Financial emotion analysis system

Abstract

The invention discloses a financial emotion analysis system which can comprehensively and efficiently process complex financial text data through accurate recognition of an emotion analysis module, dynamic optimization of a reward function and reinforcement learning and accurate report generation of a public opinion evaluation module. The emotion analysis module ensures that the system accurately recognizes emotion segmented words and trend expression words in a large amount of unstructured data, and key emotion and trend information is provided. And through combination of the reward function and reinforcement learning, the system can automatically optimize semantic vectors of emotions and trends in a continuously changing market environment, and the prediction capability of the model is continuously improved.

Description

Technical Field

[0001] The present invention belongs to the field of financial sentiment analysis systems, and in particular, relates to a financial sentiment analysis system. Background Art

[0002] In the financial market, the analysis of sentiment and trends is crucial for investment decisions and risk management. With the development of market globalization, the sources of information in the financial market are becoming more extensive and complex, including unstructured data such as news reports, social media, and market comments. This information is not only huge in quantity, but also complex and diverse in content, making it difficult for traditional analysis methods to fully capture and understand changes in market sentiment.

[0003] Traditional financial sentiment analysis methods mainly rely on manual experience and simple statistical models. These methods may provide certain reference value in specific scenarios, but they show obvious limitations when dealing with the huge and diverse data in modern financial markets. First, these methods are difficult to effectively process unstructured data, such as news reports and social media comments. Such data are not only huge in quantity, but also diverse in form, and contain complex semantics and context. In addition, since these methods are usually based on preset rules and fixed statistical patterns, when faced with the complexity of language expression and the diversity of emotional expression, it is often difficult to accurately capture the subtle differences in the text, resulting in distorted or one-sided analysis results. Furthermore, traditional methods lack the ability to dynamically optimize and cannot adapt to the rapid changes in the market environment in a timely manner. In the case of rapidly changing market sentiment, relying on static analysis models may lead to lagging predictions and missing important market signals. The lag and lack of accuracy of such analysis results further affect the scientificity and timeliness of investment decisions and increase the risks in financial operations. With the increasing complexity of the market environment and information sources, the limitations of traditional sentiment analysis methods have become increasingly prominent.

[0004] In view of this, the present invention is proposed. Summary of the invention

[0005] The technical problem to be solved by the present invention is to overcome the deficiencies of the prior art and provide a financial sentiment analysis system to solve the problems raised in the above-mentioned background technology.

[0006] In order to solve the above technical problems, the basic concept of the technical solution adopted by the present invention is:

[0007] A financial sentiment analysis system includes: a text data receiving module for receiving text data containing financial related content;

[0008] A sentiment analysis module, connected to the text data receiving module, the sentiment analysis module identifies sentiment segmentation and trend expression words for the received financial text data based on a preset large language model, and calculates initial semantic vectors of the sentiment segmentation and trend expression words;

[0009] A reward function construction module, connected to the sentiment analysis module, for generating a corresponding reward function model according to a preset sentiment score and trend score, and enhancing the initial semantic vectors of the sentiment segmentation words and trend expression words based on the reward function model to obtain an enhanced semantic vector;

[0010] A reinforcement learning optimization module, connected to the reward function construction module, for optimizing the large language model based on the reward function model and a preset reinforcement learning model to generate an optimized language model;

[0011] A financial institution public opinion assessment module is connected to the reinforcement learning optimization module. The financial risk assessment module performs regression analysis on the financial text data processed by the sentiment analysis module based on the optimized language model, calculates the sentiment score of the financial text, and generates a financial institution public opinion report based on the financial institution public opinion data of the target area according to the sentiment score statistics;

[0012] The public opinion response plan generation module is used to obtain the public opinion report of the financial institution and extract keywords based on the public opinion report of the financial institution to generate corresponding plans.

[0013] Optionally, the steps for building and training a large language model are:

[0014] Collect text data related to the financial field and divide the acquired data into training sets, validation sets, and fine-tuning sets. Use the training sets to train the language model and set the number of layers, size of hidden units, and number of attention heads of the large language model.

[0015] In combination with the characteristics of the financial field, we designed a self-supervised learning task and fine-tuned the pre-trained model on the financial field data through a fine-tuning set;

[0016] Use the validation set to evaluate the fine-tuned large language model, and deploy the trained large language model on the trading platform for use.

