Automatic extraction and reply method and system for financial problems based on convolutional neural network

Abstract

The application provides a financial problem automatic extraction reply method based on a convolutional neural network, characterized in that the financial problem automatic extraction reply method based on the convolutional neural network comprises the following steps: retrieving a financial problem containing a preset keyword from a preset channel; converting the financial problem into a question code according to a first preset algorithm; inputting the question code into a trained machine learning model to obtain answer information; and uploading and filling in the answer information through the preset channel. In addition, the application also provides a system and a computer device thereof. The technical scheme of the application effectively solves the problem that the existing network financial problem question and answer cannot be intelligently replied or requires human labor.

Description

Technical Field

[0001] This invention relates to the field of financial technology, and in particular to a method, system, and computer equipment for automatically extracting answers to financial questions based on convolutional neural networks. Background Technology

[0002] Financial services are closely related to our lives, and we inevitably encounter many financial-related questions in our daily lives. When faced with these problems, users usually choose to search for answers on platforms like Baidu and Zhihu to solve their problems themselves. However, the quality of the answers retrieved through these searches is not guaranteed and may mislead consumers. Banks typically employ staff to manually answer questions from the internet. But this method is time-consuming and labor-intensive, and cannot be widely implemented. Therefore, designing an automated question extraction and answering system applicable to the financial industry is essential. Summary of the Invention

[0003] This invention provides a method and system for automatically extracting and answering financial questions based on convolutional neural networks, enabling efficient automatic extraction and answering of financial questions.

[0004] In a first aspect, embodiments of the present invention provide a method for automatically extracting answers to financial questions based on convolutional neural networks, the method comprising:

[0005] Financial questions containing preset keywords are retrieved from preset channels;

[0006] The financial problem is converted into a problem code according to the first preset algorithm;

[0007] The question code is input into a trained machine learning model to obtain the answer information; and

[0008] The answer information is uploaded and filled in through the preset channel.

[0009] Specifically, the first preset algorithm includes a gate mechanism, a residual structure mechanism, a dilated convolution mechanism, and an attention mechanism. Transcoding the financial problem into a problem code according to the first preset algorithm includes:

[0010] The financial issues are segmented to obtain issue word blocks;

[0011] A vector model is obtained based on the question word block using the gate mechanism and the residual structure mechanism, wherein the vector model includes a vision;

[0012] The field of view is increased by utilizing the dilated convolution mechanism; and

[0013] The attention mechanism is used to transcode the vector model to form the problem code.

[0014] Specifically, inputting the question code into the trained machine learning model to obtain answer information includes:

[0015] The question is encoded and input into a trained machine learning model to obtain a document model, wherein the document model includes several answer word blocks, and each answer word block includes a one-to-one corresponding document vector and a positional encoding.

[0016] The document vector is transcoded using an attention mechanism to generate a document encoding.

[0017] Using a pointer mechanism, the document encoding is extracted based on the positional encoding to form a new document encoding;

[0018] The answer information is formed by concatenating and transcoding all the document encodings.

[0019] Specifically, concatenating and transcoding all the document encodings to form the answer information also includes:

[0020] All the document encodings are concatenated and transcoded to form at least one answer model, wherein the answer model includes several second word blocks and scoring coefficients, and the second word blocks include one-to-one document vectors;

[0021] Determine whether the document vector contains duplicates;

[0022] When the document vectors are repeated, any one of the answer models is selected as the answer information based on the repeated document vectors;

[0023] When the document vectors are not duplicated, the answer model corresponding to the largest rating coefficient is selected as the answer information.

[0024] Specifically, the WebQA dataset includes question substrings and material substrings, and the bank database includes labeled data. Before inputting the question encoding into the trained machine learning model to obtain answer information, the automatic answer extraction method for financial questions based on convolutional neural networks further includes:

[0025] Obtain the question substring as input data from the WebQA dataset;

[0026] The labeled data and the material substring are obtained from the bank database and the WebQA dataset as pre-answer data and used as output data;

[0027] The question substring, the labeled data, and the material substring are respectively divided into a training set and a test set according to a preset ratio;

[0028] Train the machine learning model based on the training set; and

[0029] The machine learning model is tested according to the test set until the trained machine learning model is obtained.

