby achieving a technical effect of improving the accuracy of generated topic information, and further solving the technical problem of low accuracy of generated topic information.

[0060] The above embodiments of the present application are further described below.

[0061] In some embodiments of the present application, step S204 integrates the feedback information, the first feature information, and the second feature information into target information, including: integrating the feedback information, the first feature information, and the second feature information into initial information; converting the initial information into vector data, and performing dimensionality reduction processing on the vector data to obtain the target information.

[0062] In this embodiment, after obtaining the feedback information, the first feature information and the second feature information, the feedback information, the first feature information and the second feature information may be integrated into the initial information. The initial information may be further converted into vector data, and the vector data may be subjected to dimensionality reduction processing to obtain the target information. The initial information may be information obtained by integrating the feedback information, the first feature information and the second feature information. The target information may be information obtained by cleaning, extracting features and subjecting the initial information to dimensionality reduction processing.

[0063] Optionally, after the feedback information, the first feature information and the second feature information are integrated into the initial information, the initial information can be cleaned, and the cleaned text data can be further mapped into a high-dimensional space, so that the text of each feedback evaluation and the corresponding structured data are converted into a numerical vector form. The text is converted into an n-dimensional numerical vector using a bidirectional encoder representation (BidirectionalEncoder Representations from Transformers, referred to as BERT) pre-training model. Among them, the choice of n depends on the length of the text data after cleaning, and it is freely selected on this basis. For example, n can be 768, which is only an example here, and there is no specific restriction on the value of n. Finally, the BERT pre-training model converts all text data into an m*n matrix, where m can be used to represent m text data, converts the text data into an n-dimensional text vector, and stores the n-dimensional text vector to provide a data source for subsequent deep neural network classification pruning.

[0064] Optionally, after converting the initial information into vector data, the dimension of the obtained n-dimensional text vector of each data can be reduced. Since too high a dimension will lead to the dimensionality disaster for clustering algorithms, whether calculating the Euclidean distance or similarity between vectors, it will lead to excessive calculation. Therefore, after converting the initial information into vector data, the vector data needs to be reduced in dimension.

[0065] In some embodiments of the present application, dimensionality reduction is performed on vector data to obtain target information, including: mean centering the vector data to obtain processed vector data; determining the covariance matrix of the processed vector data; performing eigenvalue decomposition on the covariance matrix to decompose the covariance matrix into eigenvalues ​​and eigenvectors; taking the eigenvectors corresponding to the top target number of eigenvalues ​​as principal components, and projecting the vector data onto the principal components to obtain target information.

[0066] In this embodiment, the target number may be a number set according to actual conditions, and the target number may be represented by K. After the initial information is converted into vector data, a principal component analysis (PCA) dimensionality reduction method may be used to reduce the dimensionality of the vector data.

[0067] Optionally, each feature is first mean-centered, that is, the mean of the feature is subtracted to ensure that the mean of the data is zero. Further calculate the covariance matrix cov(DD T) The covariance matrix describes the linear relationship between data features. After determining the covariance matrix of the data, eigenvalue decomposition can be performed on the covariance matrix, and the covariance matrix can be decomposed into eigenvalues and eigenvectors using singular value decomposition. Among them, the eigenvalues represent the variance in the data, and the eigenvectors represent the directions of the principal components.

[0068] Optionally, after decomposing the covariance matrix into eigenvalues and eigenvectors, they can be sorted according to the magnitudes of the eigenvalues, and the eigenvectors corresponding to the top k eigenvalues can be selected as the principal components, where k is the dimension that is desired to be retained. For example, k can be determined such that the numerical difference between the top k eigenvalues and the bottom k eigenvalues is extremely large, and k is selected to be as large as possible while retaining. Finally, the original data is projected onto the selected principal components to form a new feature space. The projection can be achieved by matrix multiplication. At this time, the text vector is k-dimensional, and the data is a matrix D of m * k p (i.e., the target information).

[0069] In some embodiments of the present application, before integrating the feedback information, the first feature information, and the second feature information into the target information, the method further includes: cleaning the feedback information, the first feature information, and the second feature information; after integrating the feedback information, the first feature information, and the second feature information into the initial information, the method further includes: cleaning the initial information.

[0070] In this embodiment, after obtaining the feedback information, the first feature information, and the second feature information, the feedback information, the first feature information, and the second feature information can be cleaned. For the feedback information, the method of cleaning text data can be adopted for processing. For the first feature information and the second feature information, the method of cleaning structured data can be adopted for processing.

[0071] Optionally, for text data, first use a function to remove special characters to remove the special characters in the text string. The special characters can include punctuation marks, Hyper Text Markup Language (HTML) tags, emojis, non-standard fonts, etc. The function for removing special characters is regularized filtering by writing a regular expression that identifies the above special characters, and replaces the special characters matched by the regular expression with empty characters, thereby playing the role of removing special characters.

[0072] Optionally, for text data, further use a function to remove common Chinese stop words to delete the stop words in the text. Chinese stop words can include words such as "de", "shi", "le" that have no actual meaning. Removing stop words can improve the efficiency of text processing and the performance of subsequent feature extraction modules. The function for removing Chinese stop words saves a list of common Chinese stop words, traverses all text data, and deletes them when these stop words exist.

[0073] Optionally, for text data, a function that combines simple and commonly used text preprocessing is used to process the text data. The functions of this function include processing repeated characters, unifying text formats, and removing redundant blank characters. Unifying the text format can convert all Chinese characters into a unified case and encoding format from the encoding level, and all characters are uniformly converted into a common format (for example, UTF-8). Unifying the text format and removing redundant blank characters can be achieved by writing rules using regular expressions, and the principle is the same as the function of removing special characters.

[0074] Optionally, for structured data, cleaning of structured data may include null value processing, type conversion, and duplicate value processing. Among them, null value processing is to check whether there is a feature in each data that is null. If it is not null, go directly to the next step. If there is a null value, the business personnel can choose to remove the null value or fill the missing value, depending on the business personnel's goals and the actual amount of analyzed data. If the amount of data is too small, the missing value filling method is used. For the missing values ​​of the category type, it can be left unprocessed. For the missing values ​​of the numerical type, you can choose to fill 0, the default value or the average value. If the amount of data is large enough, the data with missing values ​​can be discarded, so that the accuracy of the data can be ensured as much as possible. That is, the processing function opens different processing methods for business personnel to choose according to the actual situation. Type conversion mainly converts the numerical type into a string form, which is convenient for the subsequent data integration module splicing, and finally enters the duplicate value processing function. By checking whether there is the same data, if it exists, it is removed and only the first one is retained.