[0017] Optionally, the steps of identifying sentiment segmentation words and trend expression words for the received financial text data based on a preset large language model, and calculating the initial semantic vectors of the sentiment segmentation words and trend expression words are as follows:

[0018] The received financial text data is first preprocessed, and the pre-set large language model is used to perform sentiment segmentation and trend-related expression word recognition on the preprocessed text data;

[0019] For the identified sentiment segmentation words and trend expression words, the formula Calculate their initial semantic vectors respectively, where v i represents the initial semantic vector of sentiment segmentation or trend expression word i, represents the context window of word i, w j represents the weight of context word j, freq(j) represents the frequency of context word j in the entire corpus, ∈ is a smoothing parameter to prevent computational problems caused by low or zero frequency, E j represents the embedding vector of context word j, Z is the number of context words, and α is the weight coefficient of the context embedding vector. It controls the influence of the context on the initial semantic vector, P i Represents the sentiment embedding vector of word i in the sentiment word set, T i represents the trend embedding vector of word i in the trend expression word set, β is the weighting coefficient of sentiment and trend embedding vector, γ and δ are the sentiment embedding vectors p respectively. i and the trend embedding vector T i The weighting coefficient of .

[0020] Optionally, the steps of using a preset large language model to identify sentiment segmentation words and trend expression words for received financial text data and calculating initial semantic vectors of the sentiment segmentation words and trend expression words are as follows:

[0021] Use the big language model to identify sentiment words in financial texts and assign sentiment scores and trend scores based on their usage in context and their matching degree with the preset sentiment labels in the big language model;

[0022] According to the analyzed sentiment score and trend score, the formula Ω i =α·ψ i +β·τ i ·v i Construct a reward function model, which evaluates the initial semantic vectors of sentiment segmentation and trend expression words according to sentiment scores and trend scores, where Ω i Represents the reward function value of sentiment segmentation or trend expression word i. α and β are adjustment coefficients used to balance the contribution of sentiment score and trend score in the final reward function. ψ i Represents the sentiment score of sentiment participle i. τ i represents the trend score of trend expression word i;

[0023] Based on the feedback signal of the constructed reward function model, the initial semantic vectors of the sentiment segmentation words and trend expression words are adjusted and gradually optimized to make them more consistent with the emotions and trend directions represented by the sentiment scores and trend scores;

[0024] According to the adjustment results of the semantic vector, the relevant parameters in the large language model are updated and through multiple iterations and repeated optimization, the model gradually adjusts its internal structure and weights.

[0025] Optionally, the large language model is optimized based on the reward function model and the preset reinforcement learning model to generate the optimized language model in the following steps:

[0026] Based on the large language model, a reinforcement learning framework is set up, including states, actions, and rewards;

[0027] Using the previously generated reward function model, according to the performance of the model in the current state, through the formula Calculate the corresponding reward value, where the reward value reflects the accuracy of the current model output and the degree of match with the expected target. R Represents the final calculated reward value, which reflects the accuracy of the current model output and the degree of match with the expected goal. α is an adjustment factor used to control the overall magnitude of the reward value. N Indicates the number of samples, usually the number of samples used to evaluate model performance. For the i The expected target reward value of samples, which is the target value set based on the ideal output, Indicates i The absolute error between the actual output of a sample and the expected target, which measures the gap between the model output and the target value;

[0028] Feed the reward signal back to the reinforcement learning model to guide it to choose the best action in the subsequent training process to maximize the cumulative reward;

[0029] Based on the reward signal, the reinforcement learning model updates its policy. After performing the policy update, the action is selected based on the current policy and the parameters of the large language model are adjusted.

[0030] Optionally, based on the optimized language model, regression analysis is performed on the financial text data processed by the sentiment analysis module to calculate the sentiment score of the financial text, and the public opinion data of financial institutions in the target area are statistically analyzed according to the sentiment score to predict future financial volatility indicators, and then the steps of evaluating the financial risk level are as follows:

[0031] The financial text data processed by the sentiment analysis module is used as input, including the extracted enhanced semantic vectors of sentiment segmentation words and trend expression words and the obtained enhanced semantic vectors of sentiment segmentation words and trend expression words as the feature input of regression analysis, and the target sentiment score of the financial text is set as the target variable of the regression model;

[0032] Use the annotated financial text data set to build and train the regression model. After training, apply the regression model to calculate the sentiment score of new financial text data. Each input text data will generate a sentiment score to reflect its sentiment tendency and intensity.

