Method and device for determining answer to financial question, program product and electronic equipment

Abstract

The invention discloses a financial question answer determination method and device, a program product and electronic equipment, and relates to the field of artificial intelligence, and the method comprises the steps: processing a to-be-answered target financial question, obtaining the target word vector data of the target financial question, inputting the target word vector data into a preset classification model, and obtaining the answer of the target financial question; outputting a target question type of the target financial question, querying a candidate question set corresponding to a preset question type which is the same as the target question type from a preset corpus, and determining question similarity between the target financial question and each candidate question in the candidate question set by adopting a preset similarity model, and determining the candidate question indicated by the maximum question similarity as a target candidate question, and determining a target answer corresponding to the target financial question from a candidate answer set corresponding to the target candidate question by adopting a preset answer selection model. According to the method and the device, the technical problem of relatively low accuracy of answering financial questions in related technologies is solved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and in particular to a method for determining answers to financial questions and a device, program product, and electronic device thereof. Background Art

[0002] At present, when customers use financial institutions' services, they often encounter some problems such as business handling procedures, scope of application, and operating guidelines, and most of the solutions they adopt are to go to the branch for consultation or make phone consultations. However, going to the branch for consultation has a lag and requires customers to spend extra time; for telephone consultations, customers' same questions need to be answered one by one, which requires a large number of customer service resources.

[0003] In the relevant technology, intelligent customer service can be used to answer customers' questions, but the current intelligent customer service mainly focuses on short text processing, does not have sufficient scene understanding and contextual dialogue capabilities, and cannot give users a clear and high-quality answer based on complex sentence content and contextual context.

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

[0005] The embodiments of the present invention provide a method for determining an answer to a financial question and a device, program product, and electronic device thereof, so as to at least solve the technical problem of low accuracy in answering financial questions in related technologies.

[0006] According to one aspect of an embodiment of the present invention, a method for determining an answer to a financial question is provided, comprising: processing a target financial question to be answered to obtain target word vector data of the target financial question, and inputting the target word vector data into a preset classification model to output a target question type of the target financial question; querying a set of candidate questions corresponding to a preset question type that is the same as the target question type from a preset corpus, wherein the preset corpus comprises: a set of historical questions corresponding to a plurality of preset question types, each historical question in the set of historical questions corresponding to a set of historical answers; using a preset similarity model to determine the question similarity between the target financial question and each candidate question in the set of candidate questions, and determining the candidate question indicated by the maximum question similarity as the target candidate question; using a preset answer selection model to determine a target answer corresponding to the target financial question from the set of candidate answers corresponding to the target candidate question.

[0007] Furthermore, the step of processing the target financial question to be answered to obtain target word vector data of the target financial question includes: obtaining a financial stop word list and a preset character table of the financial institution, and filtering the target financial question based on the financial stop word list and the preset character table to obtain preset text data, wherein the financial stop word list includes: a plurality of preset stop words, which are pre-set words that need to be filtered; performing word segmentation on the preset text data to obtain a word sequence; and performing vectorization on the word sequence to obtain target word vector data.

[0008] Furthermore, before inputting the target word vector data into the preset classification model and outputting the target problem type of the target financial problem, it also includes: constructing a densely connected network structure and a classification structure, wherein the densely connected network structure includes at least: a convolutional layer, a pooling layer, and a fully connected layer; the classification structure is composed of N convolutional layers and N sampling layers alternating, and N is a positive integer greater than or equal to 2; based on the densely connected network structure and the classification structure, an initial classification model is constructed; the first historical data is collected, and the initial classification model is trained using the first historical data to obtain a preset classification model, wherein the first historical data includes at least: multiple historical problems and a problem type label marked for each historical problem.

[0009] Furthermore, the target word vector data is input into a preset classification model to output a target problem type of a target financial problem, including: using a densely connected network structure to perform feature extraction on the target word vector data to obtain a feature vector; and using a classification structure to process the feature vector to obtain a target problem type.

[0010] Furthermore, before querying a set of candidate questions corresponding to a preset question type that is the same as the target question type from a preset corpus, it also includes: collecting historical questions over a historical time period and a set of historical answers to the historical questions; inputting each historical question into a preset classification model, and outputting the preset question type corresponding to the historical question; adding all historical questions indicated by the same preset question type to the same set of historical questions; and constructing a preset corpus based on the set of historical questions corresponding to all preset question types and the set of historical answers corresponding to each historical question in the set of historical questions.

[0011] Furthermore, before using a preset similarity model to determine the problem similarity between the target financial problem and each candidate problem in the candidate problem set, it also includes: constructing a pre-training structure and a densely connected bidirectional gated loop structure, wherein the pre-training structure is used to perform feature processing on the word sequence of the input problem to obtain a contextual feature representation of the problem; the contextual feature representation vector includes at least: a word vector, a segment vector, and a position vector; the densely connected bidirectional gated loop structure is used to process the contextual feature representations of the two input problems to obtain the problem similarity between the two problems; based on the pre-training structure and the densely connected bidirectional gated loop structure, an initial similarity model is constructed; second historical data is collected, and the initial similarity model is trained using the second historical data to obtain a preset similarity model, wherein the second historical data includes at least: multiple historical problem pairs and similarity labels annotated for each group of historical problem pairs.

[0012] Furthermore, before the preset answer selection model is used to determine the target answer corresponding to the target financial question from the candidate answer set corresponding to the target candidate question, it also includes: marking each historical answer in the historical answer set corresponding to each historical question in the preset corpus to obtain a score label for the historical answer; constructing third historical data based on all historical answers marked with score labels and historical questions corresponding to the historical answers; using the third historical data to train the initial answer selection model, and after the initial answer selection model is trained, a preset answer selection model is obtained, wherein the initial answer selection model at least includes: an embedding layer, a dynamic attention layer, a multi-strategy interaction layer, a bidirectional gated loop layer, and an output layer; the embedding layer is used to perform feature processing on the word sequence of the input question The first vector representation of the question is obtained, and the word sequence of the input answer is feature processed to obtain the second vector representation of the answer; the dynamic attention layer is used to weight the first vector representation and the second vector representation to obtain the first weighted vector representation and the second weighted vector representation; the multi-strategy interaction layer is used to match the first weighted vector representation and the second weighted vector representation to obtain the semantic relationship between the first weighted vector representation and the second weighted vector representation; the bidirectional gated recurrent layer is used to process the first weighted vector representation and the second weighted vector representation based on the semantic relationship to obtain the first context vector representation and the second context vector representation; the output layer is used to process the first context vector representation and the second context vector representation to obtain the score between the question and the answer.

[0013] Furthermore, the step of using a preset answer selection model to determine a target answer corresponding to a target financial question from a set of candidate answers corresponding to the target candidate question includes: using the preset answer selection model to determine a score between the target financial question and each candidate answer in the candidate answer set; and determining the candidate answer corresponding to the maximum score as the target answer corresponding to the target financial question.