[0075] Optionally, after the feedback information, the first feature information, and the second feature information are integrated into the initial information, the initial information can be further cleaned. Data cleaning is performed by calling the created function for removing special characters, the function for removing common Chinese stop words, the function for common text preprocessing, and the preprocessing function for structured data. In order to clean the impact caused by data integration, since meaningless words such as stop words may be introduced when text collages are performed on different sources and structured data, the integrated initial information is cleaned again, and a text that only retains meaningful words and has a word order with context will be obtained.

[0076] In some embodiments of the present application, the target information is clustered to obtain a clustering result, including: determining the mutual reachable distances between multiple data points in the target information; based on the mutual reachable distances, determining core points among the multiple data points; based on the core points, clustering the target information to obtain a clustering result.

[0077] In this embodiment, it should be noted that the traditional Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN) is a relatively advanced hierarchical clustering algorithm. However, since most of the evaluation feedback of financial products is massive, the traditional HDBSCAN algorithm uses density clustering, core distance, mutual reachability and minimum spanning tree algorithm, and the calculation efficiency is slow on big data. In order to improve the efficiency of the algorithm and reduce the waiting time of business personnel, this embodiment splits part of the calculation and calculates in parallel on the graphics processing unit (GPU) based on the Compute Unified Device Architecture (CUDA) computing engine.

[0078] Optionally, first prepare the data for layering and parallelization, and then construct an m*k tensor on CUDA as a container for storing the results. p The matrix (i.e. target information) is converted into a tensor of the same dimension for easy calculation on CUDA. Next, the k distance of each data point and the mutual distance of all points are calculated. For each point, the distance between the point and other points is calculated, and the distances are sorted in ascending order. The kth distance is selected as the k distance. The distance from each point to other points can be calculated in parallel to improve the calculation efficiency. After the calculation is completed, the sorting algorithm is run to take the first k distances and store them in an m*k tensor. For example, the k distance of data point A can be calculated using d core-k (A) To express.

[0079] Furthermore, the mutual reachable distance of all points is calculated. The mutual reachable distance of two points can be expressed by the following formula:

[0080] d m-k (A,B)=max(d core-k (A),d core-k (B),d(A,B))

[0081] Among them, d core-k (A) can be used to represent the k distance of data point A, d core-k(B) can be used to represent the k distance of data point B, (A, B) can be used to represent the mutual distance between data point A and data point B, and the mutual reachable distance takes the maximum value of the above. In addition, since the above values ​​are stored in the tensor, the mutual reachable distance of each two points can be calculated in parallel by calling the built-in function. To determine the core point based on the mutual reachable distance, a threshold of 1 can be set in advance. When the mutual reachable distance is less than the threshold, the point is marked as a core point. Finally, based on the core point, the target information is clustered to obtain the clustering result.

[0082] In some embodiments of the present application, target information is clustered based on core points to obtain clustering results, including: constructing a core graph based on core points, and generating a minimum spanning tree based on the core graph, wherein the weights of the edges between the nodes of the core graph are the mutually reachable distances between the nodes; sorting the edges of the minimum spanning tree according to the weights of the edges of the minimum spanning tree, and clustering the sorted edges to obtain clustering results.

[0083] In this embodiment, after the core points are determined based on the mutual reachable distance, a core graph can be constructed based on the core points, and the weight of the edge between the points in the core graph is the mutual reachable distance between the points. An algorithm for finding the minimum spanning tree in the graph (e.g., Prim's algorithm) is further used to generate a minimum spanning tree based on the core graph.

[0084] Optionally, after generating a minimum spanning tree based on the core graph, the edges of the minimum spanning tree may be sorted by weight, and the sorted edges may be clustered, starting from the edge with the minimum weight, and the two connected points may be merged into one cluster, at which point a hierarchical clustering result may be generated.

[0085] In the related art, the HDBSCAN clustering algorithm has a part of compressing the clustering tree, and the computational complexity is relatively high. However, this application adopts a new way to deal with the problem of too many branches in the tree. This part of the computation can be calculated in parallel when running the above steps, thereby speeding up the response speed of the entire application.

[0086] In some embodiments of the present application, topic information of multiple sub-feedback information is generated based on clustering results, including: calling a text classification model to predict the feedback information to obtain a prediction result, wherein the prediction result is used to indicate the probability that the feedback information meets a preset topic; based on the prediction result, pruning the clustering results, and generating topic information based on the pruned clustering results.

[0087] In this embodiment, we start from the bottom layer of hierarchical clustering. If the proportion of irrelevant prediction results in the lowest layer of clustering exceeds a certain threshold (the threshold can be defined by the business personnel), it can be considered that the cluster is not the direction that the business personnel want to analyze, and it can be cut off. At the same time, a dictionary is maintained for all texts to record which cluster the text belongs to and whether it has been cut off.

[0088] Optionally, after the bottom layer of clusters is traversed, the penultimate layer of clusters is traversed. The difference is that the objects in the penultimate layer of clusters are the bottom layer of clusters. According to similar logic, if the number of the bottom layer of clusters pruned in the cluster is more than the threshold, the cluster is pruned. This rule is executed until the second layer of clusters is pruned, and finally a pruned hierarchical clustering spanning tree is obtained. Further based on the hierarchical clustering spanning tree, the large language model is guided to perform auxiliary induction to generate topic information.

[0089] In some embodiments of the present application, topic information is generated based on the pruned clustering results, including: splicing prompt information associated with the product and the pruned clustering results to obtain query information; using the prompt information to guide a large language model to generate reply information matching the query information, wherein the large language model is trained using query information samples and reply information samples; and adjusting the reply information to obtain topic information.

[0090] In this embodiment, a rule-compliant access body is implemented through the application programming interface (API) of the large language model, a method is implemented, a question to be asked to the large language model is input, the request is returned and sent to the large model, and finally a structural parsing result is returned according to the large model.