[0033] Summarize and count the sentiment scores of all processed financial text data, and classify and analyze the data based on time, geographic location, and text source dimensions to obtain public opinion data of financial institutions in the target area;

[0034] Analyze the aggregated public opinion data, identify the changing trends of sentiment scores and build a prediction model to predict future public opinion fluctuation indicators;

[0035] Based on historical data and financial institution research, a risk threshold for the public opinion fluctuation index of a financial institution is set. When the predicted public opinion fluctuation index exceeds this threshold, it indicates that the financial institution may face a higher risk.

[0036] Compare the predicted future public opinion fluctuation index with the set risk threshold. If the predicted index is higher than the threshold, it is judged that the public opinion fluctuation of the financial institution is high; if the predicted index is lower than or equal to the threshold, it is judged that the risk of public opinion fluctuation is low. Based on the prediction results, generate a public opinion assessment report for the financial institution.

[0037] Optionally, the steps of obtaining a public opinion report of a financial institution and extracting keywords based on the public opinion report of the financial institution to generate a corresponding solution are as follows:

[0038] Perform text analysis on the obtained public opinion reports, split them into paragraphs, sentences or more fine-grained word levels, and perform grammatical and semantic analysis on the parsed texts;

[0039] The large language model extracts keywords with potential meaning and predictive value through contextual relationships and semantic understanding, and clusters and classifies the extracted keywords;

[0040] Analyze the correlation between the extracted keywords and each paragraph or event in the public opinion report and determine the context and importance of each keyword in the public opinion report through semantic similarity calculation or dependency analysis. Finally, determine the type of solution to be generated based on the category and importance of the keyword;

[0041] Financial institutions verify the generated solutions through simulation or backtesting methods. They observe the performance in historical data and simulation environments to verify the effectiveness of the strategy and adjust the solutions based on the verification results to remove unreasonable parts.

[0042] Summarize the results of keyword extraction, correlation analysis, and response plan generation into a complete public opinion response report.

[0043] A financial sentiment analysis system also includes a database, which is used to store data sent by a text data receiving module, a financial institution public opinion evaluation module, and a public opinion response plan generation module via communication signals, and encrypts the data using blockchain technology during data reception and transmission.

[0044] After adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art. Of course, any product implementing the present invention does not necessarily need to achieve all the advantages described below at the same time:

[0045] Through the accurate identification of the sentiment analysis module, the dynamic optimization of the reward function and reinforcement learning, and the accurate report generation of the public opinion evaluation module, this application can comprehensively and efficiently process complex financial text data. The sentiment analysis module ensures that the system accurately identifies sentiment segmentation words and trend expression words in a large amount of unstructured data, providing key sentiment and trend information. The combination of reward functions and reinforcement learning enables the system to automatically optimize the semantic vectors of sentiment and trends in a constantly changing market environment, and continuously improve the model's predictive ability.

[0046] The specific implementation modes of the present invention are further described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The drawings described below are only some embodiments. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0048] In the figure:

[0049] Figure 1 Block diagram of the financial sentiment analysis system based on a large language model.

[0050] It should be noted that these drawings and textual descriptions are not intended to limit the conceptual scope of the present invention in any way, but are intended to illustrate the concept of the present invention for those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0051] The present invention will now be described in further detail with reference to the accompanying drawings.

[0052] See also Figure 1 As shown, in this embodiment, a financial sentiment analysis system is provided, including a text data receiving module for receiving text data containing financial related content;

[0053] A sentiment analysis module, connected to the text data receiving module, the sentiment analysis module identifies sentiment segmentation and trend expression words for the received financial text data based on a preset large language model, and calculates initial semantic vectors of the sentiment segmentation and trend expression words;

[0054] A reward function construction module, connected to the sentiment analysis module, for generating a corresponding reward function model according to a preset sentiment score and trend score, and enhancing the initial semantic vectors of the sentiment segmentation words and trend expression words based on the reward function model to obtain an enhanced semantic vector;

[0055] A reinforcement learning optimization module, connected to the reward function construction module, for optimizing the large language model based on the reward function model and a preset reinforcement learning model to generate an optimized language model;

[0056] A financial institution public opinion assessment module is connected to the reinforcement learning optimization module. The financial risk assessment module performs regression analysis on the financial text data processed by the sentiment analysis module based on the optimized language model, calculates the sentiment score of the financial text, and generates a financial institution public opinion report based on the financial institution public opinion data of the target area according to the sentiment score statistics;

[0057] The public opinion response plan generation module is used to obtain the public opinion report of the financial institution and extract keywords based on the public opinion report of the financial institution to generate corresponding plans.