[0014] According to another aspect of an embodiment of the present invention, there is also provided a device for determining answers to financial questions, comprising: a processing unit, for processing a target financial question to be answered, obtaining target word vector data of the target financial question, and inputting the target word vector data into a preset classification model, and outputting a target question type of the target financial question; a query unit, for querying a set of candidate questions corresponding to a preset question type that is the same as the target question type from a preset corpus, wherein the preset corpus comprises: a set of historical questions corresponding to a plurality of preset question types, each historical question in the set of historical questions corresponding to a set of historical answers; a first determination unit, for determining the question similarity between the target financial question and each candidate question in the set of candidate questions by adopting a preset similarity model, and determining the candidate question indicated by the maximum question similarity as the target candidate question; and a second determination unit, for determining the target answer corresponding to the target financial question from the set of candidate answers corresponding to the target candidate question by adopting a preset answer selection model.

[0015] Furthermore, the processing unit includes: a first processing module, used to obtain a financial stop word list and a preset character table of a financial institution, and based on the financial stop word list and the preset character table, filter the target financial problem to obtain preset text data, wherein the financial stop word list includes: multiple preset stop words, which are pre-set words that need to be filtered; a second processing module, used to perform word segmentation processing on the preset text data to obtain a word sequence; a third processing module, used to perform vectorization processing on the word sequence to obtain target word vector data.

[0016] Furthermore, the determination device also includes: a first construction module, which is used to construct a densely connected network structure and a classification structure before inputting the target word vector data into a preset classification model and outputting the target problem type of the target financial problem, wherein the densely connected network structure includes at least: a convolutional layer, a pooling layer, and a fully connected layer; the classification structure is composed of N convolutional layers and N sampling layers alternating, and N is a positive integer greater than or equal to 2; a second construction module, which is used to construct an initial classification model based on the densely connected network structure and the classification structure; a first training module, which is used to collect first historical data and use the first historical data to train the initial classification model to obtain a preset classification model, wherein the first historical data includes at least: multiple historical problems and a problem type label marked for each historical problem.

[0017] Furthermore, the processing unit also includes: a first extraction module, which is used to extract features from the target word vector data using a densely connected network structure to obtain a feature vector; and a fourth processing module, which is used to process the feature vector using a classification structure to obtain a target question type.

[0018] Furthermore, the determination device also includes: a first collection module, which is used to collect historical questions over a historical time period and a set of historical answers to historical questions before querying a set of candidate questions corresponding to a preset question type that is the same as the target question type from a preset corpus; a first input module, which is used to input each historical question into a preset classification model and output a preset question type corresponding to the historical question; a first adding module, which is used to add all historical questions indicated by the same preset question type to the same set of historical questions; and a preset corpus is constructed based on the set of historical questions corresponding to all preset question types and the set of historical answers corresponding to each historical question in the set of historical questions.

[0019] Furthermore, the determination device also includes: a third construction module, which is used to construct a pre-training structure and a densely connected bidirectional gated loop structure before using a preset similarity model to determine the problem similarity between the target financial problem and each candidate problem in the candidate problem set, wherein the pre-training structure is used to perform feature processing on the word sequence of the input problem to obtain the context feature representation of the problem; the context feature representation vector includes at least: a word vector, a segment vector and a position vector; the densely connected bidirectional gated loop structure is used to process the context feature representation of the two input problems to obtain the problem similarity between the two problems; a fourth construction module, which is used to construct an initial similarity model based on the pre-training structure and the densely connected bidirectional gated loop structure; a second acquisition module, which is used to collect second historical data and use the second historical data to train the initial similarity model to obtain a preset similarity model, wherein the second historical data includes at least: multiple historical problem pairs and similarity labels annotated for each group of historical problem pairs.

[0020] Furthermore, the determination device also includes: a first labeling module, which is used to use a preset answer selection model to label each historical answer in the historical answer set corresponding to each historical question in the preset corpus before determining the target answer corresponding to the target financial question from the candidate answer set corresponding to the target candidate question, so as to obtain a score label for the historical answer; a fifth construction module, which is used to construct third historical data based on all historical answers labeled with score labels and historical questions corresponding to the historical answers; a second training module, which is used to use the third historical data to train the initial answer selection model, and obtain the preset answer selection model after the initial answer selection model is trained, wherein the initial answer selection model at least includes: an embedding layer, a dynamic attention layer, a multi-strategy interaction layer, a bidirectional gated loop layer, and an output layer; the embedding layer is used to The invention is used to perform feature processing on the word sequence of the input question to obtain the first vector representation of the question, and to perform feature processing on the word sequence of the input answer to obtain the second vector representation of the answer; the dynamic attention layer is used to weight the first vector representation and the second vector representation to obtain the first weighted vector representation and the second weighted vector representation; the multi-strategy interaction layer is used to match the first weighted vector representation and the second weighted vector representation to obtain the semantic relationship between the first weighted vector representation and the second weighted vector representation; the bidirectional gated recurrent layer is used to process the first weighted vector representation and the second weighted vector representation based on the semantic relationship to obtain the first context vector representation and the second context vector representation; the output layer is used to process the first context vector representation and the second context vector representation to obtain the score between the question and the answer.

[0021] Furthermore, the second determination unit includes: a first determination module, used to use a preset answer selection model to determine the score between the target financial question and each candidate answer in the candidate answer set; and a second determination module, used to determine the candidate answer corresponding to the maximum score as the target answer corresponding to the target financial question.

[0022] According to another aspect of an embodiment of the present invention, there is also provided a computer program product, comprising a non-volatile computer-readable storage medium, wherein the non-volatile computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for determining the answer to any one of the above-mentioned financial questions is implemented.

[0023] According to another aspect of an embodiment of the present invention, there is also provided an electronic device, comprising one or more processors and a memory, wherein the memory is used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method for determining the answer to any one of the above-mentioned financial questions.

[0024] In the present invention, the target financial question to be answered is processed to obtain target word vector data of the target financial question, and the target word vector data is input into a preset classification model, the target question type of the target financial question is output, and a set of candidate questions corresponding to a preset question type that is the same as the target question type is queried from a preset corpus, and a preset similarity model is used to determine the question similarity between the target financial question and each candidate question in the candidate question set, and the candidate question indicated by the maximum question similarity is determined as the target candidate question, and a preset answer selection model is used to determine the target answer corresponding to the target financial question from the candidate answer set corresponding to the target candidate question, thereby solving the technical problem of low accuracy in answering financial questions in the related art.

[0025] In the present invention, the target question type of the target financial question to be answered can be determined by first using a preset classification model, and then the problem similarity between the target financial question and questions of the same type can be calculated by using a preset similarity model. Thereafter, a target answer corresponding to the target financial question can be determined from a set of candidate answers corresponding to the target candidate question indicated by the maximum question similarity by using a preset answer selection model, thereby effectively improving the accuracy and efficiency of automatic answers to financial questions and achieving the technical effect of quickly and accurately answering customers' questions. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0027] Figure 1 A hardware structure block diagram of a computer terminal (or mobile device) for implementing a method for determining an answer to a financial question is shown;

[0028] Figure 2 is a flow chart of a method for determining an answer to a financial question according to Embodiment 1 of the present invention;

[0029] Figure 3 is a schematic diagram of a classification structure according to Embodiment 1 of the present invention;

[0030] Figure 4 is a schematic diagram of a preset similarity model according to Embodiment 1 of the present invention;

[0031] Figure 5 is a schematic diagram of a processing flow of a preset answer selection model according to Embodiment 1 of the present invention;

[0032] Figure 6 is a schematic diagram of an optional device for determining an answer to a financial question according to an embodiment of the present invention;

[0033] Figure 7 is a structural block diagram of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0034] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0035] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention 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 interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are 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 method for determining answers to financial questions and the device thereof in the present invention can be used in the field of artificial intelligence when determining answers to financial questions, and can also be used in any field other than the field of artificial intelligence when determining answers to financial questions. The present invention does not limit the application field of the method for determining answers to financial questions and the device thereof.