[0091] For example, for the lowest level of clustering, the texts in a cluster can be organized into prompt words to prompt the large language model, such as, "I am a user feedback analyst for financial products. I have the following feedbacks. After reading these texts, what keywords do you think can be used to describe the themes of these texts? Can you help summarize the themes of the following texts?" For the second-to-last level of clustering, the clustering object is the theme, so the prompt words of the large model can be: "I am a user feedback analyst for financial products. I have the following types of feedback. After reading these types of feedback, what keywords do you think can be used to describe the commonalities of these types? Can you help summarize the commonalities and themes of the following feedback?"

[0092] Optionally, the prepared prompt words and the corresponding text after pruning are spliced ​​into a complete question, which is input into the above method, transmitted to the large language model and an answer is obtained. Based on the feedback from the large model, business personnel can compare the keywords returned by the large model with the summarized topics, and then generate the final topic based on their own business knowledge. Business personnel can repeat similar steps to obtain the topic of the penultimate clustering layer, push the categories one by one, and traverse all levels from low to high to finally obtain the topics of all levels of clustering and the topics of the topics.

[0093] In some embodiments of the present application, the method also includes: extracting sub-feedback information that meets a preset theme from the feedback information, and annotating the sub-feedback information to obtain annotated data; constructing a training set based on the annotated data; and training an initial text classification model based on the training set to obtain a text classification model.

[0094] In this embodiment, the preset topic can specify the required analysis direction or content for each analysis task for the business personnel. After obtaining the clustering results, the clustering results can be pruned, and the clusters that the business personnel need to see can be efficiently retrieved, which is the key to optimization. The pre-trained BERT model can be used for text classification tasks.

[0095] Optionally, business personnel can specify the required analysis direction or content for each analysis task, sample relevant text from the original text, and annotate the text. For example, feedback related to the investment yield of financial products and irrelevant feedback can construct a data set of more than 1,000 positive and negative samples. For general machine learning classification algorithms, the amount of data is small, but the accuracy of binary classification using the technology of fine-tuning the BERT model can reach about 89.61%.

[0096] Furthermore, the labeled data can be divided into a training set and a validation set in a ratio of 8:2. Connect the BERT model network structure to the output layer, which has only two neurons, use an activation function (e.g., sigmoid), and the loss function is a cross-entropy loss function adapted to the binary sigmoid activation function. Use an optimizer (e.g., Adam) to optimize and start the back propagation algorithm. After training with the training set, use the validation set to adjust the parameters and verify the model effect. Finally, input the full text vector into the trained binary classification model to obtain the prediction result. Based on the prediction result, pruning can be performed according to a specific algorithm.

[0097] In an embodiment of the present application, feedback information of a product, first feature information, and second feature information of an object for publishing feedback information are obtained, wherein the first feature information is used to represent the feature of the product, and the second feature information is used to represent the feature of the object; the feedback information, the first feature information, and the second feature information are integrated into target information; the target information is clustered to obtain a clustering result, wherein the clustering result is used to represent the association relationship between the topics to which multiple sub-feedback information in the feedback information belongs; based on the clustering result, topic information of multiple sub-feedback information is generated, wherein the topic information is used to represent the topic to which the object's demand information for the product belongs, thereby achieving a technical effect of improving the accuracy of generated topic information, and further solving the technical problem of low accuracy of generated topic information.

[0098] In order to facilitate those skilled in the art to better understand the technical solution of the present application, a specific embodiment is now described.

[0099] At present, with the development of the financial market and the introduction of various financial products, the concept of financial management has gradually become popular among the public. More and more users are purchasing various financial products and providing evaluation and feedback on various products. The increasing number of evaluation feedbacks has posed a huge challenge to the acceptance capacity of financial institutions. Financial institutions are facing a huge amount of data and it is increasingly difficult to find the pain points of users. In order to better improve the work efficiency and service quality of financial institutions, the financial field has put forward new requirements for batch summarizing and summarizing the evaluation feedback of financial products and timely discovering the pain points of users.

[0100] In related technologies, text clustering technology is used to achieve induction and summary, deep neural networks such as convolutional neural networks are used to extract text features, and common clustering algorithms (such as K-means) are used for clustering to determine the pain points in the feedback. Finally, based on the clustering results, representative documents in a cluster are selected for expert review based on indicators such as keywords and importance, and the theme of the cluster is finally determined. Finally, based on the themes of all clusters, common pain points are summarized to facilitate the business department to make decisions and modify opinions in the next stage.

[0101] However, the data source used in the above solution is only plain text, and ignores some structured information, such as the relevant characteristics of users who post feedback, feedback ratings and other data. Ignoring these data may ignore the differences in the feelings of different groups on a certain service, making it impossible for business personnel to conduct detailed analysis. Secondly, pain point analysis generally uses non-hierarchical clustering. For a certain text, it may belong to a certain topic, and this topic may belong to a certain large topic. It is not convenient for analysts to find the commonalities of these topics, which increases the difficulty for analysts to analyze pain points.

[0102] The current hierarchical clustering algorithm has the problem of a large number of results and difficulty in retrieval, which causes analysts to spend a lot of time analyzing all topics instead of analyzing some important and urgent pain points in a targeted manner. Sampling all clusters and reviewing the manual induction of pain points is time-consuming and laborious, and it is difficult to process large-scale data, resulting in the inability to make pain point analysis more refined, resulting in a technical problem of low accuracy in generating topic information.

[0103] In order to solve the above problems, the present application proposes a method and device for automatically summarizing the pain points of financial product users in batches based on natural language processing. Figure 3 is a schematic diagram of a device for automatically summarizing and summarizing the pain points of financial product users based on natural language processing according to an embodiment of the present application, such as Figure 3 As shown, the device may include a data collection module 301, a data cleaning module 302, a data integration module 303, a data re-cleaning module 304, a feature extraction module 305, a dimension reduction module 306, a text clustering module 307, and an automatic summarization module 308.

[0104] The data collection module 301 integrates relevant evaluation feedback data, uses crawler technology to crawl the text content of user feedback in the feedback page, and collects the text of feedback from the same user on different platforms but for the same service or product. These feedback texts can be generated from different feedback channels such as questionnaires, online comments and messages. If it is a telephone message feedback, the feedback content can be converted into text by voice-to-text technology. This embodiment is not limited to a specific feedback method, and can be generally applicable to all feedback forms that can convert feedback content into text. Among them, the content of the feedback text is unstructured text data D 反馈文本 , and then filter out the user's relevant features D from the online database based on the user ID 用户 , financial management preferences, occupation, gender, age and other structured data. Then extract the relevant features of the product being evaluated from the database D 产品 For example, the type of financial product, the name of the financial product, the issuing company, etc. The structured data collected here is related to the user or product. At the beginning of running this module, in order to meet the different analysis needs of business personnel, business personnel are allowed to select the required user feature data and product feature data.