[0058] In this embodiment, the steps of establishing and training the large language model are as follows:

[0059] Text data related to the financial field is collected and the acquired data is divided into training set, validation set and fine-tuning set. The training set is used to train the language model and set the number of layers, hidden unit size and number of attention heads of the large language model. The training set accounts for 70% of the total data and is mainly used for model training. The validation set accounts for 15% of the total data and is used to evaluate the performance of the model during training. The fine-tuning set accounts for 15% of the total data and is specifically used for fine-tuning on specific financial tasks. The number of layers of the large language model is a 12-layer Transformer encoder. The size of the hidden unit is 768 hidden units. The number of attention heads is 12.

[0060] In combination with the characteristics of the financial field, we designed a self-supervised learning task and fine-tuned the pre-trained model on the financial field data through a fine-tuning set;

[0061] Use the validation set to evaluate the fine-tuned large language model, and deploy the trained large language model to the trading platform for use. The setting of the large language model can improve the accuracy and practicality of the model in financial text analysis, thereby better supporting trading decisions and financial institution predictions.

[0062] In this embodiment, the steps of identifying sentiment segmentation words and trend expression words for received financial text data based on a preset large language model and calculating the initial semantic vectors of the sentiment segmentation words and trend expression words are as follows:

[0063] The received financial text data is first preprocessed, and the pre-set large language model is used to perform sentiment segmentation and trend-related expression word recognition on the preprocessed text data;

[0064] For the identified sentiment segmentation words and trend expression words, the formula Calculate their initial semantic vectors respectively, where v i represents the initial semantic vector of sentiment segmentation or trend expression word i, represents the context window of word i, w j represents the weight of context word j, freq(j) represents the frequency of context word j in the entire corpus, ∈ is a smoothing parameter to prevent computational problems caused by low or zero frequency, E j represents the embedding vector of context word j, Z is the number of context words, and α is the weight coefficient of the context embedding vector. It controls the influence of the context on the initial semantic vector, P i Represents the sentiment embedding vector of word i in the sentiment word set, T i represents the trend embedding vector of word i in the trend expression word set, β is the weighting coefficient of sentiment and trend embedding vector, γ and δ are the sentiment embedding vectors P respectively. i and the trend embedding vector T i By introducing smoothing parameters and weighting coefficients, the stability and robustness of the model are enhanced, and ultimately the expressive power of semantic vectors in sentiment and trend analysis is improved, which helps to improve the accuracy and application effect of the model in financial text analysis.

[0065] In this embodiment, the steps of the preset large language model identifying the sentiment segmentation words and trend expression words on the received financial text data and calculating the initial semantic vectors of the sentiment segmentation words and trend expression words are as follows:

[0066] Use the big language model to identify sentiment words in financial texts and assign sentiment scores and trend scores based on their usage in context and their matching degree with the preset sentiment labels in the big language model;

[0067] According to the analyzed sentiment score and trend score, the formula Ω i =α·ψ i +β·τ i ·v i Construct a reward function model, which evaluates the initial semantic vectors of sentiment segmentation and trend expression words according to sentiment scores and trend scores, where Ωi Represents the reward function value of sentiment segmentation or trend expression word i. α and β are adjustment coefficients used to balance the contribution of sentiment score and trend score in the final reward function. ψ i Represents the sentiment score of sentiment participle i. τ i represents the trend score of trend expression word i;

[0068] Based on the feedback signal of the constructed reward function model, the initial semantic vectors of the sentiment segmentation words and trend expression words are adjusted and gradually optimized to make them more consistent with the emotions and trend directions represented by the sentiment scores and trend scores;

[0069] According to the adjustment results of the semantic vector, the relevant parameters in the large language model are updated and the model gradually adjusts its internal structure and weights through multiple iterations and repeated optimization. Through multiple iterations and optimization, the internal structure and weights of the model are gradually adjusted to ensure the accuracy and effect of the model in sentiment and trend analysis tasks. This process helps to improve the accuracy and application value of the model in financial sentiment analysis.