[0030] Specifically, the training set includes data segments, and before training the machine learning model based on the training set, the method for automatically extracting answers to financial questions based on convolutional neural networks further includes:

[0031] The data segment is randomly cut to obtain segments;

[0032] Repeatedly concatenate the data segment and the segmented data to obtain a new data segment; and

[0033] Add the new data segment to the training set.

[0034] Specifically, before training the machine learning model based on the training set, the method for automatically extracting answers to financial questions based on convolutional neural networks further includes:

[0035] The machine learning model is trained using the Adam optimizer and a warm-start strategy to optimize the machine learning model.

[0036] In a second aspect, embodiments of the present invention provide a computer device, the computer device comprising:

[0037] Memory, used to store program instructions; and

[0038] A processor for executing the program instructions to implement the automatic response extraction method for financial questions based on convolutional neural networks as described in any of the above-described methods.

[0039] Thirdly, embodiments of the present invention provide an automatic response system, the automatic response system comprising:

[0040] The search module is used to retrieve financial questions containing preset keywords from preset channels;

[0041] The transcoding module is used to transcode the financial problem into a problem code according to a first preset algorithm;

[0042] The prediction module is used to input the encoded question into a trained machine learning model to obtain answer information; and

[0043] The fill-in module is used to upload and fill in the answer information through the preset channel.

[0044] Specifically, the transcoding module includes:

[0045] The word segmentation module is used to segment the financial issues into word blocks to obtain the issue word blocks;

[0046] The computation module is used to obtain a vector model based on the problem word block using a gate mechanism and a residual structure mechanism, wherein the vector model includes a vision;

[0047] A dilation module is used to increase the field of view using dilated convolution; and

[0048] The sub-transcoding module is used to transcode the vector model into the problem code using an attention mechanism.

[0049] The aforementioned method and system for automatically extracting and answering financial questions based on convolutional neural networks involves: retrieving financial questions containing preset keywords from preset channels; transcoding the financial questions using a first preset algorithm to form question codes; inputting the question codes into a trained machine learning model to obtain answer information; and uploading and filling in the answer information through the preset channels. By automatically retrieving financial questions related to companies from online channels, transcoding them, generating answers through a machine learning model, and automatically uploading and filling them in on the website or page where the question was posted, this method effectively improves the efficiency of online financial consultation and inquiry, saves human resources, and increases overall efficiency. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0051] Figure 1 A flowchart of a method for automatically extracting answers to financial questions based on convolutional neural networks, provided in an embodiment of this application.

[0052] Figure 2 for Figure 1 The first sub-flowchart of the provided method for automatically extracting answers to financial questions based on convolutional neural networks.

[0053] Figure 3 for Figure 1 The second sub-flowchart of the provided method for automatically extracting answers to financial questions based on convolutional neural networks.

[0054] Figure 4 for Figure 1 The third sub-flowchart of the provided method for automatically extracting answers to financial questions based on convolutional neural networks.

[0055] Figure 5 for Figure 1 The fourth sub-flowchart of the provided method for automatically extracting answers to financial questions based on convolutional neural networks.

[0056] Figure 6 for Figure 1 The fifth sub-flowchart of the provided method for automatically extracting answers to financial questions based on convolutional neural networks.

[0057] Figure 7 This is a schematic diagram of the internal structure of the automatic response system provided in the embodiments of this application.

[0058] Figure 8 This is a schematic diagram of the internal structure of the transcoding module provided in an embodiment of this application.

[0059] Figure 9 This is a schematic diagram of the internal structure of a computer device provided in an embodiment of this application.