[0105] The data cleaning module 302 cleans the text data and structured data respectively. For the text data, first, a function for removing special characters is used to remove the special characters in the text string. The special characters may include punctuation marks, HTML tags, emojis, non-standard fonts, etc. The principle of the function for removing special characters is to perform regular filtering by writing a regular expression that recognizes the above special characters, and replace the special characters matched by the regular expression with empty characters, thus playing the role of removing special characters.

[0106] Secondly, a function for removing common Chinese stop words is used to delete the stop words in the text. Chinese stop words include words without practical meaning such as "de", "shi", "le", etc. Removing these stop words can improve the efficiency of text processing and the performance of the subsequent feature extraction module. The principle of this function is to save a table of common Chinese stop words, traverse all the text data, and delete them when these stop words exist.

[0107] Finally, a function that combines simple and common text preprocessing is used to process the text data. The functions of this function include processing repeated characters, unifying the text format, and removing redundant white space characters. Unifying the text format can convert all Chinese characters into a unified case and encoding format at the encoding level. All characters are converted into UTF-8 format in this case. Unifying the text format and removing redundant white space characters are implemented by writing rules using regular expressions, and the principle is the same as that of the function for removing special characters.

[0108] For the preprocessing of structured data, it can include null value processing, type conversion, and duplicate value processing. Among them, null value processing is to check whether there are any empty features in each piece of data. If not, it directly goes to the next step. If there are null values, the business personnel can choose to remove the null values or fill in the missing values. This depends on the goals of the business personnel and the actual number of analyzed data. If the data volume is too small, the method of filling in the missing values is used. For the missing values of the categorical type, they can be left unprocessed. For the missing values of the numerical type, 0, the default value, or the average value can be chosen to fill in. If the data volume is large enough, the data with missing values can be discarded, which can ensure the accuracy of the data as much as possible. This processing function provides different processing methods for the business personnel to choose according to the actual situation. Type conversion mainly converts the numerical type into a string form to facilitate the subsequent splicing of the data integration module. Finally, it enters the duplicate value processing function, and its processing logic is to check whether there are the same data. If so, duplicate data is removed and the first one is retained.

[0109] Figure 4 is a schematic diagram of a data integration module according to an embodiment of the present application, as Figure 4As shown, the data integration module 303 includes a structured data key value confirmation unit 3031, a structured data key value splicing unit 3032, and a structured data and unstructured data splicing unit 3033. The data integration module splices together text feedback from the same user for the same product but from different sources by text splicing, and then splices all the structured data cleaned in the above modules in the form of key value pairs {K, V} and puts them at the beginning of a data.

[0110] The structured data key value confirmation unit 3031 confirms the key value of the structured data, for example, if the user name is the key, then the specific name is the value. The structured data key value splicing unit 3032 performs splicing operations on the key value of the structured data, for example, splicing the string of the user name with the actual name, and the final result may be: "the user purchased a certain type of a certain product on a certain date of a certain year".

[0111] The structured data and unstructured data concatenation unit 3033 concatenates the structured data and the unstructured data to obtain a formal description of a complete string, which can be expressed by the following formula:

[0112]

[0113] After this operation, the structured data related to the feedback is also integrated into the text data, which makes it easier for the subsequent deep neural network to learn data features that were not integrated by previous technologies. These structured data can assist in subsequent clustering and make the text clustering results more accurate.

[0114] The data re-cleaning module 304 respectively calls the functions for removing special characters, removing common Chinese stop words, simple common text pre-processing functions and structured data pre-processing functions created by the data cleaning module 302 to perform data cleaning again, so as to clean the impact caused by the integration of the data integration module 303. When text collage is performed on different sources and structured data, meaningless words such as stop words may be introduced. After the data is cleaned again by the data re-cleaning module 304, a text with only meaningful words and a word order with context will be obtained.

[0115] The feature extraction module 305 maps the text data processed by the data cleaning module 304 to a high-dimensional space, thereby converting the text of each feedback evaluation and its corresponding structured data into a numerical vector form. The BERT pre-training model is used to convert the text into an n-dimensional numerical vector. Among them, the choice of n depends on the length of the text data after cleaning, and it is freely selected on this basis, for example, n is 768. Finally, the BERT pre-training model will convert all text data into an m*n matrix, where m represents m text data. All text data will be converted into n-dimensional text vectors, and these n-dimensional text vectors will be saved to provide a data source for subsequent deep neural network classification pruning.

[0116] The dimension reduction module 306 reduces the dimension of the n-dimensional text vector of each data. Since too high a dimension will lead to the dimensional disaster for the subsequent clustering algorithm, whether it is calculating the Euclidean distance or similarity between vectors, it will lead to excessive calculation, so the principal component analysis PCA dimension reduction method is used. Each feature is mean-centered, that is, the mean of the feature is subtracted to ensure that the mean of the data is zero. Calculate the covariance matrix cov(DD T ). The covariance matrix describes the linear relationship between data features. Perform eigenvalue decomposition on the covariance matrix. Here, singular value decomposition is used to decompose the covariance matrix into eigenvalues ​​and eigenvectors. The eigenvalues ​​represent the variance in the data, and the eigenvectors represent the direction of the principal components. Sort by the size of the eigenvalues, and select the eigenvectors corresponding to the first k eigenvalues ​​as the principal components. Among them, k is the dimension that you want to retain. For example, you can determine k so that the numerical difference between the first k eigenvalues ​​and the last k eigenvalues ​​is extremely large, then you can choose to retain such k as much as possible. Finally, project the original data onto the selected principal component to form a new feature space. Projection can be achieved by matrix multiplication. At this time, the text vector is k-dimensional, and the data will be an m*k matrix D p .