[0070] In this embodiment, the steps of optimizing the large language model based on the reward function model and the preset reinforcement learning model to generate the optimized language model are as follows:

[0071] Based on the large language model, a reinforcement learning framework is set up, including states, actions, and rewards;

[0072] Using the previously generated reward function model, according to the performance of the model in the current state, through the formula Calculate the corresponding reward value, where the reward value reflects the accuracy of the current model output and the degree of match with the expected target. R Represents the final calculated reward value, which reflects the accuracy of the current model output and the degree of match with the expected goal. α is an adjustment factor used to control the overall magnitude of the reward value. N Indicates the number of samples, usually the number of samples used to evaluate model performance. For the i The expected target reward value of samples, which is the target value set based on the ideal output, Indicates i The absolute error between the actual output of a sample and the expected target, which measures the gap between the model output and the target value;

[0073] Feed the reward signal back to the reinforcement learning model to guide it to choose the best action in the subsequent training process to maximize the cumulative reward;

[0074] Based on the reward signal, the reinforcement learning model updates its strategy. After executing the strategy update, the action is selected based on the current strategy to adjust the parameters of the large language model. The reinforcement learning model can use the policy gradient method and Q-learning in the existing technology to update the model. Through reinforcement learning, the model can dynamically adjust its parameters and strategies, so that it can better optimize the output results when processing complex financial text tasks and gradually approach the optimal solution. This process enables the model to not only perform well in initial training, but also to continuously improve in actual applications, adapt to the changing environment of financial institutions, and ultimately improve the application value and accuracy of the model in the financial field.

[0075] In actual use, first, set the current prediction state of the reinforcement learning model, such as the model predicting the future rise and fall trend of a certain stock, and the operations that the model can adjust, such as changing the prediction time window, adjusting the weight of the predicted technical indicators, etc. According to the actual financial institution data feedback, judge the correctness of the model prediction and give a reward value.

[0076] Use the previously generated reward function model to calculate the accuracy of the model in predicting stock price increases and decreases. For example, if the model predicts that stock A will increase by 10%, and the actual result is an increase of 8%, a reward value R is calculated based on the error. The higher the reward value, the more accurate the prediction. If the model's prediction matches the actual result, the reward value is higher. At this point, the reward signal is fed back to the reinforcement learning model to guide the model to continue using the current strategy; if the reward value is low, the model will adjust the strategy, such as changing the weight of technical indicators or changing the prediction method.

[0077] Based on the feedback reward signal, the reinforcement learning model updates its strategy. For example, the model may find that certain technical indicators are more effective under certain financial institution conditions, and thus give priority to these indicators in the next forecast. At the same time, the model will adjust the parameters of the large language model to make it more adaptable to the new strategy.

[0078] In this embodiment, based on the optimized language model, regression analysis is performed on the financial text data processed by the sentiment analysis module, the sentiment score of the financial text is calculated, and the public opinion data of the financial institutions in the target area are counted according to the sentiment score to predict the future financial volatility index, and then the steps of evaluating the financial risk level are as follows:

[0079] The financial text data processed by the sentiment analysis module is used as input, including the extracted enhanced semantic vectors of sentiment segmentation words and trend expression words and the obtained enhanced semantic vectors of sentiment segmentation words and trend expression words as the feature input of regression analysis, and the target sentiment score of the financial text is set as the target variable of the regression model;

[0080] Use the annotated financial text data set to build and train the regression model. After training, apply the regression model to calculate the sentiment score of new financial text data. Each input text data will generate a sentiment score to reflect its sentiment tendency and intensity.

[0081] Summarize and count the sentiment scores of all processed financial text data, and classify and analyze the data based on time, geographic location, and text source dimensions to obtain public opinion data of financial institutions in the target area;

[0082] Analyze the aggregated public opinion data, identify the changing trends of sentiment scores and build a prediction model to predict future public opinion fluctuation indicators;

[0083] Based on historical data and financial institution research, a risk threshold for the public opinion fluctuation index of a financial institution is set. When the predicted public opinion fluctuation index exceeds this threshold, it indicates that the financial institution faces high risk.

[0084] Compare the predicted future public opinion fluctuation index with the set risk threshold. If the predicted index is higher than the threshold, the public opinion fluctuation of the financial institution is judged to be high; if the predicted index is lower than or equal to the threshold, the public opinion fluctuation risk is judged to be low. Based on the prediction results, a public opinion assessment report for the financial institution is generated. By using the enhanced semantic vectors of sentiment segmentation and trend expression words as feature inputs for regression analysis, the sentiment score can be accurately calculated and the future public opinion fluctuation index can be predicted. In this way, potential financial institution risks can be identified and risk assessment can be performed according to the set risk threshold. Finally, the public opinion assessment report for the financial institution generated based on the prediction results provides decision makers with reliable financial institution analysis.