[0037] It should be noted that the relevant information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) collected and involved in the present invention are all 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 the relevant laws, regulations and standards of the relevant regions, 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 between the system and the relevant users or organizations. Before obtaining relevant information, it is necessary to send an acquisition request to the aforementioned user or organization through the interface, and obtain relevant information after receiving the consent information fed back by the aforementioned user or organization.

[0038] In the present invention, through deep learning technology, automatic classification, similarity matching and answer selection of financial problems are realized, which improves the service efficiency and customer satisfaction of financial institutions and reduces labor costs, which is of great significance for promoting the digital and intelligent development of financial institutions.

[0039] The present invention is described in detail below in conjunction with various embodiments.

[0040] Example 1

[0041] According to an embodiment of the present application, an embodiment of a method for determining answers to financial questions is also 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.

[0042] The method embodiment provided in the first embodiment of the present application can be executed in a mobile terminal, a computer terminal or a similar computing device. Figure 1 The hardware structure block diagram of a computer terminal (or mobile device) for implementing a method for determining an answer to a financial question is shown. Figure 1 As shown, the computer terminal 10 (or mobile device) may include one or more ( Figure 1 102a, 102b, ..., 102n are used to illustrate) processor 102 (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, a keyboard, a cursor control device, 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, wherein the network interface may be connected to a wired and / or wireless network. A person of ordinary skill in the art will understand 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.

[0043] 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).

[0044] 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 determining answers to financial questions 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 determining answers to financial questions described above is implemented. 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.

[0045] 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.

[0046] 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).

[0047] Under the above operating environment, this application provides Figure 2 Method for determining answers to the financial questions shown. Figure 2 is a flow chart of a method for determining an answer to a financial question according to Embodiment 1 of the present invention. Figure 2 As shown, the method comprises the following steps:

[0048] Step S201, processing the target financial question to be answered, obtaining target word vector data of the target financial question, and inputting the target word vector data into a preset classification model to output the target question type of the target financial question.

[0049] In the embodiment of the present invention, when handling financial services, customers can ask questions to the intelligent question-answering system of the financial institution. The intelligent question-answering system can pre-process the questions asked by customers (i.e., the target financial questions to be answered), i.e., remove, clean and filter useless information in the questions, so as to facilitate the subsequent rapid and accurate processing of the questions.

[0050] Here, useless information refers to the noise that affects feature extraction, such as frequently used but meaningless words (i.e., stop words) in the financial field and special characters in Chinese and English.

[0051] In an embodiment of the present invention, the target word vector data obtained after processing the target financial question may be input into a preset classification model to obtain a target question type of the target financial question.

[0052] Here, the problem types corresponding to financial problems include but are not limited to: account management business type, transaction business type, financial management business type, financial product business type, etc.

[0053] Step S202, querying a set of candidate questions corresponding to a preset question type that is the same as the target question type from a preset corpus, wherein the preset corpus includes: a set of historical questions corresponding to a plurality of preset question types, each historical question in the set of historical questions corresponds to a set of historical answers.

[0054] In an embodiment of the present invention, a preset corpus may be queried to obtain a set of candidate questions corresponding to a preset question type that is the same as the target question type, and the set of candidate questions includes multiple candidate questions of the same question type.

[0055] In the embodiment of the present invention, the preset corpus is a pre-constructed corpus, including: a set of historical questions corresponding to a plurality of preset question types, and each historical question in each set of historical questions corresponds to a set of historical answers.

[0056] Step S203: using a preset similarity model, determine the problem similarity between the target financial problem and each candidate problem in the candidate problem set, and determine the candidate problem indicated by the maximum problem similarity as the target candidate problem.

[0057] In an embodiment of the present invention, the target financial question and each candidate question in the candidate question set may be input into a preset similarity model in turn to obtain the question similarity between the target financial question and each candidate question, and then the candidate question indicated by the maximum question similarity is determined as the target candidate question, so as to facilitate the subsequent determination of the target answer corresponding to the target financial question.

[0058] Step S204: using a preset answer selection model, determine a target answer corresponding to the target financial question from a set of candidate answers corresponding to the target candidate question.

[0059] In an embodiment of the present invention, the target financial question and each candidate answer in the candidate answer set may be sequentially input into a preset answer selection model to determine a target answer corresponding to the target financial question based on the score between the target financial question and each candidate answer.

[0060] In summary, by presetting the classification model, the target question type of the target financial question to be answered is determined, and then by presetting the similarity model, the problem similarity between the target financial question and questions of the same type is calculated. After that, by presetting the answer selection model, the target answer corresponding to the target financial question is determined from the candidate answer set corresponding to the target candidate question indicated by the maximum question similarity, thereby effectively improving the accuracy and efficiency of automatic answering to financial questions, achieving the technical effect of quickly and accurately answering customers' questions, and thus solving the technical problem of low accuracy in answering financial questions in related technologies.

[0061] In order to accurately handle financial issues, in the method for determining answers to financial questions provided in Example 1 of the present application, a financial stop word list and a preset character table of a financial institution are obtained, and based on the financial stop word list and the preset character table, the target financial issue is filtered to obtain preset text data, wherein the financial stop word list includes: a plurality of preset stop words, which are pre-set words that need to be filtered; the preset text data is segmented to obtain a word sequence; the word sequence is vectorized to obtain target word vector data.

[0062] In an embodiment of the present invention, a stop word list specifically for the financial field (i.e., a financial stop word list) and a preset character list can be first obtained from a financial institution. The financial stop word list lists high-frequency words that need to be ignored when processing text (i.e., preset stop words, which are very common in language but usually do not carry much specific information or semantics, and therefore need to be ignored or filtered out in text processing, information retrieval, and natural language processing). The preset character list lists special characters, including: punctuation marks, numbers, specific financial symbols, etc.

[0063] Then, the target financial questions (i.e., questions asked by customers) are filtered using the obtained financial stop word list and preset character list to remove stop words and preset characters in the text, thereby eliminating text noise, retaining core information, and obtaining preset text data. After that, a word segmentation tool can be used to segment the preset text data, that is, to divide the continuous text into words one by one to obtain a word sequence. After that, the obtained word sequence can be converted into a vector form that can be understood by the machine learning model. For example, using a word vector model, each word segmentation is converted into its corresponding word vector to obtain the target word vector data for subsequent deep learning model processing and analysis.