[0117] Figure 5 is a schematic diagram of a text clustering module according to an embodiment of the present application, such as Figure 5As shown, the text clustering module 307 includes a data stratification unit 3071, a k-distance calculation unit 3072, a mutual reachability distance calculation unit 3073, a core point determination unit 3074, a minimum spanning tree construction unit 3075, a clustering unit 3076 and a pruning unit 3077. The text clustering module 307 uses the self-improved HDBSCAN algorithm for text clustering. The traditional HDBSCAN algorithm is a relatively advanced hierarchical clustering algorithm. However, since most of the evaluation feedback of financial products is massive, the traditional HDBSCAN algorithm uses density clustering, core distance, mutual reachability and minimum spanning tree algorithms, and the calculation efficiency is slow on big data. In order to improve the efficiency of the algorithm and reduce the waiting time of business personnel, the text clustering module 307 splits part of the calculation and puts it on the GPU for parallel calculation based on the CUDA computing engine.

[0118] The data stratification unit 3071 performs data stratification and parallelization preparation, and can construct m*k tensors on CUDA as containers for storing results, and then D p The matrix is ​​converted into a tensor of the same dimension to facilitate calculation on CUDA.

[0119] The k-distance calculation unit 3072 calculates the k-distance of each data point. For each point, the distance between the point and other points is calculated, and the distances are sorted in ascending order. The kth distance is selected as the k-distance. The distance between each point and other points is calculated in parallel for each feature when calculating the distance, thereby improving the calculation efficiency. After the calculation is completed, the sorting algorithm is run, and the first k distances are taken and stored in the m*k tensor. The k-distance of data point A is counted as d core-k (A).

[0120] The mutual reachable distance calculation unit 3073 calculates the mutual reachable distance of all points. The mutual reachable distance of two points is defined as d m-k (A,B)=max(d core-k (A),d core-k (B),d(A,B)), where the mutual reachable distance takes the maximum value, d core-k (A),d core-k (B), d(A, B) has been calculated in the previous step, and all values ​​are stored in the tensor. In this step, the distance between each two points can be calculated in parallel by calling the built-in function.

[0121] The core point determination unit 3074 determines the core point based on the mutual reachable distance. Here, a threshold of 1 is pre-set. When the distance is less than the threshold, the point is marked as a core point, otherwise it is not a core point. Then a core graph is constructed using the core points, in which the weight of the edge of the point is their mutual reachable distance.

[0122] The minimum spanning tree construction unit 3075 constructs a minimum spanning tree, and uses the Prim algorithm to generate the minimum spanning tree from the core graph obtained in the previous step.

[0123] The clustering unit 3076 sorts the edges of the minimum spanning tree by weight, clusters the sorted edges, and starts from the edge with the minimum weight, merging the two connected points into a cluster, and a hierarchical clustering result can be generated. For the traditional HDBSCAN clustering algorithm, there is also a part of compressing the clustering tree, and the calculation complexity is relatively high. Therefore, this embodiment adopts a new way to deal with the problem of too many branches of the tree. This part of the calculation amount can be calculated in parallel when running the above steps, thereby speeding up the response speed of the entire application.

[0124] The pruning unit 3077 is a deep neural network classification pruning. This pruning technology is customized and optimized for this embodiment. Business personnel do not need a large and comprehensive feedback analysis, which will only make business personnel deal with a lot of meaningless information. Business personnel need to start from a specific perspective, and improved analysis can be provided in the future. Therefore, the clusters that business personnel need to see are actually less than the clusters generated in practice. Therefore, pruning can efficiently retrieve the clusters that business personnel need to see, which is the key to optimization. Figure 6 is a schematic diagram of a pruning unit according to an embodiment of the present application, such as Figure 6 As shown, the pruning unit 3077 may include a sampling data annotation subunit 30771, a data set construction subunit 30772, a deep learning network construction subunit 30773, a tuning subunit 30774, a neural network prediction subunit 30775, and a pruning subunit 30776. The pre-trained BERT model may be used for text classification tasks, and in the feature extraction module, all texts have been converted into vectors of a specific length.

[0125] The sampling data annotation subunit 30771 annotates the sampling data. Business personnel can specify the required analysis direction or content for each analysis task, sample relevant text from the original text, and annotate it. For example, feedback related to the investment yield of financial products and irrelevant feedback. At this time, a data set with more than or equal to 1,000 data containing positive and negative samples can be constructed. For general machine learning classification algorithms, these data are relatively small, but using the technology of fine-tuning the BERT model for binary classification, the accuracy rate can reach about 89.61%.

[0126] Constructing the data set subunit 30772 includes constructing a training set and a validation set, and the previously annotated text data can be divided into a training set and a validation set in a ratio of 8:2.

[0127] Construct deep learning network subunit 30773 to connect the BERT model network structure with an output layer, which has only two neurons, and use the sigmoid activation function. The loss function is the cross entropy loss function adapted to the binary sigmoid activation function. Use the Adam optimizer to optimize and start the back propagation algorithm.

[0128] The tuning subunit 30774 uses the validation set for tuning. After training with the training set, the validation set is used to adjust parameters and verify the model effect.

[0129] The neural network prediction subunit 30775 inputs the full text vector into the trained binary classification model to obtain the prediction result. According to the prediction result, pruning is performed according to a specific algorithm.

[0130] The pruning subunit 30776 prunes branches according to the rules, starting from the bottom layer of the hierarchical clustering. If the proportion of irrelevant prediction results in the lowest layer of clustering is greater than a certain threshold (this threshold can be defined by the business personnel), then it can be regarded as this cluster is not the direction that the business personnel want to analyze. At this time, it can be pruned. At the same time, a dictionary is maintained for all texts to record which cluster they belong to and whether they are pruned. After the bottom layer of clusters is traversed, the penultimate layer of clusters is traversed. The difference is that the objects in the penultimate layer of clusters are the bottom layer of clusters. According to similar logic, if the number of the bottom layer of clusters pruned in this cluster is more than the threshold, this cluster will also be pruned. This rule is executed until the second layer of clusters is pruned. The final result is a pruned hierarchical clustering spanning tree.

[0131] The automatic summarization module 308 generates a summary according to the similar logic of the pruning rule. Since the text clustering algorithm only aggregates similar texts and cannot tell the business personnel what the pain points are, an advanced large language model can be used to assist in summarization in this step. Figure 7 is a schematic diagram of an automatic summarization module according to an embodiment of the present application, such as Figure 7 As shown, the automatic summarization and induction module 308 may include an integrated large language model API unit 3081 , a writing prompt word unit 3082 , a splicing prompt word and text unit 3083 , and a generating topic summary unit 3084 .