[0060] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0062] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar planned objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data are interchangeable where appropriate; in other words, the described embodiments are implemented according to a sequence other than that illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, may also include other content; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0063] It should be noted that the descriptions involving "first," "second," etc., in this invention are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined with "first" or "second" may explicitly or implicitly include one or more of that feature. Furthermore, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0064] Please refer to the following: Figure 1 This is a flowchart of a method for automatically extracting answers to financial questions based on convolutional neural networks, provided in an embodiment of the present invention. The method includes the following steps.

[0065] Step S101: Retrieve financial questions containing preset keywords from preset channels.

[0066] Specifically, the system searches for relevant questions on the internet and then uses web crawlers to extract these questions and store them locally. Preset keywords include, but are not limited to, "bank," "loan," "finance," "credit limit," and "term." Preset channels include search engines and social media platforms such as Baidu Knows, Zhihu, and Xiaohongshu.

[0067] Step S102: The financial problem is transcoded into a problem code according to the first preset algorithm.

[0068] Specifically, the process of transcoding the financial problem into a problem code according to the first preset algorithm will be described in detail below.

[0069] Step S103: Input the question code into the trained machine learning model to obtain the answer information.

[0070] Specifically, the process of transcoding the financial problem into a problem code according to the first preset algorithm will be described in detail below.

[0071] Step S104: Upload and fill in the answer information through the preset channel.

[0072] Specifically, the obtained answer information will be filled into the question box by a specific account using the original channel through which the financial questions were obtained.

[0073] In summary, this method involves retrieving relevant keywords from popular search engines or social media platforms to obtain financial questions; transcoding these questions into question codes and inputting them into a trained machine learning model to retrieve answers; and then using a specific account to answer the question through the original channel. This approach is beneficial for large financial and banking institutions serving a large number of users, improving the quality of answers and service levels while reducing manpower. Answering questions related to and unrelated to the company's business on well-known internet channels not only serves the public but also promotes the company's products.

[0074] Please refer to the following: Figure 2 This is the first sub-flowchart of the automatic question extraction and response method for financial problems based on convolutional neural networks provided in the embodiments of the present invention.

[0075] The first preset algorithm includes a gate mechanism, a residual structure mechanism, a dilated convolution mechanism, and an attention mechanism. Step S102 includes the following steps.

[0076] Step S202: Segment the financial issue to obtain the issue word block.

[0077] Step S204: Obtain the vector model based on the problem word block using the gate mechanism and the residual structure mechanism.

[0078] Specifically, the gate mechanism: Assuming the vector sequence to be processed is X=[x1,x2,…,xn], a gate is added to the ordinary one-dimensional convolution. The specific formula is as follows:

[0079]

[0080] Where Y is the output value and X is the input value.

[0081] The two 1D convolutions here are identical in form, but they do not share weights. The purpose of the sigmoid gate is to add a "valve" to each output to control the flow of information.

[0082] Residual Structure: To avoid gradient insignificance, a residual structure is introduced to convey more information. The specific formula is as follows:

[0083]

[0084] Where Y is the output value and X is the input value.

[0085] The vector model includes a field of view, which determines the interval at which information can be acquired and the accuracy of that information.

[0086] In summary, gating mechanisms are introduced to limit the amount of input data to improve efficiency and accuracy, while residual structures are introduced to enhance the gradient of information.

[0087] Step S206: Increase the field of view using the dilated convolution mechanism.

[0088] Specifically, when using CNNs to encode text, the field of view of the vector model depends on the size of the CNN's convolutional kernels. To enable the vector model to capture farther distances without increasing the number of model parameters, dilated convolutions are used to provide a larger field of view with the same kernel size. To achieve more comprehensive coverage, the dilation of the dilated convolution increases geometrically with increments of 1, 2, 4, 8, etc.

[0089] Step S208: Use the attention mechanism to transcode the vector model to form the problem code.

[0090] Attention Mechanism: An attention mechanism is introduced to replace simple pooling to integrate sequence information. For example, the sequence of question vectors is encoded into a total question vector, and the sequence of answers is encoded into a total material vector. The specific formula is as follows:

[0091]

[0092]

[0093] Where a and w are trainable parameters; activation is the activation function, which is the tanh function.