[0085] In actual use: A piece of news about a company may contain sentiment words and trend expressions such as "declining profits" and "turbulence in financial institutions". After model analysis, the enhanced semantic vectors of these words will be used as input features for regression analysis. Use historical financial data sets to train regression models. The model learns how to generate sentiment scores that reflect sentiment tendencies based on the features of the input sentiment words and trend expressions. After training, when new financial data is input, the model will generate sentiment scores for each piece of data, such as "the optimism of financial institutions is 0.7" or "the panic index of financial institutions is 0.8".

[0086] Aggregate the resulting sentiment scores and categorize them by time, geography, or source (e.g., news, social media). For example, you might find that financial institution sentiment gradually shifts from "optimistic" to "pessimistic" in a specific region or over a period of time, which could indicate potential financial institution volatility.

[0087] By analyzing the trend of sentiment scores, a prediction model is established to predict future public opinion fluctuation indicators. Based on financial institution research and historical data, a risk threshold is set, such as "when the financial institution panic index exceeds 0.6, the financial institution may face a higher risk." If the predicted public opinion fluctuation index is higher than this threshold, it is judged that the risk of the financial institution has increased.

[0088] In this embodiment, the steps of obtaining the public opinion report of the financial institution and extracting keywords according to the public opinion report of the financial institution to generate the corresponding solution are as follows:

[0089] Perform text analysis on the obtained public opinion reports, split them into paragraphs, sentences or more fine-grained word levels, and perform grammatical and semantic analysis on the parsed texts;

[0090] The large language model extracts keywords with potential meaning and predictive value through contextual relationships and semantic understanding, and clusters and classifies the extracted keywords;

[0091] Analyze the correlation between the extracted keywords and each paragraph or event in the public opinion report and determine the context and importance of each keyword in the public opinion report through semantic similarity calculation or dependency analysis. Finally, determine the type of solution to be generated based on the category and importance of the keyword;

[0092] Financial institutions verify the generated solutions through simulation or backtesting methods. They observe the performance in historical data and simulation environments to verify the effectiveness of the strategy and adjust the solutions based on the verification results to remove unreasonable parts.

[0093] The results of keyword extraction, correlation analysis, and response plan generation are summarized into a complete public opinion response report. Through in-depth grammatical and semantic analysis of public opinion reports by large language models, keywords with potential significance and predictive value can be extracted, and effective response plans can be generated based on the correlation of these keywords. Subsequently, financial institutions verify the effectiveness of the plans through simulation or backtesting to ensure that the generated strategies have practical application value. Finally, all results are summarized into a complete public opinion response report, providing decision makers with reliable financial institution analysis and response strategies, thereby improving the accuracy of financial institutions' predictions and the success rate of strategy implementation.

[0094] A public opinion report about a company's financial crisis was received. By parsing the public opinion report, the system splits the text into multiple paragraphs, sentences, and even specific words, such as "bankruptcy" and "debt restructuring". Then, the large language model performs grammatical and semantic analysis on these texts to extract keywords with potential meanings, such as "crisis management" and "financial institution response".

[0095] The large language model extracts key terms from the report based on contextual relationships and semantic understanding. For example, during the analysis process, the model extracts keywords such as "financial institution confidence" and "investor panic" and clusters them into relevant risk factor categories.

[0096] Next, the system analyzes the relevance of these keywords to specific paragraphs or events in the public opinion report, and determines the importance of each keyword in the report through semantic similarity and dependency analysis. For example, the model may identify that "confidence of financial institutions" is mentioned repeatedly in multiple key paragraphs in the report, indicating that this is an issue that needs to be addressed as a priority. Based on this, a response plan for "restoring confidence in financial institutions" is generated, such as "strengthening financial institutions' communication strategies."

[0097] Finally, all the results of keyword extraction, correlation analysis and response plans are summarized into a complete public opinion response report. The report not only details the current status of financial institutions’ confidence issues, but also provides optimized response strategies and suggests how to implement these strategies to minimize risks and restore confidence in financial institutions.