[0132] The integrated large language model API unit 3081 implements a rule-compliant access body through the API of the large language model, which is similar to the request of an http request. A method is implemented to input the question to be asked to the large language model, return and send the request to the large model, and finally parse the result according to the structure returned by the large model.

[0133] Writing prompt word unit 3082 For the lowest level of clustering, the text in a cluster can be organized into prompt words in this way and fed to the large language model: "I am a user feedback analyst for financial products. I have the following feedback. After reading these texts, what keywords do you think can describe the themes of these texts? Can you help summarize the themes of the following texts?" For the second-to-last level of clustering, the clustering object is the theme, so the prompt words when feeding to the large model can be as follows: "I am a user feedback analyst for financial products. I have the following types of feedback. After reading these types of feedback, what keywords do you think can describe the commonalities of these types? Can you help summarize the commonalities and themes of the following feedback?"

[0134] The prompt word and text concatenation unit 3083 concatenates the above-prepared prompt words and the corresponding text after pruning into a complete question, inputs it into the method mentioned in the integrated large language model API unit 3081, transmits it to the large language model and obtains the answer.

[0135] The subject summary generating unit 3084 can generate the subject summary unit 3084 according to the feedback from the large model. The business personnel can compare the keywords returned by the large model with the summarized subjects, and then generate the final subject according to their business knowledge. The business personnel can obtain the subject of the penultimate cluster by repeating the similar steps. By pushing the categories one by one and traversing all levels from low to high, the themes of all levels of clustering and the themes of the subjects can be obtained.

[0136] Figure 8 is a flow chart of a method for automatically summarizing and summarizing the pain points of financial product users based on natural language processing according to an embodiment of the present application, such as Figure 8 As shown, the process of the method for automatically summarizing the pain points of financial product users based on natural language processing may include the following steps:

[0137] Step S801, collecting feedback data, user feature data and product feature data.

[0138] In the above steps, business personnel collect product feedback from various channels, such as questionnaires and industry feedback, read structured data from online databases, that is, relevant characteristics of the feedback personnel, such as financial preferences, occupation, gender, etc., and collect all unstructured text data from the feedback text.

[0139] Step S802, performing data cleaning on the feedback data, user feature data and product feature data.

[0140] In the above steps, the collected feedback data, user feature data and product feature data are cleaned to remove special characters in the text, such as HTML tags, etc., remove common Chinese stop words, remove unnecessary blank characters, unify the encoding format, and finally perform null value processing, duplicate value processing and convert numbers into strings.

[0141] Step S803, integrating the cleaned feedback data, user feature data and product feature data.

[0142] In the above steps, the cleaned feedback data, user feature data and product feature data are integrated, the key values ​​of the structured data are confirmed and concatenated into strings respectively, and finally the structured data and the unstructured data are concatenated into the final data source.

[0143] Step S804, cleaning the integrated data again.

[0144] In the above steps, the data source obtained after integration is again subjected to the data cleaning process as in step S802 to eliminate data pollution caused by data integration.

[0145] Step S805, converting the cleaned text data into a numerical vector.

[0146] In the above steps, the BERT pre-trained model is used to convert the cleaned text data into an n-dimensional numerical vector.

[0147] Step S806, performing dimensionality reduction processing on the numerical vector.

[0148] In the above steps, the PCA dimensionality reduction technique is used to transform the original high-dimensional n-dimensional numerical vector into a relatively low-dimensional numerical vector retaining the principal components.

[0149] Step S807, clustering the numerical vector after dimensionality reduction to obtain a clustering result.

[0150] In the above steps, the obtained low-dimensional numerical vector is converted into a tensor and placed on the GPU to prepare for parallel computing. The k distance of each data point is calculated, and the mutual reachable distance of all points is calculated. The threshold is defined based on the mutual reachable distance and the core point is determined. The minimum spanning tree is generated based on the core point using the Prim algorithm, and then the edges of the minimum spanning tree are sorted by weight and the sorted edges are clustered. Finally, based on the results, the deep neural network classification and pruning technology is used to screen out the topics that business personnel focus on and the specific feedback text.

[0151] Step S808: Generate a topic summary based on the clustering results.

[0152] In the above steps, the method of constructing the request and parsing the returned big language model's answer using the big language model's API is used. Different prompt words are constructed according to the different objects of summary, whether they are text or topics, and the actual texts are spliced ​​together and sent to the big language model. Finally, feedback opinions are automatically summarized. Business personnel can use these feedback opinions to promptly discover the pain points of users after purchasing financial products and make timely and rapid responses and adjustments.

[0153] This embodiment combines structured data and unstructured data features into a sentence through key-value pairs and pastes them into text, which cleverly expands the text data of the existing technology which only has unstructured data, and brings more information into the feedback information, making pain point discovery more personalized and accurate than the existing technology, and helping business personnel discover more detailed pain points.

[0154] This embodiment uses an improved version of the HDBSCAN hierarchical clustering algorithm, puts the calculation into the CUDA tensor, uses the GPU to parallelize the calculation of the part that can be calculated in parallel, and discards the tedious calculation of the compressed clustering tree, so that the time complexity of this part is reduced, thereby making the response time of the entire system faster, allowing business personnel to analyze the feedback in time after it is given, thereby improving user satisfaction. In the scenario of pain point analysis of financial product feedback, hierarchical clustering is used to facilitate business personnel to discover the commonality of pain point themes and reduce the difficulty of business personnel's analysis.

[0155] This embodiment uses a neural network classification algorithm to prune the spanning tree of hierarchical clustering, reuses the text vector extracted from the previous text features, effectively reduces the system response time, and establishes a rule for pruning objects under a cluster if they are determined by the deep neural network to not meet the requirements of business personnel. At the same time, this rule can be used for clusters at any level. Such a pruning algorithm can quickly help business personnel screen out feedback text and clusters that they are not concerned about, thereby improving work efficiency.

[0156] This embodiment uses the natural language big model to use appropriate prompt words to ask the big prediction model and let it return a summary of the text that the business personnel are interested in. It splices the prompt words and objects in the cluster based on specific rules and is applicable to clusters at all levels. It allows the system to automatically summarize the themes of the clusters. It is more efficient and less time-consuming than the prior art of having business experts manually summarize, and allows business personnel to obtain the themes, types and corresponding analysis of pain points in a timely manner.