[0094] In summary, unlike models built on BERT, which are relatively cumbersome, inefficient, and require a large amount of knowledge base data for information extraction, replacing LSTM, commonly used in NLP tasks, with a CNN model offers the advantage of using a lighter model to improve speed and achieve high efficiency.

[0095] Please refer to the following: Figure 3 This is the second sub-flowchart of the automatic question extraction and response method for financial problems based on convolutional neural networks provided in the embodiments of the present invention.

[0096] Specifically, step S103 includes the following steps.

[0097] Step S302: Input the problem code into the trained machine learning model to obtain the document model.

[0098] Specifically, the document model includes several answer word blocks, each containing a corresponding document vector and a positional encoding. Positional information is added to the document model during input, with the specific formula as follows:

[0099]

[0100]

[0101] Adding location information reduces redundant text and improves the accuracy of the output when selecting answers later.

[0102] Step S304: Use an attention mechanism to transcode the document vector to generate document encoding.

[0103] Step S306: Use a pointer mechanism to extract the document encoding based on the position encoding to form a new document encoding.

[0104] The document model sets up two outputs to represent the start and end positions of the answer, as shown in the following formula:

[0105]

[0106]

[0107] In addition, considering that some questions have no answers, a control item is added to determine whether an answer exists. The specific formula is as follows:

[0108]

[0109] If there is no matching answer, the output value will be zero.

[0110] Step S308: Concatenate and transcode all the document encodings to form the answer information.

[0111] In conclusion, adding positional encoding and pointer output to the document model input effectively improves the accuracy of the answers and the readability of the sentences.

[0112] Please refer to the following: Figure 4 This is the third sub-flowchart of the automatic question extraction and response method for financial problems based on convolutional neural networks provided in the embodiments of the present invention.

[0113] Specifically, step S308 includes the following steps.

[0114] Step S402: Encode and concatenate all the document codes to form at least one answer model.

[0115] Specifically, the answer model includes several second word blocks and scoring coefficients, where each second word block comprises a corresponding document vector. The model's output is passed through a sigmoid gate to obtain probability values ​​p(start) and p(end), with the goal of selecting the start and end positions in the document in the form of [0,1]. p(start) and p(end) are used as ranking metrics; in practice, when a question is input, the model will predict from multiple documents.

[0116] Step S404: Determine whether the document vector is duplicated. If the document vector is duplicated, proceed to step S406; if the document vector is not duplicated, proceed to step S408.

[0117] Step S406: Select any one of the answer models as the answer information based on the repeated document vectors.

[0118] Specifically, when duplicate document vectors exist, only the document vector with the highest score is taken as the answer information.

[0119] Step S408: Select the answer model corresponding to the largest scoring coefficient as the answer information.

[0120] Specifically, the answer model with the highest score coefficient among the different answers is selected as the answer information.

[0121] In summary, using repetition and a scoring system as the basis for answer selection, repetition can be considered key information. When no repetition occurs, the highest-scoring answer is used as the key information. All key information is then assembled into a sentence as the answer output. This effectively improves the accuracy of the answer and the readability of the sentence.

[0122] Please refer to the following: Figure 5 This is the fourth sub-flowchart of the automatic question extraction and response method for financial questions based on convolutional neural networks provided in this embodiment of the invention. The WebQA dataset includes question substrings and material substrings, and the bank database includes labeled data. Before inputting the question encoding into the trained machine learning model to obtain answer information, the automatic question extraction and response method for financial questions based on convolutional neural networks further includes the following steps.

[0123] Step S502: Obtain the question substring from the WebQA dataset as input data.

[0124] Step S504: Obtain the labeled data and the material substring from the bank database and the WebQA dataset as pre-answer data as output data.

[0125] Step S506: Divide the question substring, the labeled data, and the material substring into a training set and a test set according to a preset ratio.

[0126] Step S508: Train the machine learning model based on the training set.

