Model construction method and device applied to bond data processing and electronic equipment

Abstract

The invention discloses a model construction method and device applied to bond data processing and electronic equipment, and the method comprises the steps: carrying out the recognition of target bond comment data based on a large model, obtaining a first data set, and enabling each piece of data in the first data set to comprise the bond comment data and a corresponding emotion label; and training the initial recognition model based on a bond dictionary in the bond field and the first data set until an output result of the initial recognition model meets a target condition, and determining the current initial recognition model as a target recognition model. According to the method, the model for performing emotion recognition on the bond comment data is constructed, and a large model is adopted for sample labeling, so that the cost of manual labeling is reduced, the accuracy of model recognition is improved, and an accurate data basis is provided for bond data processing.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a model building method, device and electronic equipment for bond data processing. Background Art

[0002] With the popularization of the Internet and social media, the impact of investor sentiment on the bond market has become increasingly significant. Discussions on financial forums and social media platforms contain a wealth of information about investors’ emotions and opinions, and analysis of this information may reveal non-traditional factors that affect bond price fluctuations.

[0003] Currently, investors’ opinions on bonds are mainly analyzed through machine learning models. However, due to the large amount of investor comment information data, a large number of machine learning samples need to be labeled, which increases the cost of manual labeling. If the number of samples is small, it will also reduce the accuracy of machine learning model recognition. Summary of the invention

[0004] In order to solve the above problems, this application provides the following technical solutions:

[0005] A model building method for bond data processing, the method comprising:

[0006] Identify the target bond review data based on the large model to obtain a first data set, where each data in the first data set includes bond review data and a corresponding sentiment label;

[0007] Based on the bond dictionary in the bond field and the first data set, the initial recognition model is trained until the output result of the initial recognition model meets the target condition, and the current initial recognition model is determined as the target recognition model;

[0008] The bond dictionary includes several words in the bond field.

[0009] Optionally, the method further comprises:

[0010] Collect initial commentary data on the bond sector;

[0011] Performing data preprocessing on the initial review data to determine a bond code corresponding to each review data;

[0012] Based on the review data labeled with each bond code, target bond review data is generated.

[0013] Optionally, the large model includes a first model and a second model, and the step of identifying the target bond review data based on the large model to obtain a first data set includes:

[0014] Identify each data in the target bond review data based on the first model to obtain a first sentiment label corresponding to the data;

[0015] Identify each data in the target bond review data based on the second model to obtain a second sentiment label corresponding to the data;

[0016] Based on the first emotion label and the second emotion label corresponding to each data, determining a target emotion label corresponding to each data;

[0017] Determine the data marked with the target emotion label as the first data set;

[0018] The first model and the second model have different model structures.

[0019] Optionally, the method further comprises:

[0020] Perform word segmentation processing on each bond review data to obtain the vocabulary of each bond review data;

[0021] The frequency of occurrence of each word is counted, and the words with a frequency higher than the target frequency threshold are combined into a bond dictionary.

[0022] Optionally, the training of the initial recognition model based on the bond dictionary in the bond field and the first data set until the output result of the initial recognition model meets the target condition, and determining the current initial recognition model as the target recognition model includes:

[0023] Performing unsupervised training on the initial recognition model based on a bond dictionary in the bond field to obtain a first initial model after unsupervised training;

[0024] Performing supervised training on the first initial model based on the first data set to obtain a supervised trained second initial model;

[0025] Verify the second initial model based on the test set, and if the deviation range between the result output by the second initial model and the annotated label of the test data in the test set meets the target condition, determine the current second initial model as the target recognition model;

[0026] If the deviation range does not meet the target condition, the model parameters of the second initial model are adjusted until the output result of the adjusted second initial model meets the target condition, and the currently adjusted second initial model is determined as the target recognition model.

[0027] Optionally, the method further comprises:

[0028] Based on the target recognition model, the bond review data to be recognized is recognized to obtain a sentiment label corresponding to the bond review data, wherein the sentiment label includes a positive sentiment label and a negative sentiment label;

[0029] Based on the sentiment label, the influencing factors of bond returns are analyzed to obtain a target influencing factor, and the target influencing factor is used to evaluate the return characteristics of the bond.