[0064] In order to accurately construct a preset classification model, in the method for determining the answer to a financial question provided in Example 1 of the present application, before inputting the target word vector data into the preset classification model and outputting the target question type of the target financial question, a densely connected network structure and a classification structure are constructed, wherein the densely connected network structure includes at least: a convolutional layer, a pooling layer, and a fully connected layer; the classification structure is composed of N convolutional layers and N sampling layers alternatingly, where N is a positive integer greater than or equal to 2; based on the densely connected network structure and the classification structure, an initial classification model is constructed; first historical data is collected, and the initial classification model is trained using the first historical data to obtain a preset classification model, wherein the first historical data includes at least: multiple historical questions and a question type label marked for each historical question.

[0065] In an embodiment of the present invention, a densely connected network structure can be constructed first. The densely connected network structure is a network structure that uses dense connections in a convolutional neural network, that is, each layer is directly connected to all subsequent layers to form a dense feature transfer path, which helps to reduce the number of parameters, improve the training efficiency of the model, and promote feature reuse, so that the model can learn a richer feature representation. In this embodiment, the densely connected network structure includes at least: a convolution layer, a pooling layer, and a fully connected layer. In addition, a classification structure can be constructed, which is composed of N (N is a positive integer greater than or equal to 2) convolution layers and N sampling layers alternately. The purpose of this structural design is to extract the key features of the question from different levels, first capture local features through the convolution layer, and then reduce the feature dimension and maintain important information through the sampling layer. After multiple alternations, a compact and information-rich feature vector is finally obtained. Then, these feature vectors can be further integrated through the fully connected layer for the final classification decision.

[0066] Figure 3 is a schematic diagram of a classification structure according to Embodiment 1 of the present invention, such as Figure 3 As shown, the classification structure can be composed of two convolutional layers (C1, C2) and two sampling layers (S1, S2) alternating.

[0067] In an embodiment of the present invention, an initial classification model can be constructed based on a densely connected network structure and a classification structure. Before training, the weights and parameters of this model are randomly initialized. Then, the first historical data is collected, which may include: multiple historical questions and their corresponding question type labels. The historical questions may be questions asked by customers before, and the corresponding question type labels are pre-marked to indicate which specific category the questions belong to. Afterwards, the initial classification model can be trained using these labeled historical data, for example, by adjusting the model parameters through back propagation so that the model can learn the mapping relationship from questions to question types. After multiple iterations and optimizations, a preset classification model is obtained. The model can also make accurate classification predictions on new questions that have never been seen.

[0068] In order to improve the accuracy of determining the question type, in the method for determining the answer to a financial question provided in Example 1 of the present application, a densely connected network structure is used to extract features of the target word vector data to obtain a feature vector; and a classification structure is used to process the feature vector to obtain the target question type.

[0069] In an embodiment of the present invention, the target word vector data can be input into a densely connected network structure to extract features from the target word vector data through the densely connected network structure to obtain a feature vector. In this embodiment, the densely connected network can make the output of each layer serve as the input of all subsequent layers, forming a dense connection pattern, which helps the model learn the complex relationships and deep features of the words in the question.

[0070] Then, after feature extraction, the obtained feature vector is input into the classification structure for processing to obtain the target question type. In this embodiment, the classification structure specifically refers to a part consisting of multiple convolutional layers and pooling layers alternating, as well as the final fully connected layer and classification layer. The classification structure converts the feature vector into a probability prediction for various predefined question types through a series of mathematical operations. The output layer usually uses a softmax function (i.e., a mathematical function) to map the feature vector to a probability distribution of the question type classification, thereby obtaining the target question type.

[0071] In order to accurately construct a preset corpus, in the method for determining answers to financial questions provided in Example 1 of the present application, before querying a set of candidate questions corresponding to a preset question type that is the same as the target question type from the preset corpus, historical questions over a historical time period and a set of historical answers to the historical questions are collected; each historical question is input into a preset classification model, and the preset question type corresponding to the historical question is output; all historical questions indicated by the same preset question type are added to the same set of historical questions; and a preset corpus is constructed based on the set of historical questions corresponding to all preset question types and the set of historical answers corresponding to each historical question in the set of historical questions.

[0072] In an embodiment of the present invention, historical questions (i.e., questions previously raised by customers through telephone, online consultation, and other channels) and their corresponding historical answer sets within a historical time period can be collected. These historical data include various questions raised by customers in different situations and the corresponding accurate answers given by officials or customer service managers. Then, each collected historical question is input into the preset classification model previously constructed and trained, and the model outputs the preset question type corresponding to the historical question. Here, the preset question type refers to a question category predefined according to business needs and common scenarios of customer consultation, such as account information inquiry, loan business, credit card service, financial product consultation, and other types. Afterwards, all historical questions of the same category indicated by the preset question type are added to the same historical question set. For example, all historical questions about "account information inquiry" will be grouped into one set, and all historical questions about "financial product consultation" will be grouped into another set, etc. Afterwards, a preset corpus is constructed based on the historical question sets corresponding to all the above preset question types, and the historical answer sets corresponding to the historical questions in each set. This corpus contains historical questions grouped by category and their correct answers, providing rich training and reference data for subsequent question similarity detection and answer selection.

[0073] In order to accurately construct a preset similarity model, in the method for determining the answer to a financial question provided in Example 1 of the present application, before using the preset similarity model to determine the problem similarity between the target financial question and each candidate question in the candidate question set, a pre-training structure and a densely connected bidirectional gated loop structure are constructed, wherein the pre-training structure is used to perform feature processing on the word sequence of the input question to obtain a context feature representation of the question; the context feature representation vector includes at least: a word vector, a segment vector, and a position vector; the densely connected bidirectional gated loop structure is used to process the context feature representations of the two input questions to obtain the problem similarity between the two questions; based on the pre-training structure and the densely connected bidirectional gated loop structure, an initial similarity model is constructed; second historical data is collected, and the initial similarity model is trained using the second historical data to obtain a preset similarity model, wherein the second historical data includes at least: multiple historical question pairs and similarity labels annotated for each group of historical question pairs.

[0074] In an embodiment of the present invention, a pre-training structure can be constructed first, for example, a pre-training model BERT (Bidirectional Encoder Representations from Transformers), which can perform feature processing on the word sequence of the input question (i.e., the word sequence obtained by segmenting the question), so as to obtain a feature representation that can capture the semantic context. In this process, the pre-training model generates word vectors, segment vectors, and position vectors, which constitute the context feature representation of the input question, wherein the word vector represents the semantic features of each word; the segment vector represents the semantic features of the question fragment, which helps to understand the local structure of the sentence; the position vector represents the position information of the word in the sentence, which is crucial to capturing the sentence structure and semantic order.

[0075] In addition, a densely connected bidirectional gated recurrent structure (BiGRU (Birectional Gated Recurrent Unit)) can be constructed, which is used to process the context feature representation of the two input questions, and calculate the similarity between them by comparing and analyzing the features of the two questions. In this embodiment, the bidirectional GRU (Gated Recurrent Unit) can process sequence data in two directions, and can more effectively utilize historical information and context by combining dense connections, so as to more accurately evaluate the similarity of the two questions.