[0157] In an embodiment of the present application, feedback information of a product, first feature information, and second feature information of an object for publishing feedback information are obtained, wherein the first feature information is used to represent the feature of the product, and the second feature information is used to represent the feature of the object; the feedback information, the first feature information, and the second feature information are integrated into target information; the target information is clustered to obtain a clustering result, wherein the clustering result is used to represent the association relationship between the topics to which multiple sub-feedback information in the feedback information belongs; based on the clustering result, topic information of multiple sub-feedback information is generated, wherein the topic information is used to represent the topic to which the object's demand information for the product belongs, thereby achieving a technical effect of improving the accuracy of generated topic information, and further solving the technical problem of low accuracy of generated topic information.

[0158] The embodiment of the present application also provides a device for generating topic information. It should be noted that the device for generating topic information of this embodiment can be used to execute the method for generating topic information of the embodiment of the present invention.

[0159] Fig. 9 is a structural diagram of a device for generating topic information according to an embodiment of the present application, such as Fig. 9 As shown, the device 900 for generating topic information may include: an acquisition unit 902 , an integration unit 904 , a clustering unit 906 and a generation unit 908 .

[0160] The acquisition unit 902 is used to acquire feedback information of the product, first characteristic information, and second characteristic information of the object that publishes the feedback information, wherein the first characteristic information is used to represent the characteristics of the product, and the second characteristic information is used to represent the characteristics of the object.

[0161] The integration unit 904 is used to integrate the feedback information, the first feature information and the second feature information into target information.

[0162] The clustering unit 906 is used to cluster the target information to obtain a clustering result, wherein the clustering result is used to represent the association relationship between the topics to which the multiple sub-feedback information in the feedback information belong.

[0163] The generating unit 908 is used to generate topic information of a plurality of sub-feedback information based on the clustering result, wherein the topic information is used to indicate the topic to which the object's demand information for the product belongs.

[0164] Optionally, the integration unit 904 includes: an integration module, used to integrate the feedback information, the first feature information and the second feature information into initial information; a processing module, used to convert the initial information into vector data, and perform dimensionality reduction processing on the vector data to obtain target information.

[0165] Optionally, the processing module includes: a processing submodule, used to perform mean centering processing on the vector data to obtain processed vector data; a determination submodule, used to determine the covariance matrix of the processed vector data; a decomposition submodule, used to perform eigenvalue decomposition on the covariance matrix, and decompose the covariance matrix into eigenvalues ​​and eigenvectors; a projection submodule, used to take the eigenvectors corresponding to the top target number of eigenvalues ​​as principal components, and project the vector data onto the principal components to obtain target information.

[0166] Optionally, before the integration unit 904 is used to integrate the feedback information, the first characteristic information and the second characteristic information into the initial information, the device also includes: a first cleaning unit for cleaning the feedback information, the first characteristic information and the second characteristic information; after the integration unit 904 is used to integrate the feedback information, the first characteristic information and the second characteristic information into the initial information, the device also includes: a second cleaning unit for cleaning the initial information.

[0167] Optionally, the clustering unit 906 includes: a first determination module, used to determine the mutual reachable distance between multiple data points in the target information; a second determination module, used to determine the core points among the multiple data points based on the mutual reachable distance; and a clustering module, used to cluster the target information based on the core points to obtain clustering results.

[0168] Optionally, the clustering module includes: a construction submodule, used to construct a core graph based on core points, and to generate a minimum spanning tree based on the core graph, wherein the weights of the edges between the nodes of the core graph are the mutually reachable distances between the nodes; a clustering submodule, used to sort the edges of the minimum spanning tree according to the weights of the edges of the minimum spanning tree, and to cluster the sorted edges to obtain clustering results.

[0169] Optionally, the generation unit 908 includes: a prediction module, used to call the text classification model, predict the feedback information, and obtain a prediction result, wherein the prediction result is used to indicate the probability that the feedback information satisfies a preset topic; a generation module, used to prune the clustering results based on the prediction results, and generate topic information based on the pruned clustering results.

[0170] Optionally, the generation module includes: a splicing submodule, used to splice the prompt information associated with the product and the pruned clustering results to obtain query information; a generation submodule, used to use the prompt information to guide the large language model to generate reply information matching the query information, wherein the large language model is trained using query information samples and reply information samples; an adjustment submodule, used to adjust the reply information to obtain topic information.

[0171] Optionally, the device also includes: a labeling unit, used to extract sub-feedback information that meets a preset theme from the feedback information, and label the sub-feedback information to obtain labeled data; a construction unit, used to construct a training set based on the labeled data; and a training unit, used to train an initial text classification model based on the training set to obtain a text classification model.

[0172] In the device, feedback information of the product, first characteristic information, and second characteristic information of the object for publishing the feedback information are obtained by the acquisition unit 902, wherein the first characteristic information is used to represent the characteristics of the product, and the second characteristic information is used to represent the characteristics of the object. The feedback information, the first characteristic information, and the second characteristic information are integrated into the target information by the integration unit 904. The target information is clustered by the clustering unit 906 to obtain the clustering result, wherein the clustering result is used to represent the association relationship between the topics to which the multiple sub-feedback information in the feedback information belongs. The topic information of the multiple sub-feedback information is generated based on the clustering result by the generation unit 908, wherein the topic information is used to represent the topic to which the object's demand information for the product belongs, thereby achieving the technical effect of improving the accuracy of the generated topic information, thereby solving the technical problem of low accuracy of the generated topic information.

[0173] It should be noted that the acquisition unit 902, integration unit 904, clustering unit 906 and generation unit 908 correspond to steps S202 to S208, and the four modules are the same as the examples and application scenarios implemented by the corresponding steps, but are not limited to the contents disclosed in the above embodiments. It should be noted that the above modules or units can be hardware components or software components stored in a memory (e.g., memory 104) and processed by one or more processors (e.g., processors 102a, 102b, ..., 102n), and the above modules can also be part of the device and can be run in the computer terminal 10 provided in the embodiment.

[0174] An embodiment of the present application may provide an electronic device, Fig.10 is a structural block diagram of an electronic device according to an embodiment of the present application, such as Fig.10 As shown, the electronic device may include: one or more ( Fig.10 (only one is shown) processor 1002, memory 1004, storage controller, and peripheral interface, wherein the peripheral interface is connected to the radio frequency module, audio module and display.