[0127] Step S510: Test the machine learning model according to the test set until the trained machine learning model is obtained.

[0128] In summary, the training model uses data annotated in the bank's internal knowledge base and Baidu's open-source WebQA dataset. The bank's internal annotated dataset contains approximately 8,000 entries, while the WebQA dataset contains over 40,000 entries. The data format is "question + answer + materials." These two databases are the primary source of data, with other databases potentially added as supplementary sources.

[0129] Please refer to the following: Figure 6 This is the fifth sub-flowchart of the automatic question extraction and response method for financial questions based on convolutional neural networks provided in this embodiment of the invention. The training set includes a data segment. Before training the machine learning model based on the training set, the automatic question extraction and response method for financial questions based on convolutional neural networks further includes the following steps.

[0130] Step S602: Randomly trim the data segment to obtain segments.

[0131] Step S604: Repeatedly splice the data segment and the segment to obtain a new data segment.

[0132] Step S606: Add the new data segment to the training set.

[0133] In summary, due to the small amount of data in the bank database, data augmentation was used to increase the training samples. The same material was repeatedly spliced ​​and randomly cut to obtain new material (the number and position of the answers also changed accordingly).

[0134] Specifically, before training the machine learning model based on the training set, the machine learning model is trained using the Adam optimizer and a warm-start strategy to optimize the machine learning model.

[0135] Specifically, to make the trained model more stable, the Exponential Moving Average (EMA) method is used. This method can improve the performance of the solution at almost zero additional cost. The specific formula is as follows:

[0136]

[0137] in It is a "shadow variable" that is maintained. It is the attenuation factor, and its value can be 0.999.

[0138] Please refer to the following: Figure 9 This is a schematic diagram of the internal structure of a computer device provided in an embodiment of this application. The computer device 10 includes a memory 11 and a processor 12. The memory 11 is used to store program instructions, and the processor 12 is used to execute the program instructions to implement the above-described method for automatically extracting and answering financial questions based on convolutional neural networks.

[0139] In some embodiments, the processor 12 may be a central processing unit (CPU), controller, microcontroller, microprocessor or other data processing chip, used to run program instructions stored in the memory 11.

[0140] The memory 11 includes at least one type of readable storage medium, including flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 11 may be an internal storage unit of a computer device, such as a hard disk. In other embodiments, the memory 11 may be an external storage device of a computer device, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., provided on the computer device. Furthermore, the memory 11 may include both internal and external storage units of the computer device. The memory 11 can be used not only to store application software and various types of data installed on the computer device, such as code for an automatic answer extraction method for financial questions based on convolutional neural networks, but also to temporarily store data that has been output or will be output.

[0141] Please refer to the following: Figure 7 This is a schematic diagram of the internal structure of the automatic response system provided in this application embodiment. The automatic response system 1000 includes a retrieval module 100, a transcoding module 200, a prediction module 300, and a filling module 400.

[0142] The retrieval module 100 is used to retrieve financial questions containing preset keywords from preset channels.

[0143] Specifically, the system searches for relevant questions on the internet and then uses web crawlers to extract these questions and store them locally. Preset keywords include, but are not limited to, "bank," "loan," "finance," "credit limit," and "term." Preset channels include search engines and social media platforms such as Baidu Knows, Zhihu, and Xiaohongshu.

[0144] The transcoding module 200 is used to transcode the financial problem into a problem code according to a first preset algorithm.

[0145] The prediction module 300 is used to input the question encoding into a trained machine learning model to obtain answer information.

[0146] The filling module 400 is used to upload and fill in the answer information through the preset channel.

[0147] Please refer to the following: Figure 8 This is a schematic diagram of the internal structure of the transcoding module provided in an embodiment of this application. The transcoding module 200 includes:

[0148] The word segmentation module 201 is used to segment the financial issue into word blocks.

[0149] The calculation module 202 is used to obtain a vector model based on the problem word block using the gate mechanism and the residual structure mechanism.