[0030] Optionally, the analyzing the influencing factors of bond returns based on the sentiment tags to obtain target influencing factors includes:

[0031] Based on the sentiment tags, determining the scores of various impact factors of the bonds corresponding to each bond review data;

[0032] Based on the scores of the impact factors, sorting the impact factors, and determining candidate impact factors based on the sorting results;

[0033] A correlation test between the impact factor and the bond yield is performed on each of the candidate impact factors, and a target impact factor is determined based on the test results.

[0034] Optionally, the model structure of the initial recognition model is a FinBERT model structure, and the FinBERT model structure can transfer knowledge on financial field corpus to downstream financial text application tasks.

[0035] A model building device for bond data processing, the device comprising:

[0036] An acquisition unit, configured to identify the target bond review data based on the large model to obtain a first data set, wherein each data in the first data set includes the bond review data and a corresponding sentiment label;

[0037] A training unit, configured to train an initial recognition model based on a bond dictionary in the bond field and the first data set until an output result of the initial recognition model meets a target condition, and determine the current initial recognition model as a target recognition model;

[0038] The bond dictionary includes several words in the bond field.

[0039] An electronic device, comprising:

[0040] Memory, used to store programs;

[0041] A processor is used to execute the program, and the program is specifically used to implement the model building method applied to bond data processing as described in any one of the above.

[0042] Compared with the prior art, the present application provides a model construction method, device and electronic device for bond data processing, the method comprising: identifying the target bond review data based on a large model to obtain a first data set, each data in the first data set includes bond review data and corresponding sentiment tags; based on the bond dictionary in the bond field and the first data set, the initial recognition model is trained until the output result of the initial recognition model meets the target condition, and the current initial recognition model is determined as the target recognition model. The present application constructs a model for sentiment recognition of bond review data, and uses a large model for sample annotation, which reduces the cost of manual annotation, improves the accuracy of model recognition, and provides an accurate data basis for bond data processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0044] Figure 1 A flow chart of a model building method for bond data processing provided in an embodiment of the present application;

[0045] Figure 2 A schematic diagram of the structure of a model building device for bond data processing provided in an embodiment of the present application. DETAILED DESCRIPTION

[0046] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0047] The terms "first" and "second" and the like in the specification and claims of this application and the above drawings are used to distinguish different objects, rather than to describe a specific order. 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 limited to the listed steps or units, but may include steps or units that are not listed.

[0048] In the embodiment of the present application, a model building method for bond data processing is provided. The method builds a target recognition model by deeply analyzing and mining the comment data of investors in relevant forums. The method can determine the commentator's evaluation sentiment on the relevant bonds based on the target recognition model, thereby assisting in obtaining the corresponding bond investment scenario and improving the accuracy of data processing. In the construction of the target recognition model, the annotation information is mainly automatically generated through a large model, which reduces the cost of manual annotation, can provide more data samples for the construction of the target model, and improves the accuracy and efficiency of model construction.

[0049] In order to facilitate the description of the relevant implementation methods of the embodiments of the present application, the relevant terms used in the present application are now explained.

[0050] Public opinion information: the public’s views, emotions and attitudes on a specific topic or event, usually expressed through channels such as social media, forums, and news comments.

[0051] Sentiment analysis: Use natural language processing (NLP) technology to identify and extract subjective information from text and determine its emotional tendency, such as positive, negative, or neutral.

[0052] FinBERT is a domain-specific BERT obtained by using multi-task learning on financial corpus. It can transfer the knowledge of financial domain corpus to downstream financial text-related application tasks.

[0053] IC test: Information Coefficient test, used to measure the correlation between bond selection factors and future bond yields.

[0054] Quantile portfolio testing: An investment strategy testing method that divides bonds into multiple groups (quantiles) based on the value of a certain factor, then constructs an investment portfolio and tests its performance.

[0055] Long-Short Combination: An investment strategy that simultaneously holds long (buy) and short (sell) positions in an effort to profit from price differences.

[0056] Sharpe ratio: measures the excess return generated by a portfolio for every unit of total risk it bears, and is a risk-adjusted return indicator.

[0057] Maximum Drawdown Rate: The maximum drop in portfolio value from peak to trough, used to describe investment risk.