[0098] A financial sentiment analysis system also includes a database, which is used to store data sent by a text data receiving module, a financial institution public opinion evaluation module, and a public opinion response plan generation module via communication signals, and encrypts the data using blockchain technology during data reception and transmission.

[0099] The present invention is not limited to the above-mentioned embodiments. Anyone should be aware that any structural changes made under the enlightenment of the present invention, and any technical solutions that are the same or similar to the present invention, fall within the protection scope of the present invention. The technology, shape, and structural parts not described in detail in the present invention are all well-known technologies.

Claims

1. A financial sentiment analysis system, characterized in that: include: A text data receiving module, used for receiving text data containing financial related content; A sentiment analysis module, connected to the text data receiving module, the sentiment analysis module identifies sentiment segmentation and trend expression words for the received financial text data based on a preset large language model, and calculates initial semantic vectors of the sentiment segmentation and trend expression words; A reward function construction module, connected to the sentiment analysis module, for generating a corresponding reward function model according to a preset sentiment score and trend score, and enhancing the initial semantic vectors of the sentiment segmentation words and trend expression words based on the reward function model to obtain an enhanced semantic vector; A reinforcement learning optimization module, connected to the reward function construction module, for optimizing the large language model based on the reward function model and a preset reinforcement learning model to generate an optimized language model; A financial institution public opinion assessment module is connected to the reinforcement learning optimization module. The financial risk assessment module performs regression analysis on the financial text data processed by the sentiment analysis module based on the optimized language model, calculates the sentiment score of the financial text, and generates a financial institution public opinion report based on the financial institution public opinion data of the target area according to the sentiment score statistics; The public opinion response plan generation module is used to obtain the public opinion report of the financial institution and extract keywords based on the public opinion report of the financial institution to generate corresponding plans.

2. A financial sentiment analysis system according to claim 1, characterized in that: The steps for building and training a large language model are: Collect text data related to the financial field and divide the acquired data into training sets, validation sets, and fine-tuning sets. Use the training sets to train the language model and set the number of layers, size of hidden units, and number of attention heads of the large language model. In combination with the characteristics of the financial field, we designed a self-supervised learning task and fine-tuned the pre-trained model on the financial field data through a fine-tuning set; Use the validation set to evaluate the fine-tuned large language model, and deploy the trained large language model on the trading platform for use.

3. A financial sentiment analysis system according to claim 2, characterized in that: The steps of identifying sentiment segmentation words and trend expression words for the received financial text data based on the preset large language model and calculating the initial semantic vectors of the sentiment segmentation words and trend expression words are as follows: The received financial text data is first preprocessed, and the pre-set large language model is used to perform sentiment segmentation and trend-related expression word recognition on the preprocessed text data; For the identified sentiment segmentation words and trend expression words, the formula Calculate their initial semantic vectors respectively, where v i represents the initial semantic vector of sentiment segmentation or trend expression word i, represents the context window of word i, w j represents the weight of context word j, freq(j) represents the frequency of context word j in the entire corpus, ∈ is a smoothing parameter to prevent computational problems caused by low or zero frequency, E j represents the embedding vector of context word j, Z is the number of context words, α is the weighting coefficient of the context embedding vector, which controls the influence of the context on the initial semantic vector, and P i Represents the sentiment embedding vector of word i in the sentiment word set, T i represents the trend embedding vector of word i in the trend expression word set, β is the weighting coefficient of sentiment and trend embedding vector, γ and δ are the sentiment embedding vectors P respectively. i and the trend embedding vector T i The weighting coefficient of .

4. A financial sentiment analysis system according to claim 3, characterized in that: The steps of the preset large language model identifying sentiment segmentation words and trend expression words for the received financial text data and calculating the initial semantic vectors of the sentiment segmentation words and trend expression words are as follows: Use the big language model to identify sentiment words in financial texts and assign sentiment scores and trend scores based on their usage in context and their matching degree with the preset sentiment labels in the big language model; According to the analyzed sentiment score and trend score, the formula Ω i =α·ψ i +β·τ i ·v i Construct a reward function model, which evaluates the initial semantic vectors of sentiment segmentation and trend expression words according to sentiment scores and trend scores, where Ω i represents the reward function value of the sentiment segmentation or trend expression word i, α and β are adjustment coefficients used to balance the contribution of the sentiment score and trend score in the final reward function, ψ i represents the sentiment score of sentiment participle i, τ i represents the trend score of trend expression word i; Based on the feedback signal of the constructed reward function model, the initial semantic vectors of the sentiment segmentation words and trend expression words are adjusted and gradually optimized to make them more consistent with the emotions and trend directions represented by the sentiment scores and trend scores; According to the adjustment results of the semantic vector, the relevant parameters in the large language model are updated and through multiple iterations and repeated optimization, the model gradually adjusts its internal structure and weights.