[0175] Among them, the memory can be used to store software programs and modules, such as program instructions / modules corresponding to the methods and devices in the embodiments of the present application, and the processor executes various functional applications and data processing by running the software programs and modules stored in the memory, that is, realizing the above-mentioned method. The memory 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 may further include a memory remotely arranged relative to the processor, and these remote memories may be connected to the terminal 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.

[0176] It can be understood by those skilled in the art that Fig.10 The structure shown is for illustration only, and the electronic device may also be a smart phone (such as an Android phone, an iOS phone, etc.), a tablet computer, a PDA, a mobile Internet device (Mobile Internet Devices, MID), a PAD, or other terminal devices. Fig.10 The structure of the electronic device is not limited. Fig.10 More or fewer components (such as network interfaces, display devices, etc.) shown in, or having Fig.10 Different configurations are shown.

[0177] A person of ordinary skill in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing the hardware related to the terminal device through a program, and the program can be stored in a computer-readable storage medium, and the storage medium may include: a flash drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0178] The embodiment of the present application further provides a storage medium. Optionally, in this embodiment, the storage medium can be used to store the program code executed by the method for generating subject information provided in the first embodiment.

[0179] Optionally, in this embodiment, the above storage medium may be located in any computer terminal in a computer terminal group in a computer network, or in any mobile terminal in a mobile terminal group.

[0180] The present application also provides a computer program product, which, when executed on a data processing device, is suitable for executing the steps of the method for generating subject information.

[0181] In the above embodiments of the present application, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0182] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0183] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0184] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0185] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, a server or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, disk or optical disk and other media that can store program codes.

[0186] The above is only a preferred implementation of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A method for generating topic information, characterized in that: include: Acquire feedback information of a product, first characteristic information, and second characteristic information of an object for publishing the feedback information, wherein the first characteristic information is used to represent the characteristics of the product, and the second characteristic information is used to represent the characteristics of the object; Integrate the feedback information, the first feature information, and the second feature information into target information; Clustering the target information to obtain a clustering result, wherein the clustering result is used to represent the association relationship between the topics to which the plurality of sub-feedback information in the feedback information belongs; Based on the clustering result, topic information of the plurality of sub-feedback information is generated, wherein the topic information is used to indicate the topic to which the demand information of the object for the product belongs.

2. The method according to claim 1, characterized in that Integrating the feedback information, the first feature information, and the second feature information into target information includes: Integrate the feedback information, the first feature information and the second feature information into initial information; The initial information is converted into vector data, and dimension reduction is performed on the vector data to obtain the target information.

3. The method according to claim 2, characterized in that Performing dimensionality reduction processing on the vector data to obtain the target information includes: Performing mean centering processing on the vector data to obtain processed vector data; Determining a covariance matrix of the processed vector data; Performing eigenvalue decomposition on the covariance matrix to decompose the covariance matrix into eigenvalues ​​and eigenvectors; The eigenvectors corresponding to the top target number of eigenvalues ​​are taken as principal components, and the vector data are projected onto the principal components to obtain the target information.

4. The method according to claim 2, characterized in that: Before integrating the feedback information, the first feature information, and the second feature information into initial information, the method further includes: cleaning the feedback information, the first characteristic information, and the second characteristic information; After integrating the feedback information, the first characteristic information and the second characteristic information into initial information, the method further includes: cleaning the initial information.

5. The method according to claim 1, characterized in that Clustering the target information to obtain a clustering result includes: Determining mutually reachable distances between a plurality of data points in the target information; Determining a core point among the plurality of data points based on the mutual reachable distance; Based on the core points, the target information is clustered to obtain the clustering result.

6. The method according to claim 5, characterized in that Based on the core points, the target information is clustered to obtain a clustering result, including: Based on the core points, a core graph is constructed, and based on the core graph, a minimum spanning tree is generated, wherein the weights of the edges between the nodes of the core graph are the mutually reachable distances between the nodes; The edges of the minimum spanning tree are sorted according to the weights of the edges of the minimum spanning tree, and the sorted edges are clustered to obtain the clustering result.

7. The method according to claim 1, characterized in that Generating topic information of the plurality of sub-feedback information based on the clustering result includes: Calling a text classification model to predict the feedback information and obtain a prediction result, wherein the prediction result is used to indicate the probability that the feedback information meets a preset topic; Based on the prediction result, the clustering result is pruned, and based on the pruned clustering result, the topic information is generated.

8. The method according to claim 7, characterized in that Based on the clustering results after pruning, the topic information is generated, including: splicing the prompt information associated with the product and the pruned clustering result to obtain query information; Using the prompt information, guiding the large language model to generate reply information matching the query information, wherein the large language model is trained using the query information sample and the reply information sample; The reply information is adjusted to obtain the subject information.

9. The method according to claim 7, characterized in that: The method further comprises: Extracting sub-feedback information that meets the preset theme from the feedback information, and marking the sub-feedback information to obtain marked data; Based on the labeled data, construct a training set; The initial text classification model is trained based on the training set to obtain the text classification model.

10. A device for generating topic information, characterized in that: include: an acquisition unit, used to acquire feedback information of a product, first characteristic information, and second characteristic information of an object for publishing the feedback information, wherein the first characteristic information is used to represent the characteristics of the product, and the second characteristic information is used to represent the characteristics of the object; an integration unit, configured to integrate the feedback information, the first feature information and the second feature information into target information; A clustering unit, used for clustering the target information to obtain a clustering result, wherein the clustering result is used to represent the association relationship between the topics to which the plurality of sub-feedback information in the feedback information belongs; A generating unit is used to generate topic information of the plurality of sub-feedback information based on the clustering result, wherein the topic information is used to indicate the topic to which the demand information of the object for the product belongs.

11. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored executable program, wherein when the executable program is executed, the device where the computer-readable storage medium is located is controlled to execute the method according to any one of claims 1 to 9.

12. An electronic device, characterized in that: include: A memory storing an executable program; A processor, configured to run the program, wherein the program executes the method according to any one of claims 1 to 9 when running.

13. A computer program product comprising computer instructions, characterized in that: When the computer instructions are executed by a processor, the steps of the method according to any one of claims 1 to 9 are implemented.

Images