[0058] Excess return: The portion of a portfolio's return that exceeds the benchmark or market average.

[0059] Turnover rate: The frequency with which bonds in an investment portfolio are bought and sold over a certain period of time.

[0060] See also Figure 1 , is a flow chart of a model building method for bond data processing provided in an embodiment of the present application, the method may include the following steps:

[0061] S101. Identify target bond review data based on a large model to obtain a first data set.

[0062] Each data in the first data set includes bond review data and corresponding sentiment labels. The target bond review data is obtained by preprocessing the bond review data collected on the Internet, which may include multiple review data. The corresponding sentiment labels may include positive sentiment labels, negative sentiment labels, and neutral sentiment labels.

[0063] In the embodiment of the present application, the target bond review data is not manually labeled with sentiment labels, but a large model is used to label the target bond review data with sentiment labels, wherein the large model is a model trained based on data samples in the bond field, which can identify the bond review data based on the user and determine the corresponding sentiment label. In this way, the target bond review data is identified by a large model (large language model) and the situation labels are automatically labeled, which solves the problems of low efficiency and poor accuracy of manual labeling.

[0064] In one implementation of the embodiment of the present application, the process of processing the initial comment data to obtain the target bond comment data includes:

[0065] Collect initial review data for the bond field; perform data preprocessing on the initial review data to determine the bond code corresponding to each review data; and generate target bond review data based on each review data marked with a bond code.

[0066] First, we can select relevant data in the bond field in the target time period as the initial comment data, such as the data of stockholder posts in various financial or bond forums from January 2018 to December 2023. Eliminate information that is not related to the expression of investor sentiment, such as news, research reports, Q&A with the secretary of the board, company announcements, etc. Match the post data with the bond code one by one, and use data preprocessing methods to remove specific format texts and merge the title and text content to improve data quality. Data preprocessing can include data cleaning, format unification, deduplication and other operations.

[0067] Furthermore, in order to improve the accuracy of the sentiment labeling of the target bond review data by the large model, at least two models can be selected for labeling in the embodiment of the present application, and the data with consistent labeling results are used as the subsequent data set. In one embodiment, the large model includes a first model and a second model, and the target bond review data is identified based on the large model to obtain a first data set, including:

[0068] Based on the first model, each data in the target bond review data is identified to obtain a first sentiment label corresponding to the data; based on the second model, each data in the target bond review data is identified to obtain a second sentiment label corresponding to the data; based on the first sentiment label and the second sentiment label corresponding to each data, a target sentiment label corresponding to each data is determined; the data marked with the target sentiment label is determined as a first data set; wherein the model structures of the first model and the second model are different.

[0069] For example, the first model can be a complex Transformer architecture that can capture the deep features and patterns of language and has strong text analysis capabilities. The large model is pre-trained based on a massive data set, and does not require labeled samples. These data sets contain rich emotional expressions, so they have good generalization capabilities, can be generalized to new samples, and can perform emotional recognition of bond rating information. In order to improve the quality and accuracy of sample annotations and avoid inaccurate scoring results using a single large language model, the second model, such as a large language model (FinGPT), can be used to perform sentiment classification annotations on randomly selected sample data to generate preliminary sentiment labels. 50,000 bond rating data are randomly selected from the post data set (ensuring that the number of bond ratings for each bond is not too large) as a pre-annotated training set. Then, the large language model is used to perform sentiment classification annotations on the above 50,000 data sets, and the sentiment classification is divided into three categories: "positive", "negative", and "neutral". Since the bond market investment strategy model focuses on bullish and bearish, the neutral label has little to do with the strategy, so positive and negative comments are selected. Finally, 30,509 pieces of data were screened out, of which 45.03% were labeled “negative” and 22.05% were labeled “positive”.

[0070] S102, based on the bond dictionary in the bond field and the first data set, train the initial recognition model until the output result of the initial recognition model meets the target condition, and determine the current initial recognition model as the target recognition model.