[0076] In the embodiment of the present invention, based on the pre-training structure and the densely connected bidirectional gated loop structure, an initial similarity model can be constructed, which can predict the similarity between question pairs based on the comparison of feature vectors. Then, the second historical data is collected and the model is trained to learn the similarity patterns between different question pairs, and finally a preset similarity model is obtained, which can automatically and accurately detect semantically similar questions.

[0077] In the embodiment of the present invention, the second historical data includes: a plurality of historical question pairs and a similarity label marked for each group of historical question pairs.

[0078] In this embodiment, the label of a similar question pair may be 1, and the label of a dissimilar question pair may be 1. As shown in the following Table 1, labels between question pairs (question 1 and question 2) are exemplarily shown.

[0079] Table 1

[0080]

[0081] Figure 4 is a schematic diagram of a preset similarity model according to Embodiment 1 of the present invention, such as Figure 4 As shown, the preset similarity model can be divided into: text preprocessing layer (i.e., pre-training structure), densely connected BiGRU and interactive classification layer (i.e., densely connected bidirectional gated loop structure). Among them, the text preprocessing layer can segment the input questions (such as question 1 and question 2), and then obtain the context feature representation vector corresponding to the segmentation through BERT (including: word vector, segment vector and position vector), and then the BiGRU in the densely connected BiGRU can process the context feature representation vector, and through the pooling layer, the processing result corresponding to the question is input into the interactive layer on the interactive classification layer, and after the features of the two questions are interactively fused, the similarity label between the two questions is calculated through the softmax function.

[0082] In order to accurately construct a preset answer selection model, in the method for determining the answer to a financial question provided in Example 1 of the present application, a preset answer selection model is used. Before determining the target answer corresponding to the target financial question from the candidate answer set corresponding to the target candidate question, each historical answer in the historical answer set corresponding to each historical question in the preset corpus is labeled to obtain a score label for the historical answer; based on all historical answers labeled with score labels and historical questions corresponding to the historical answers, a third historical data is constructed; the third historical data is used to train an initial answer selection model, and after the initial answer selection model is trained, a preset answer selection model is obtained, wherein the initial answer selection model at least includes: an embedding layer, a dynamic attention layer, a multi-strategy interaction layer, a bidirectional gated loop layer, and an output layer; the embedding layer is used to The word sequence of the input question is feature processed to obtain the first vector representation of the question, and the word sequence of the input answer is feature processed to obtain the second vector representation of the answer; the dynamic attention layer is used to weight the first vector representation and the second vector representation to obtain the first weighted vector representation and the second weighted vector representation; the multi-strategy interaction layer is used to match the first weighted vector representation and the second weighted vector representation to obtain the semantic relationship between the first weighted vector representation and the second weighted vector representation; the bidirectional gated recurrent layer is used to process the first weighted vector representation and the second weighted vector representation based on the semantic relationship to obtain the first context vector representation and the second context vector representation; the output layer is used to process the first context vector representation and the second context vector representation to obtain the score between the question and the answer.

[0083] In an embodiment of the present invention, each historical answer in a set of historical answers corresponding to each historical question in a preset corpus can be labeled first. The labeling here mainly refers to assigning a score label to each historical answer. This label is usually based on the accuracy and relevance of the answer to indicate the degree of match between the answer and the question.

[0084] As shown in the following Table 2, the labels between the question and the answer are shown as an example.

[0085] Table 2

[0086]

[0087] In the embodiment of the present invention, the third historical data is constructed based on all historical answers marked with score tags and the historical questions corresponding thereto, so as to train the initial answer selection model using the third historical data.

[0088] In this embodiment of the present invention, the model at least includes: an embedding layer, a dynamic attention layer, a multi-strategy interaction layer, a bidirectional gated recurrent layer (BiGRU) and an output layer, and the functions of each layer are as follows:

[0089] Embedding layer: Perform feature processing on the word sequences of the input question and answer, and convert each word into a vector representation through word embedding technology (such as BERT) to obtain the first vector representation of the question and the second vector representation of the answer.

[0090] Dynamic attention layer: The first vector representation and the second vector representation are weighted, and the key parts of the question and answer are emphasized through the dynamic attention mechanism to obtain the first weighted vector representation and the second weighted vector representation.

[0091] Multi-strategy interaction layer: The first weighted vector representation and the second weighted vector representation are matched, and multiple strategies (such as complete matching and attention matching) are used to extract the deep semantic relationship between the question and the answer.

[0092] Bidirectional gated recurrent layer: Based on the semantic relationship obtained by the multi-strategy interaction layer, the first weighted vector representation and the second weighted vector representation are further processed to obtain the first context vector representation and the second context vector representation. These two representations contain richer context information.

[0093] Output layer: Use classification or regression technology (such as softmax) to process the first context vector representation and the second context vector representation to obtain the score between the question and the candidate answer. This score reflects the degree of match between the answer and the question.

[0094] In an embodiment of the present invention, after the initial answer selection model is trained, the model parameters are adjusted and optimized to ultimately obtain a preset answer selection model, which can automatically evaluate the match between customer questions and candidate answers and select the most appropriate answer.

[0095] Figure 5 is a schematic diagram of the processing flow of the preset answer selection model according to Embodiment 1 of the present invention. Figure 5 As shown, the input question and answer can be segmented, and then the vector representation corresponding to the segmentation can be obtained through the embedding layer (BERT) (that is, the first vector representation of the question and the second vector of the answer, each vector includes: word vector, segment vector and position vector), and then the first vector representation of the question and the second vector representation of the answer are weighted through the dynamic attention layer to obtain the first weighted vector representation and the second weighted vector representation, and then the first weighted vector representation and the second weighted vector representation are matched through the multi-strategy interaction layer to obtain the semantic relationship, and then the first weighted vector representation and the second weighted vector representation are processed based on the semantic relationship through the bidirectional gated recurrent layer to obtain the first context vector representation and the second context vector representation, and finally, the first context vector representation and the second context vector representation are processed through the output layer to obtain the score between the question and the answer.

[0096] In order to improve the accuracy of determining the answer corresponding to the question, in the method for determining the answer to the financial question provided in Example 1 of the present application, a preset answer selection model is used to determine the score between the target financial question and each candidate answer in the candidate answer set; the candidate answer corresponding to the maximum score is determined as the target answer corresponding to the target financial question.

[0097] In an embodiment of the present invention, the target financial question and each candidate answer in the candidate answer set can be sequentially input into a preset answer selection model to obtain the score between the target financial question and each candidate answer through the processing of the embedding layer, the dynamic attention layer, the multi-strategy interaction layer, the bidirectional gated loop layer and the output layer. Then, the candidate answer corresponding to the maximum score is determined as the target answer corresponding to the target financial question.