[0150] Specifically, the gate mechanism: Assuming the vector sequence to be processed is X=[x1,x2,…,xn], a gate is added to the ordinary one-dimensional convolution. The specific formula is as follows:

[0151]

[0152] Where Y is the output value and X is the input value.

[0153] The two 1D convolutions here are identical in form, but they do not share weights. The purpose of the sigmoid gate is to add a "valve" to each output to control the flow of information.

[0154] Residual Structure: To avoid gradient insignificance, a residual structure is introduced to convey more information. The specific formula is as follows:

[0155]

[0156] Where Y is the output value and X is the input value.

[0157] The vector model includes a field of view, which determines the interval at which information can be acquired and the accuracy of that information. In summary, a gating mechanism is introduced to limit the amount of input data to improve efficiency and accuracy, while a residual structure is introduced to enhance the gradient of information.

[0158] The dilation module 203 is used to increase the field of view by dilation convolution.

[0159] Specifically, the dilation module 203 uses a CNN to encode the text, and the field of view of the vector model depends on the size of the CNN's convolutional kernel. To enable the vector model to capture farther distances without increasing the number of model parameters, dilated convolutions are used to provide a larger field of view with the same kernel size. For more comprehensive coverage, the dilation of the dilated convolution increases geometrically at increments of 1, 2, 4, 8, etc.

[0160] Sub-transcoding module 204 is used to transcode the vector model to form the problem code using an attention mechanism.

[0161] Attention Mechanism: An attention mechanism is introduced to replace simple pooling to integrate sequence information. For example, the sequence of question vectors is encoded into a total question vector, and the sequence of answers is encoded into a total material vector. The specific formula is as follows:

[0162]

[0163]

[0164] Where a and w are trainable parameters; activation is the activation function, which is the tanh function.

[0165] In summary, unlike models built on BERT, which are relatively cumbersome, inefficient, and require a large amount of knowledge base data for information extraction, replacing LSTM, commonly used in NLP tasks, with a CNN model offers the advantage of using a lighter model to improve speed and achieve high efficiency.

[0166] In the above embodiments, financial questions containing preset keywords are retrieved from preset channels; the financial questions are transcoded into question codes according to a first preset algorithm; the question codes are input into a trained machine learning model to obtain answer information; and the answer information is uploaded and filled in through the preset channels. By automatically retrieving financial questions related to companies from online channels, transcoding them, generating answers through a machine learning model, and automatically uploading and filling them in on the website or page where the question was posted, the efficiency of online financial consultation and inquiry is effectively improved, saving human resources and increasing efficiency.

[0167] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

[0168] The above-listed embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the scope of the present invention. Therefore, any equivalent variations made in accordance with the claims of the present invention are still within the scope of the present invention.

Claims

1. A method for automatically extracting answers to financial questions based on convolutional neural networks, characterized in that, The method for automatically extracting answers to financial questions based on convolutional neural networks includes: Financial questions containing preset keywords are retrieved from preset channels; The financial problem is transcoded into a problem code according to a first preset algorithm, which includes a gate mechanism, a residual structure mechanism, a dilated convolution mechanism, and an attention mechanism. The question code is input into a trained machine learning model to obtain the answer information; and Upload and fill in the answer information through the preset channel; The process of transcoding the financial problem into a problem code according to the first preset algorithm includes: The financial issues are segmented to obtain issue word blocks; A vector model is obtained based on the question word block using the gate mechanism and the residual structure mechanism, wherein the vector model includes a vision; The field of view is increased by utilizing the dilated convolution mechanism; and The attention mechanism is used to transcode the vector model to form the problem code; The process of encoding the question and inputting it into a trained machine learning model to obtain the answer information includes: The question is encoded and input into a trained machine learning model to obtain a document model, wherein the document model includes several answer word blocks, and each answer word block includes a one-to-one corresponding document vector and a positional encoding. The document vector is transcoded using an attention mechanism to generate a document encoding. Using a pointer mechanism, the document encoding is extracted based on the positional encoding to form a new document encoding; The answer information is formed by concatenating and transcoding all the document encodings.