[0071] The model structure of the initial recognition model is the FinBERT model structure, which can transfer the knowledge of the financial field corpus to the application tasks of downstream financial texts. That is, an improved Bert model is used, namely albert_chinese_small in the ALBERT model series. This model is a small version of albert, which aims to reduce the number of model parameters while maintaining or improving the performance of the model. The albert_chinese_small model is pre-trained on a large amount of Chinese corpus and is specially designed and optimized for the Chinese language.

[0072] In order to enable the initial recognition model to be better applied in the field of bond data, it is first pre-trained using a bond dictionary, and then trained using the first data set annotated by the large model to obtain a target recognition model. Specifically, in one embodiment, the process of constructing a bond dictionary includes: performing word segmentation processing on each bond review data to obtain the vocabulary of each bond review data; counting the frequency of occurrence of each vocabulary, and forming a bond dictionary with vocabulary whose frequency of occurrence is higher than the target frequency threshold.

[0073] Among them, the bond review data can be the unscreened bond review data, and then each review data is segmented to obtain a number of words corresponding to each review data, and the frequency of occurrence of each word is counted. A target frequency threshold can be set, such as 75%, and words with a frequency higher than this frequency are regarded as high-frequency words, so as to construct a bond dictionary based on these high-frequency words, that is, to construct a dictionary in the field of bond selection.

[0074] Correspondingly, in one implementation of the embodiment of the present application, based on the bond dictionary in the bond field and the first data set, the initial recognition model is trained until the output result of the initial recognition model meets the target condition, and the current initial recognition model is determined as the target recognition model, including:

[0075] Performing unsupervised training on the initial recognition model based on a bond dictionary in the bond field to obtain a first initial model after unsupervised training;

[0076] Performing supervised training on the first initial model based on the first data set to obtain a supervised trained second initial model;

[0077] Verify the second initial model based on the test set, and if the deviation range between the result output by the second initial model and the annotated label of the test data in the test set meets the target condition, determine the current second initial model as the target recognition model;

[0078] If the deviation range does not meet the target condition, the model parameters of the second initial model are adjusted until the output result of the adjusted second initial model meets the target condition, and the currently adjusted second initial model is determined as the target recognition model.

[0079] The initial model architecture is pre-trained through the bond dictionary, so that the model can learn the prior knowledge in the bond field, such as financial concepts, stockholder sentiment, etc., thereby enhancing the learning effect of the model. Before using the first data set for supervised training, you can also randomly select the posting data provided by each forum in the target time period. After preprocessing these data, you can get several pieces of data for unsupervised learning to further enrich the model's knowledge of the financial market. Then supervised training is performed based on the first data set to build the FinBERT model. In the process of supervised training using the first data set, multiple iterative updates can be performed, and the current output of the model can be compared with the label annotated by the test set to determine the deviation between the two. If 7 out of 10 data have deviations, you need to continue to adjust the model parameters until the deviation is within the target condition range, and the current model is determined as the target recognition model.

[0080] For example, the first data set can be divided into training set and test set in a ratio of 6:4. The out-of-sample accuracy of the FinBERT model in the bond market sentiment recognition task is 75.63%. The trained FinBERT model is used to reclassify the sentiment of more than 100,000 bond reviews in the posting data set. According to statistics, under the annotation of the FinBERT model, in the bond market bond review posting data from 2022 to 2023, the "positive" label accounts for 29.32%, the "negative" label accounts for 48.02%, and the data with neutral labels are excluded.

[0081] After obtaining the target recognition model, the target recognition model can be used to identify the relevant bond review data to obtain the corresponding sentiment labels, so that the influence factors for providing auxiliary investment advice can be constructed based on the sentiment labels. Correspondingly, the relevant sentiment labels can be corrected based on the determined relationship between the influence factors and bond returns, and the corrected sentiment labels can be used in the optimization process of the target recognition model to obtain a more accurate target recognition model.

[0082] In one implementation of the embodiment of the present application, the method further includes:

[0083] Based on the target recognition model, the bond review data to be recognized is recognized to obtain a sentiment label corresponding to the bond review data, wherein the sentiment label includes a positive sentiment label and a negative sentiment label;

[0084] Based on the sentiment label, the influencing factors of bond returns are analyzed to obtain a target influencing factor, and the target influencing factor is used to evaluate the return characteristics of the bond.