[0098] The method for determining the answer to a financial question provided in the embodiment of the present application determines the target question type of the target financial question to be answered through a preset classification model, then calculates the question similarity between the target financial question and questions of the same type through a preset similarity model, and then determines the target answer corresponding to the target financial question from the set of candidate answers corresponding to the target candidate question indicated by the maximum question similarity through a preset answer selection model, thereby effectively improving the accuracy and efficiency of automatic answers to financial questions, achieving the technical effect of quickly and accurately answering customers' questions, and thereby solving the technical problem of low accuracy in answering financial questions in related technologies.

[0099] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0100] Example 2

[0101] The embodiment of the present application also provides a device for determining an answer to a financial question. It should be noted that the device for determining an answer to a financial question in the embodiment of the present application can be used to execute the method for determining an answer to a financial question provided in the embodiment of the present application. The device for determining an answer to a financial question provided in the embodiment of the present application is introduced below.

[0102] According to an embodiment of the present application, a device for implementing the above-mentioned method for determining the answer to the financial question is also provided. Figure 6 is a schematic diagram of an optional device for determining an answer to a financial question according to an embodiment of the present invention, such as Figure 6 As shown, the determination device may include: a processing unit 60, a query unit 61, a first determination unit 62, and a second determination unit 63.

[0103] The processing unit 60 is used to process the target financial question to be answered, obtain target word vector data of the target financial question, input the target word vector data into a preset classification model, and output the target question type of the target financial question;

[0104] A query unit 61 is used to query a set of candidate questions corresponding to a preset question type that is the same as the target question type from a preset corpus, wherein the preset corpus includes: a set of historical questions corresponding to a plurality of preset question types, each historical question in the set of historical questions corresponding to a set of historical answers;

[0105] A first determination unit 62 is used to determine the problem similarity between the target financial problem and each candidate problem in the candidate problem set by using a preset similarity model, and determine the candidate problem indicated by the maximum problem similarity as the target candidate problem;

[0106] The second determination unit 63 is used to determine a target answer corresponding to the target financial question from a set of candidate answers corresponding to the target candidate question by using a preset answer selection model.

[0107] The device for determining answers to financial questions provided in the embodiments of the present application determines the target question type of the target financial question to be answered through a preset classification model, and then calculates the question similarity between the target financial question and questions of the same type through a preset similarity model. Thereafter, through a preset answer selection model, the target answer corresponding to the target financial question is determined from a set of candidate answers corresponding to the target candidate question indicated by the maximum question similarity, thereby effectively improving the accuracy and efficiency of automatic answers to financial questions, achieving the technical effect of quickly and accurately answering customers' questions, and thereby solving the technical problem of low accuracy in answering financial questions in related technologies.

[0108] Optionally, the processing unit includes: a first processing module, used to obtain a financial stop word list and a preset character table of a financial institution, and based on the financial stop word list and the preset character table, filter the target financial issue to obtain preset text data, wherein the financial stop word list includes: multiple preset stop words, which are pre-set words that need to be filtered; a second processing module, used to perform word segmentation processing on the preset text data to obtain a word sequence; a third processing module, used to perform vectorization processing on the word sequence to obtain target word vector data.

[0109] Optionally, the determination device also includes: a first construction module, used to construct a densely connected network structure and a classification structure before inputting the target word vector data into a preset classification model and outputting the target problem type of the target financial problem, wherein the densely connected network structure includes at least: a convolution layer, a pooling layer, and a fully connected layer; the classification structure is composed of N convolution layers and N sampling layers alternating, and N is a positive integer greater than or equal to 2; a second construction module, used to construct an initial classification model based on the densely connected network structure and the classification structure; a first training module, used to collect first historical data, and use the first historical data to train the initial classification model to obtain a preset classification model, wherein the first historical data includes at least: multiple historical problems and a problem type label marked for each historical problem.

[0110] Optionally, the processing unit also includes: a first extraction module, used to use a densely connected network structure to perform feature extraction on the target word vector data to obtain a feature vector; and a fourth processing module, used to use a classification structure to process the feature vector to obtain a target question type.

[0111] Optionally, the determination device also includes: a first collection module, used to collect historical questions over a historical time period and a set of historical answers to historical questions before querying a set of candidate questions corresponding to a preset question type that is the same as the target question type from a preset corpus; a first input module, used to input each historical question into a preset classification model and output a preset question type corresponding to the historical question; a first adding module, used to add all historical questions indicated by the same preset question type to the same set of historical questions; and construct a preset corpus based on the set of historical questions corresponding to all preset question types and the set of historical answers corresponding to each historical question in the set of historical questions.

[0112] Optionally, the determination device also includes: a third construction module, which is used to construct a pre-trained structure and a densely connected bidirectional gated loop structure before using a preset similarity model to determine the problem similarity between the target financial problem and each candidate problem in the candidate problem set, wherein the pre-trained structure is used to perform feature processing on the word sequence of the input problem to obtain a context feature representation of the problem; the context feature representation vector includes at least: a word vector, a segment vector and a position vector; the densely connected bidirectional gated loop structure is used to process the context feature representation of the two input problems to obtain the problem similarity between the two problems; a fourth construction module, which is used to construct an initial similarity model based on the pre-trained structure and the densely connected bidirectional gated loop structure; a second acquisition module, which is used to collect second historical data and use the second historical data to train the initial similarity model to obtain a preset similarity model, wherein the second historical data includes at least: multiple historical problem pairs and similarity labels annotated for each group of historical problem pairs.

[0113] Optionally, the determination device also includes: a first labeling module, which is used to use a preset answer selection model to label each historical answer in the historical answer set corresponding to each historical question in the preset corpus before determining the target answer corresponding to the target financial question from the candidate answer set corresponding to the target candidate question, so as to obtain a score label for the historical answer; a fifth construction module, which is used to construct third historical data based on all historical answers labeled with score labels and historical questions corresponding to the historical answers; a second training module, which is used to use the third historical data to train the initial answer selection model, and obtain the preset answer selection model after the initial answer selection model is trained, wherein the initial answer selection model at least includes: an embedding layer, a dynamic attention layer, a multi-strategy interaction layer, a bidirectional gated loop layer, and an output layer; the embedding layer is used to The word sequence of the input question is feature processed to obtain the first vector representation of the question, and the word sequence of the input answer is feature processed to obtain the second vector representation of the answer; the dynamic attention layer is used to weight the first vector representation and the second vector representation to obtain the first weighted vector representation and the second weighted vector representation; the multi-strategy interaction layer is used to match the first weighted vector representation and the second weighted vector representation to obtain the semantic relationship between the first weighted vector representation and the second weighted vector representation; the bidirectional gated recurrent layer is used to process the first weighted vector representation and the second weighted vector representation based on the semantic relationship to obtain the first context vector representation and the second context vector representation; the output layer is used to process the first context vector representation and the second context vector representation to obtain the score between the question and the answer.

[0114] Optionally, the second determination unit includes: a first determination module, used to use a preset answer selection model to determine the score between the target financial question and each candidate answer in the candidate answer set; and a second determination module, used to determine the candidate answer corresponding to the maximum score as the target answer corresponding to the target financial question.

[0115] The above-mentioned determination device may also include a processor and a memory. The above-mentioned processing unit 60, query unit 61, first determination unit 62, second determination unit 63, etc. are all stored in the memory as program units, and the processor executes the above-mentioned program units stored in the memory to realize corresponding functions.