2. The method for automatically extracting answers to financial questions based on convolutional neural networks as described in claim 1, characterized in that, The process of encoding and transcoding all the aforementioned documents to form the answer information also includes: All the document encodings are concatenated and transcoded to form at least one answer model, wherein the answer model includes several second word blocks and scoring coefficients, and the second word blocks include one-to-one document vectors; Determine whether the document vector contains duplicates; When the document vectors are repeated, any one of the answer models is selected as the answer information based on the repeated document vectors; When the document vectors are not duplicated, the answer model corresponding to the largest rating coefficient is selected as the answer information.

3. The method for automatically extracting answers to financial questions based on convolutional neural networks as described in claim 1, characterized in that, The WebQA dataset includes question substrings and material substrings, and the bank database includes labeled data. Before inputting the question encoding into the trained machine learning model to obtain answer information, the automatic answer extraction method for financial questions based on convolutional neural networks further includes: Obtain the question substring as input data from the WebQA dataset; The labeled data and the material substring are obtained from the bank database and the WebQA dataset as pre-answer data and as output data, respectively. The question substring, the labeled data, and the material substring are respectively divided into a training set and a test set according to a preset ratio; Train the machine learning model based on the training set; and The machine learning model is tested according to the test set until the trained machine learning model is obtained.

4. The method for automatically extracting answers to financial questions based on convolutional neural networks as described in claim 3, characterized in that, The training set includes a data segment. Before training the machine learning model based on the training set, the method for automatically extracting answers to financial questions based on convolutional neural networks further includes: The data segment is randomly cut to obtain segments; Repeatedly concatenate the data segment and the segmented data to obtain a new data segment; and Add the new data segment to the training set.

5. The method for automatically extracting answers to financial questions based on convolutional neural networks as described in claim 3, characterized in that, Before training the machine learning model based on the training set, the method for automatically extracting answers to financial questions based on convolutional neural networks further includes: The machine learning model is trained using the Adam optimizer and a warm-start strategy to optimize the machine learning model.

6. A computer device, characterized in that, The computer device includes: Memory, used to store program instructions; and A processor for executing the program instructions to implement the automatic response extraction method for financial questions based on convolutional neural networks as described in any one of claims 1 to 5.

7. An automatic response system, characterized in that, The automated response system includes: The search module is used to retrieve financial questions containing preset keywords from preset channels; The transcoding module is used to transcode the financial problem into a problem code according to a first preset algorithm. The first preset algorithm includes a gate mechanism, a residual structure mechanism, a dilated convolution mechanism, and an attention mechanism. Transcoding the financial problem into a problem code according to the first preset algorithm includes: The financial issues are segmented to obtain issue word blocks; A vector model is obtained based on the question word block using the gate mechanism and the residual structure mechanism, wherein the vector model includes a vision; The field of view is increased by utilizing the dilated convolution mechanism; and The attention mechanism is used to transcode the vector model to form the problem code; The prediction module is used to input the question code into a trained machine learning model to obtain answer information; wherein, inputting the question code into the trained machine learning model to obtain answer information includes: The question is encoded and input into a trained machine learning model to obtain a document model, wherein the document model includes several answer word blocks, and each answer word block includes a one-to-one corresponding document vector and a positional encoding. The document vector is transcoded using an attention mechanism to generate a document encoding. Using a pointer mechanism, the document encoding is extracted based on the positional encoding to form a new document encoding; The answer information is formed by concatenating and transcoding all the document encodings. The fill-in module is used to upload and fill in the answer information through the preset channel.

8. The automatic response system as described in claim 7, characterized in that, The transcoding module includes: The word segmentation module is used to segment the financial issues into word blocks to obtain the issue word blocks; The computation module is used to obtain a vector model based on the problem word block using a gate mechanism and a residual structure mechanism, wherein the vector model includes a vision; A dilation module is used to increase the field of view using dilated convolution; and The sub-transcoding module is used to transcode the vector model into the problem code using an attention mechanism.

Images