[0085] Furthermore, the influencing factors of bond returns are analyzed based on the sentiment labels to obtain target influencing factors, including:

[0086] Based on the sentiment tags, determining the scores of various impact factors of the bonds corresponding to each bond review data;

[0087] Based on the scores of the impact factors, sorting the impact factors, and determining candidate impact factors based on the sorting results;

[0088] A correlation test between the impact factor and the bond yield is performed on each of the candidate impact factors, and a target impact factor is determined based on the test results.

[0089] Based on the target recognition model, the bond review data to be identified is identified to obtain the text sentiment score results, and the post labels are mapped to the bonds in order to construct a multi-dimensional influencing factor (also known as a public opinion factor) related to the bond market investment strategy. Specifically, users of financial forums may still post on holidays and after the market, so the constructed monthly frequency influencing factor (bond selection factor) uses the posting data from t-31 to t-1 as the bond selection factor on day t (beginning of the month).

[0090] In order to evaluate the effectiveness of stock selection factors, the factor IC test and the method of constructing quantile combinations are used for research. For the quantile combination test, the bonds are divided into 10 groups according to the factor values ​​from high to low, and the top combination to the bottom combination are constructed with equal weights. The long combination Top and the short combination Bottom are obtained to obtain the long-short combination (LS combination). The predictive ability of the factor is measured by the performance of this combination. The backtest frequency is monthly, and the position adjustment time point is the second trading day of each month. The backtest is performed at the opening price. The backtest time period is from January 1, 2022 to December 30, 2023. When selecting bonds, screen in the bond market.

[0091] The homogeneity of stockholder sentiment is characterized by the number of monthly positive posts / negative posts, and the number of monthly posts is used to characterize the attention of stockholders to the bond. The more posts, the more stockholders pay attention to the bond. The variance of the monthly daily attention (positive number is positive, negative number is negative) is calculated. If the factor value is large, it means that the attention of stockholders to the bond has changed a lot, rather than the attention remains unchanged. The three factors are synthesized in proportion to obtain the synthetic factor, and backtested. The IC of the synthetic factor reaches 5.14%, the annualized return of long and short is 6.59%, the Sharpe ratio of the long and short combination is 7.02, and the maximum drawdown rate of the long and short combination is 0.74%. The synthetic factor has a low correlation with traditional factors such as growth, momentum, value, quality, volatility, analyst consensus expectations, and technology, and the correlation coefficient does not exceed 0.3. Therefore, the bond selection factor based on forum sentiment analysis is a good supplement to the traditional bond selection factor and is conducive to providing additional information.

[0092] The target recognition model in the embodiment of the present application adopts the FinBERT architecture. Since FinBERT is a variant of the BERT model designed specifically for the financial field, it has better understanding and processing capabilities in financial terms and language characteristics. The FinBERT model can more accurately capture and understand professional discussions and emotional expressions in financial forums. The FinBERT model has been pre-trained and fine-tuned in the financial field, and can better extract deep semantic information from texts. This improves the accuracy of sentiment classification, thereby improving the quality of public opinion bond selection factors. Accurate sentiment analysis and professional understanding of the financial field make the constructed public opinion bond selection factors more reliable. This may lead to the constructed investment strategy showing a higher annualized rate of return in backtesting and practical applications. The public opinion factor constructed by the FinBERT model has a low correlation with existing factors and can provide additional information beyond traditional bond selection factors. This helps investment strategies reduce risks and increase risk-adjusted returns while diversifying their portfolios. The specialized training of the FinBERT model enables it to adapt to the rapid changes in financial texts and the emergence of new terms. This enables the strategy to quickly adapt to market changes and maintain the timeliness of its predictive capabilities.

[0093] In another embodiment of the present application, a model building device for bond data processing is provided. Figure 2 , the device comprises:

[0094] An acquisition unit 201 is used to identify the target bond review data based on the large model to obtain a first data set, wherein each data in the first data set includes bond review data and a corresponding sentiment label;

[0095] A training unit 202 is used to train the initial recognition model based on the bond dictionary in the bond field and the first data set until the output result of the initial recognition model meets the target condition, and determine the current initial recognition model as the target recognition model;

[0096] The bond dictionary includes several words in the bond field.