[0116] The processor includes a kernel, which retrieves the corresponding program unit from the memory. One or more kernels may be provided, and a preset answer selection model is adopted by adjusting kernel parameters to determine the target answer corresponding to the target financial question from the candidate answer set corresponding to the target candidate question.

[0117] The above-mentioned memory may include non-permanent memory in a computer-readable medium, random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash memory (fl ash RAM), and the memory includes at least one storage chip.

[0118] It should be noted that the above processing unit 60, query unit 61, first determination unit 62, and second determination unit 63 correspond to steps S201 to S204 in Embodiment 1, and the above units and corresponding steps implement the same instances and application scenarios, but are not limited to the contents disclosed in Embodiment 1. It should be noted that the above units may 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 units may also be part of a device and run in the computer terminal 10 provided in Embodiment 1.

[0119] Example 3

[0120] An embodiment of the present application may provide an electronic device, Figure 7 is a structural block diagram of an electronic device according to an embodiment of the present invention. Figure 7 As shown, the electronic device may include: one or more ( Figure 7 (only one is shown) processor 702, memory 704, storage controller, and peripheral interface, wherein the peripheral interface is connected to the radio frequency module, audio module and display.

[0121] Among them, the memory can be used to store software programs and modules, such as the program instructions / modules corresponding to the method and device for determining the answer to the financial question in the embodiment of the present application. 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 for determining the answer to the financial question. 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.

[0122] The processor can call the information and application programs stored in the memory through the transmission device to perform the following steps: process the target financial question to be answered to obtain the target word vector data of the target financial question, input the target word vector data into the preset classification model, and output the target question type of the target financial question; query the candidate question set corresponding to the preset question type that is the same as the target question type from the preset corpus, wherein the preset corpus includes: a historical question set corresponding to multiple preset question types, and each historical question in the historical question set corresponds to a historical answer set; use a preset similarity model to determine the question similarity between the target financial question and each candidate question in the candidate question set, and determine the candidate question indicated by the maximum question similarity as the target candidate question; use a preset answer selection model to determine the target answer corresponding to the target financial question from the candidate answer set corresponding to the target candidate question.

[0123] The processor can also call the information and application programs stored in the memory through the transmission device to perform the following steps: obtain the financial stop word list and preset character table of the financial institution, and filter the target financial problem based on the financial stop word list and the preset character table to obtain preset text data, wherein the financial stop word list includes: multiple preset stop words, which are pre-set words that need to be filtered; perform word segmentation on the preset text data to obtain a word sequence; and perform vectorization on the word sequence to obtain target word vector data.

[0124] The processor can also call the information and application programs stored in the memory through the transmission device to perform the following steps: construct a densely connected network structure and a classification structure, wherein the densely connected network structure includes at least: a convolutional layer, a pooling layer, and a fully connected layer; the classification structure is composed of N convolutional layers and N sampling layers alternating, and N is a positive integer greater than or equal to 2; based on the densely connected network structure and the classification structure, an initial classification model is constructed; the first historical data is collected, and the initial classification model is trained using the first historical data to obtain a preset classification model, wherein the first historical data includes at least: multiple historical questions and a question type label marked for each historical question.

[0125] The processor can also call the information and application programs stored in the memory through the transmission device to perform the following steps: use a densely connected network structure to extract features from the target word vector data to obtain a feature vector; use a classification structure to process the feature vector to obtain a target question type.

[0126] The processor can also call the information and application programs stored in the memory through the transmission device to perform the following steps: collect historical questions and a set of historical answers to historical questions over a historical time period; input each historical question into a preset classification model and output a preset question type corresponding to the historical question; add all historical questions indicated by the same preset question type to the same historical question set; and construct a preset corpus based on the set of historical questions corresponding to all preset question types and the set of historical answers corresponding to each historical question in the historical question set.

[0127] The processor can also call the information and application programs stored in the memory through the transmission device to perform the following steps: construct a pre-training structure and a densely connected bidirectional gated loop structure, wherein the pre-training structure is used to perform feature processing on the word sequence of the input question to obtain the context feature representation of the question; the context feature representation vector includes at least: a word vector, a segment vector and a position vector; the densely connected bidirectional gated loop structure is used to process the context feature representation of the two input questions to obtain the question similarity between the two questions; based on the pre-training structure and the densely connected bidirectional gated loop structure, an initial similarity model is constructed; the second historical data is collected, and the initial similarity model is trained using the second historical data to obtain a preset similarity model, wherein the second historical data includes at least: multiple historical question pairs and similarity labels annotated for each group of historical question pairs.

[0128] The processor can also call the information and application programs stored in the memory through the transmission device to perform the following steps: mark each historical answer in the historical answer set corresponding to each historical question in the preset corpus to obtain a score label for the historical answer; construct third historical data based on all historical answers marked with score labels and historical questions corresponding to the historical answers; use the third historical data to train the initial answer selection model, and after the initial answer selection model is trained, obtain the preset answer selection model, wherein the initial answer selection model at least includes: an embedding layer, a dynamic attention layer, a multi-strategy interaction layer, a bidirectional gated loop layer, and an output layer; the embedding layer is used to perform feature processing on the word sequence of the input question to obtain the first vector of the question. The first weighted vector representation is used to represent the question and the second weighted vector representation is used to perform feature processing on the word sequence of the input answer to obtain a second vector representation of the answer; the dynamic attention layer is used to weight the first vector representation and the second vector representation to obtain a first weighted vector representation and a second weighted vector representation; the multi-strategy interaction layer is used to match the first weighted vector representation and the second weighted vector representation to obtain a semantic relationship between the first weighted vector representation and the second weighted vector representation; the bidirectional gated recurrent layer is used to process the first weighted vector representation and the second weighted vector representation based on the semantic relationship to obtain a first context vector representation and a second context vector representation; the output layer is used to process the first context vector representation and the second context vector representation to obtain a score between the question and the answer.

[0129] The processor can also call the information and application programs stored in the memory through the transmission device to perform the following steps: using a preset answer selection model to determine the score between the target financial question and each candidate answer in the candidate answer set; determining the candidate answer corresponding to the maximum score as the target answer corresponding to the target financial question.

[0130] By adopting the embodiment of the present application, a solution for determining the answer to a financial question is provided. By presetting a classification model, the target question type of the target financial question to be answered is determined, and then the problem similarity between the target financial question and the same type of question is calculated by a preset similarity model, and then the target answer corresponding to the target financial question is determined from the candidate answer set corresponding to the target candidate question indicated by the maximum problem similarity by a preset answer selection model, which effectively improves the accuracy and efficiency of automatic answering of financial questions, achieves the technical effect of quickly and accurately answering the customer's questions, and thus solves the technical problem of low accuracy in answering financial questions in related technologies.

[0131] It can be understood by those skilled in the art that Figure 7The structure shown is for illustration only, and the electronic device may also be a terminal device such as a smart phone, a tablet computer, a PDA, and a Mobile Internet Device (MID). Figure 7 The structure of the electronic device is not limited. Figure 7 More or fewer components (such as network interfaces, display devices, etc.) shown in, or having Figure 7 Different configurations are shown.