5. A financial sentiment analysis system according to claim 4, characterized in that: The steps of optimizing the large language model based on the reward function model and the preset reinforcement learning model to generate the optimized language model are as follows: Based on the large language model, a reinforcement learning framework is set up, including states, actions, and rewards; Using the previously generated reward function model, according to the performance of the model in the current state, through the formula Calculate the corresponding reward value, where the reward value reflects the accuracy of the current model output and the degree of match with the expected goal. R represents the final calculated reward value, which reflects the accuracy of the current model output and the degree of match with the expected goal. α is an adjustment coefficient used to control the overall amplitude of the reward value. N represents the number of samples, which is usually the number of samples used to evaluate model performance. is the expected target reward value of the i-th sample, which is the target value set according to the ideal output. Represents the absolute error between the actual output of the i-th sample and the expected target; Feed the reward signal back to the reinforcement learning model to guide it to choose the best action in the subsequent training process to maximize the cumulative reward; Based on the reward signal, the reinforcement learning model updates its policy. After performing the policy update, the action is selected based on the current policy and the parameters of the large language model are adjusted.

6. A financial sentiment analysis system according to claim 5, characterized in that: Based on the optimized language model, regression analysis is performed on the financial text data processed by the sentiment analysis module to calculate the sentiment score of the financial text, and the public opinion data of the financial institutions in the target area are counted according to the sentiment score to predict the future public opinion fluctuation index of the financial institutions, and then the steps of evaluating the public opinion level of the financial institutions are as follows: The financial text data processed by the sentiment analysis module is used as input and the enhanced semantic vectors of sentiment segmentation and trend expression words are obtained as feature inputs for regression analysis. The target sentiment score of the financial text is set as the target variable of the regression model. Use the annotated financial text data set to build and train the regression model. After training, apply the regression model to calculate the sentiment score of new financial text data. Each input text data will generate a sentiment score to reflect its sentiment tendency and intensity. Summarize and count the sentiment scores of all processed financial text data, and classify and analyze the data based on time, geographic location, and text source dimensions to obtain public opinion data of financial institutions in the target area; Analyze the aggregated public opinion data, identify the changing trends of sentiment scores and build a prediction model to predict future public opinion fluctuation indicators; Based on historical data and financial institution research, a risk threshold for the public opinion fluctuation index of a financial institution is set. When the predicted public opinion fluctuation index exceeds this threshold, it indicates that the financial institution faces high risk. Compare the predicted future public opinion fluctuation index with the set risk threshold. If the predicted index is higher than the threshold, it is judged that the public opinion fluctuation of the financial institution is high; if the predicted index is lower than or equal to the threshold, it is judged that the risk of public opinion fluctuation is low. Based on the prediction results, generate a public opinion assessment report for the financial institution.

7. A financial sentiment analysis system according to claim 6, characterized in that: The steps to obtain the public opinion report of financial institutions and extract keywords based on the public opinion report of financial institutions and generate the corresponding solution are as follows: Perform text analysis on the obtained public opinion reports, split them into paragraphs, sentences or more fine-grained word levels, and perform grammatical and semantic analysis on the parsed texts; The large language model extracts keywords with potential meaning and predictive value through contextual relationships and semantic understanding, and clusters and classifies the extracted keywords; Analyze the correlation between the extracted keywords and each paragraph or event in the public opinion report and determine the context and importance of each keyword in the public opinion report through semantic similarity calculation. Finally, determine the type of solution to be generated based on the category and importance of the keyword. Financial institutions verify the generated solutions through simulations, observe the performance in historical data and simulation environments to verify the effectiveness of the strategies and adjust the solutions based on the verification results to remove unreasonable parts; Summarize the results of keyword extraction, correlation analysis, and response plan generation into a complete public opinion response report.

8. A financial sentiment analysis system according to claim 1, characterized in that: It also includes a database, which is used to store data sent by the text data receiving module, the financial institution public opinion evaluation module, and the public opinion response plan generation module via communication signals, and encrypts the data using blockchain technology during the data reception and transmission process.

Images