[0097] Optionally, the device further comprises:

[0098] The data preprocessing unit is used to collect initial review data for the bond field; perform data preprocessing on the initial review data to determine the bond code corresponding to each review data; and generate target bond review data based on each review data marked with a bond code.

[0099] Optionally, the large model includes a first model and a second model, and the acquisition unit includes:

[0100] A first identification subunit is used to identify each data in the target bond review data based on the first model to obtain a first sentiment label corresponding to the data;

[0101] A second identification subunit is used to identify each data in the target bond review data based on the second model to obtain a second sentiment label corresponding to the data;

[0102] A first determining subunit, configured to determine a target emotion label corresponding to each data based on the first emotion label and the second emotion label corresponding to each data;

[0103] A second determination subunit is used to determine the data marked with the target emotion label as a first data set;

[0104] The first model and the second model have different model structures.

[0105] Optionally, the method further comprises:

[0106] The dictionary construction unit is used to perform word segmentation processing on each bond review data to obtain the vocabulary of each bond review data; count the frequency of occurrence of each vocabulary, and form a bond dictionary with vocabulary whose frequency of occurrence is higher than the target frequency threshold.

[0107] Optionally, the training unit includes:

[0108] A first training subunit is used to perform unsupervised training on the initial recognition model based on a bond dictionary in the bond field to obtain a first initial model after unsupervised training;

[0109] A second training subunit is used to perform supervised training on the first initial model based on the first data set to obtain a second initial model after supervised training;

[0110] A third determination subunit is used to verify the second initial model based on a test set, and if the deviation range between the result output by the second initial model and the annotated label of the test data in the test set meets the target condition, determine the current second initial model as a target recognition model;

[0111] An adjustment subunit is used to adjust the model parameters of the second initial model if the deviation range does not meet the target condition, until the output result of the adjusted second initial model meets the target condition, and determine the currently adjusted second initial model as the target recognition model.

[0112] Optionally, the device further comprises:

[0113] an identification unit, configured to identify the bond review data to be identified based on the target identification model, and obtain a sentiment label corresponding to the bond review data, wherein the sentiment label includes a positive sentiment label and a negative sentiment label;

[0114] An analysis unit is used to analyze the influencing factors of bond returns based on the sentiment tags to obtain target influencing factors, and the target influencing factors are used to evaluate the return characteristics of the bonds.

[0115] Optionally, the analysis unit is configured to:

[0116] Based on the sentiment tags, determining the scores of various impact factors of the bonds corresponding to each bond review data;

[0117] Based on the scores of the impact factors, sorting the impact factors, and determining candidate impact factors based on the sorting results;

[0118] A correlation test between the impact factor and the bond yield is performed on each of the candidate impact factors, and a target impact factor is determined based on the test results.

[0119] Optionally, the model structure of the initial recognition model is a FinBERT model structure, and the FinBERT model structure can transfer knowledge on financial field corpus to downstream financial text application tasks.

[0120] It should be noted that the specific implementation of each unit and sub-unit in this embodiment can refer to the corresponding content in the previous text and will not be described in detail here.

[0121] In another embodiment of the present application, a storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the model building method for bond data processing as described in any one of the above items is implemented.

[0122] In another embodiment of the present application, an electronic device is further provided, and the electronic device may include:

[0123] A memory, used to store applications and data generated by the operation of the applications;

[0124] A processor is used to execute the application program to implement the model building method applied to bond data processing as described in any of the above items.

[0125] It should be noted that the specific implementation of the processor in this embodiment can refer to the corresponding content in the previous text and will not be described in detail here.

[0126] It should be noted that the above-mentioned processor or CPU can be at least one of an Application Specific Integrated Circuit (ASIC), a Digital Signal Processor (DSP), a Digital Signal Processing Device (DSPD), a Programmable Logic Device (PLD), a Field Programmable Gate Array (FPGA), a Central Processing Unit (CPU), a controller, a microcontroller, and a microprocessor. It can be understood that the electronic device that realizes the function of the above-mentioned processor can also be other, and the embodiments of the present application are not specifically limited.