[0132] 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.

[0133] Example 4

[0134] 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 determining the answer to the financial question provided in the first embodiment.

[0135] 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.

[0136] The present application also provides a computer program product, which, when executed on a data processing device, is suitable for executing the program steps of the method for determining answers to financial questions.

[0137] The serial numbers of the above-mentioned embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0138] 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.

[0139] 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.

[0140] 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.

[0141] 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.

[0142] 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 is essentially 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, server or 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 CD-ROM and other media that can store program codes.

[0143] 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 determining an answer to a financial question, characterized in that: include: Processing the target financial question to be answered to obtain target word vector data of the target financial question, and inputting the target word vector data into a preset classification model to output a target question type of the target financial question; Querying a set of candidate questions corresponding to a preset question type that is the same as the target question type from a preset corpus, wherein the preset corpus includes: a set of historical questions corresponding to a plurality of the preset question types, each historical question in the set of historical questions corresponding to a set of historical answers; Using a preset similarity model, determine the problem similarity between the target financial problem and each candidate problem in the candidate problem set, and determine the candidate problem indicated by the maximum problem similarity as the target candidate problem; A preset answer selection model is used to determine a target answer corresponding to the target financial question from a set of candidate answers corresponding to the target candidate question.

2. The determination method according to claim 1, characterized in that: The step of processing the target financial question to be answered to obtain target word vector data of the target financial question includes: Obtaining a financial stop word list and a preset character table of a financial institution, and filtering the target financial question based on the financial stop word list and the preset character table to obtain preset text data, wherein the financial stop word list includes: a plurality of preset stop words, and the preset stop words are pre-set words that need to be filtered; Performing word segmentation processing on the preset text data to obtain a word sequence; The word sequence is vectorized to obtain the target word vector data.

3. The determination method according to claim 1, characterized in that: Before inputting the target word vector data into a preset classification model and outputting the target problem type of the target financial problem, the method further includes: Constructing a densely connected network structure and a classification structure, wherein the densely connected network structure at least includes: a convolution layer, a pooling layer, and a fully connected layer; the classification structure is composed of N convolution layers and N sampling layers alternately, where N is a positive integer greater than or equal to 2; Based on the densely connected network structure and the classification structure, construct an initial classification model; First historical data is collected, and the initial classification model is trained using the first historical data to obtain the preset classification model, wherein the first historical data at least includes: a plurality of historical questions and a question type label marked for each of the historical questions.

4. The determination method according to claim 3, characterized in that: The step of inputting the target word vector data into a preset classification model to output the target problem type of the target financial problem comprises: Using the densely connected network structure to extract features from the target word vector data to obtain a feature vector; The characteristic vector is processed by using the classification structure to obtain the target question type.

5. The determination method according to claim 1, characterized in that: Before searching the preset corpus for a set of candidate questions corresponding to the preset question type that is the same as the target question type, the method further includes: Collect the historical questions and historical answers to the historical questions over a historical period of time; Input each of the historical questions into the preset classification model, and output the preset question type corresponding to the historical question; Adding all the historical questions indicated by the same preset question type to the same historical question set; The preset corpus is constructed based on the historical question set corresponding to all the preset question types and the historical answer set corresponding to each of the historical questions in the historical question set.

6. The determination method according to claim 1, characterized in that: Before using a preset similarity model to determine the problem similarity between the target financial problem and each candidate problem in the candidate problem set, the method further includes: Constructing a pre-training structure and a densely connected bidirectional gated loop structure, wherein the pre-training structure is used to perform feature processing on a word sequence of an input question to obtain a context feature representation of the question; the context feature representation vector includes at least: a word vector, a segment vector, and a position vector; the densely connected bidirectional gated loop structure is used to process the context feature representations of two input questions to obtain the question similarity between the two questions; Based on the pre-trained structure and the densely connected bidirectional gated loop structure, construct an initial similarity model; The second historical data is collected, and the initial similarity model is trained using the second historical data to obtain the preset similarity model, wherein the second historical data at least includes: a plurality of historical question pairs and a similarity label annotated for each group of the historical question pairs.

7. The determination method according to claim 1, characterized in that: Before determining the target answer corresponding to the target financial question from the set of candidate answers corresponding to the target candidate question using the preset answer selection model, the method further includes: Marking each historical answer in the historical answer set corresponding to each historical question in the preset corpus to obtain a score label of the historical answer; constructing third historical data based on all the historical answers marked with the score labels and the historical questions corresponding to the historical answers; The third historical data is used to train an initial answer selection model, and after the initial answer selection model is trained, the preset answer selection model is obtained, wherein the initial answer selection model at least includes: an embedding layer, a dynamic attention layer, a multi-strategy interaction layer, a bidirectional gated loop layer, and an output layer; the embedding layer is used to perform feature processing on the word sequence of the input question to obtain a first vector representation of the question, and to perform feature processing on the word sequence of the input answer to obtain a second vector representation of the answer; the dynamic attention layer is used to weight the first vector representation and the second vector representation to obtain a first weighted vector representation and a second weighted vector representation; the multi-strategy interaction layer is used to match the first weighted vector representation and the second weighted vector representation to obtain a semantic relationship between the first weighted vector representation and the second weighted vector representation; the bidirectional gated loop layer is used to process the first weighted vector representation and the second weighted vector representation based on the semantic relationship to obtain a first context vector representation and a second context vector representation; the output layer is used to process the first context vector representation and the second context vector representation to obtain a score between the question and the answer.

8. The determination method according to claim 1, characterized in that: The step of using a preset answer selection model to determine a target answer corresponding to the target financial question from a set of candidate answers corresponding to the target candidate question includes: Using a preset answer selection model, determining a score between the target financial question and each candidate answer in the candidate answer set; The candidate answer corresponding to the maximum score is determined as the target answer corresponding to the target financial question.

9. A device for determining answers to financial questions, characterized in that: include: A processing unit, configured to process the target financial question to be answered, obtain target word vector data of the target financial question, input the target word vector data into a preset classification model, and output a target question type of the target financial question; A query unit, configured to query a set of candidate questions corresponding to a preset question type that is the same as the target question type from a preset corpus, wherein the preset corpus includes: a set of historical questions corresponding to a plurality of the preset question types, each historical question in the set of historical questions corresponding to a set of historical answers; A first determination unit is configured to determine the problem similarity between the target financial problem and each candidate problem in the candidate problem set by using a preset similarity model, and determine the candidate problem indicated by the maximum problem similarity as the target candidate problem; The second determination unit is used to determine a target answer corresponding to the target financial question from a set of candidate answers corresponding to the target candidate question by using a preset answer selection model.

10. A computer program product, characterized in that It comprises a non-volatile computer-readable storage medium, wherein the non-volatile computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for determining an answer to a financial question as claimed in any one of claims 1 to 8 is implemented.

11. An electronic device, characterized in that: It includes one or more processors and a memory, wherein the memory is used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method for determining the answer to a financial question as described in any one of claims 1 to 8.

Images