[0127] It should be noted that the above-mentioned computer storage medium / memory can be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), a magnetic random access memory (FRAM), a flash memory (Flash Memory), a magnetic surface memory, an optical disc, or a compact disc read-only memory (CD-ROM) and other memories; it can also be various terminals including one or any combination of the above-mentioned memories, such as mobile phones, computers, tablet devices, personal digital assistants, etc.

[0128] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

[0129] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0130] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A model building method for bond data processing, characterized in that: The method comprises: Identify the target bond review data based on the large model to obtain a first data set, wherein each data in the first data set includes bond review data and a corresponding sentiment label; Based on the bond dictionary in the bond field and the first data set, the initial recognition model is trained until the output result of the initial recognition model meets the target condition, and the current initial recognition model is determined as the target recognition model; The bond dictionary includes several words in the bond field.

2. The method according to claim 1, characterized in that The method further comprises: Collect initial commentary data on the bond sector; Performing data preprocessing on the initial review data to determine a bond code corresponding to each review data; Based on the review data labeled with each bond code, target bond review data is generated.

3. The method according to claim 1, characterized in that The large model includes a first model and a second model. The method of identifying the target bond review data based on the large model to obtain a first data set includes: Identify each data in the target bond review data based on the first model to obtain a first sentiment label corresponding to the data; Identify each data in the target bond review data based on the second model to obtain a second sentiment label corresponding to the data; Based on the first emotion label and the second emotion label corresponding to each data, determining a target emotion label corresponding to each data; Determine the data marked with the target emotion label as the first data set; The first model and the second model have different model structures.

4. The method according to claim 1, characterized in that The method further comprises: Perform word segmentation processing on each bond review data to obtain the vocabulary of each bond review data; The frequency of occurrence of each word is counted, and the words with a frequency higher than the target frequency threshold are combined into a bond dictionary.

5. The method according to claim 1 or 4, characterized in that: The bond dictionary based on the bond field and the first data set, training the initial recognition model until the output result of the initial recognition model meets the target condition, and determining the current initial recognition model as the target recognition model, includes: Performing unsupervised training on the initial recognition model based on a bond dictionary in the bond field to obtain a first initial model after unsupervised training; Performing supervised training on the first initial model based on the first data set to obtain a supervised trained second initial model; Verify the second initial model based on the test set, and if the deviation range between the result output by the second initial model and the annotated label of the test data in the test set meets the target condition, determine the current second initial model as the target recognition model; If the deviation range does not meet the target condition, the model parameters of the second initial model are adjusted until the output result of the adjusted second initial model meets the target condition, and the currently adjusted second initial model is determined as the target recognition model.

6. The method according to claim 1, characterized in that The method further comprises: Based on the target recognition model, the bond review data to be recognized is recognized to obtain a sentiment label corresponding to the bond review data, wherein the sentiment label includes a positive sentiment label and a negative sentiment label; Based on the sentiment label, the influencing factors of bond returns are analyzed to obtain a target influencing factor, and the target influencing factor is used to evaluate the return characteristics of the bond.

7. The method according to claim 6, characterized in that The influencing factors of bond returns are analyzed based on the sentiment labels to obtain target influencing factors, including: Based on the sentiment tags, determining the scores of various impact factors of the bonds corresponding to each bond review data; Based on the scores of the impact factors, sorting the impact factors, and determining candidate impact factors based on the sorting results; A correlation test between the impact factor and the bond yield is performed on each of the candidate impact factors, and a target impact factor is determined based on the test results.

8. The method according to claim 1, characterized in that The model structure of the initial recognition model is the FinBERT model structure, which can transfer the knowledge on the financial field corpus to the application tasks of downstream financial texts.

9. A model building device for bond data processing, characterized in that: The device comprises: An acquisition unit, configured to identify the target bond review data based on the large model to obtain a first data set, wherein each data in the first data set includes the bond review data and a corresponding sentiment label; A training unit, configured to train an initial recognition model based on a bond dictionary in the bond field and the first data set until an output result of the initial recognition model meets a target condition, and determine the current initial recognition model as a target recognition model; The bond dictionary includes several words in the bond field.

10. An electronic device, characterized in that: include: Memory, used to store programs; A processor is used to execute the program, wherein the program is specifically used to implement the model building method applied to bond data processing as described in any one of claims 1 to 8.

Images