Financial data intelligent interaction method based on AI big model dialogue

Abstract

The invention relates to the field of AI dialogues, and discloses an intelligent financial data interaction method based on AI large model dialogues, which comprises the following steps: analyzing input content, performing formatting processing and semantic analysis on the analyzed text, identifying and correcting logic errors in the text, and performing feature extraction on the current input text in combination with historical dialogues, so as to obtain the intelligent financial data interaction method based on the AI large model dialogues. Analyzing and deducing the intention of the user, identifying an entity related to financial data, querying in a financial knowledge graph, generating a dialogue for understanding the intention of the user in combination with a historical dialogue context, sending the dialogue to an interactive interface, inquiring whether the intention understood by the user is correct or not, waiting for further confirmation of the user, and receiving confirmation information of the user. The intention is divided into the atomic tasks, the user tasks are recognized, then the corresponding financial data interface is called, the needed financial data are obtained, the financial query information is generated and fed back, the accuracy of semantic understanding and the user intention is improved, and the recognition result is more accurate.

Description

Technical Field

[0001] This invention relates to the field of AI dialogue technology, and more specifically to a method for intelligent interaction of financial data based on AI large-scale model dialogue. Background Technology

[0002] In today's world, where data-driven decision-making is increasingly important, the complexity of financial data and the limitations of traditional interaction methods create a contradiction with users' demands for convenient, intelligent, and efficient data access. Meanwhile, breakthroughs in natural language processing using AI big data models offer unprecedented technological possibilities for interacting with complex systems through more natural and human-like dialogue. Therefore, the "Intelligent Interaction Method for Financial Data Based on AI Big Data Model Dialogue" has emerged, aiming to bridge the professional barriers of financial data and allow users to easily query, analyze, and understand financial data through natural language, just like conversing with a human, thereby improving the efficiency of financial work and the quality of decision-making. The core value of this method lies in transforming complex financial data into easily understood and utilized knowledge, empowering a wider user base. However, the above process still has the following drawbacks: Firstly, existing AI large-scale model dialogue methods for intelligent interaction of financial data do not perform semantic analysis of user input text in a deep enough way, making it difficult to accurately identify logical errors and failing to effectively combine historical dialogue context to understand user intent, which can easily lead to misunderstandings. Secondly, existing AI large-scale model dialogue methods for intelligent interaction of financial data may not be able to accurately identify entities related to financial data after understanding the user's intent, resulting in inaccurate subsequent queries and feedback. Furthermore, after generating a dialogue that understands the user's intent, subsequent operations may be performed directly without user confirmation, which may easily lead to the execution of incorrect tasks due to misunderstanding. Third, existing AI large-scale model dialogue methods for intelligent interaction of financial data may not be able to effectively break down user intent into atomic tasks after receiving user confirmation information, resulting in inaccurate identification of user tasks and affecting subsequent operations. Summary of the Invention

[0003] To overcome the aforementioned deficiencies in the prior art, this invention provides a method for intelligent interaction of financial data based on AI large model dialogue, in order to solve the problems existing in the background art.

[0004] This invention provides the following technical solution: a method for intelligent interaction of financial data based on AI large-scale model dialogue, comprising: S1: Used to receive voice and text information input by the user, convert the voice information into text, parse the input content, and format the parsed text. S2: By performing semantic analysis on the formatted text, mapping the content of the semantic analysis to the financial knowledge graph, identifying and correcting logical errors, and then combining historical dialogue context information, extracting features from the current input text, analyzing and inferring the user's intent; S3: Based on the results of user intent analysis, identify entities related to financial data; S4: By querying the financial knowledge graph for the entities identified that are related to financial data, and combining them with the historical dialogue context, a dialogue that understands the user's intent is generated and sent to the interactive interface. The user is then asked whether the understanding of the intent is correct and is asked to confirm the user's further confirmation. S5: By receiving the user's confirmation information, the intent is broken down into atomic tasks using the intent decomposition model to identify the user's task; S6: Based on the recognition results of the user task, call the corresponding financial data interface to obtain the required financial data, generate financial query information, feed the financial query information back to the user, and complete this interaction.

[0005] Preferably, step S1 acquires the user's voice input via a microphone and obtains the user's text input via an input interface. Simultaneously, it continuously listens for or waits for user input to capture the original voice signal or text string. When voice input is detected, speech recognition is immediately initiated. The captured raw voice signal is preprocessed, and the speech recognition model is used to convert the preprocessed voice signal into the corresponding text string. When a user provides both voice and text input simultaneously or sequentially, the text converted from voice input is merged with the directly input text, and then the merged text is converted into the corresponding text string. The input text is then segmented using a word segmentation tool and labeled with part-of-speech tags. Key entities are extracted using named entity recognition technology. Dependency parsing or constituent parsing is used to identify sentence structure and semantic roles. The parsing results are then mapped to a predefined format and standardized.

[0006] Preferably, step S2 uses an AI big data model to perform semantic understanding on the formatted text, extracts core semantic units, and then obtains the company's financial information and industry standard information to construct a financial knowledge graph. The financial data and rules are mapped to nodes and edges of the knowledge graph, and the semantic units are matched with the nodes in the financial knowledge graph to verify logical relationships and correct any detected logical errors. By extracting key information from historical dialogues, natural language processing techniques are used to generate contextual feature vectors. Then, features are extracted from the current input text and combined with the historical dialogue feature vectors to form context-enhanced feature representations. A classification model is used to classify these context-enhanced feature representations, resulting in classification labels including query, analysis, and prediction categories. Based on the classification results, corresponding intent responses are generated.

[0007] Preferably, step S3, based on the intent response results, uses named entity recognition technology, combined with the contextual understanding capabilities of a large model and a predefined financial entity dictionary, to identify key entities related to financial data from the intent response results.

[0008] Preferably, step S4 uses the identified entities as query conditions, links the entity names to a financial knowledge graph, queries the financial knowledge graph based on entity relationships, and obtains query results from the knowledge graph. The query results include a data value, a data range, a descriptive text, and an information link. It also retrieves and analyzes the historical dialogue history preceding the current dialogue to understand the relationship between the current query and previous discussions. The knowledge graph query results and contextual information are integrated to form a complete understanding of the user's intent. Based on the integrated intent, a natural and fluent dialogue response is generated. The generated dialogue text is then appropriately formatted and sent. The message is sent to the user's current interactive interface. After sending the understood dialogue, a confirmation request is immediately sent to explicitly ask the user whether the system's understanding of their intent is accurate. A multimodal confirmation request containing key entities is generated. This request supports text highlighting, voice restatement, and visual entity annotation. The system then enters a waiting state and does not actively perform any further operations. When entering the waiting state, a timeout management mechanism is activated. If no response is received within 90 seconds, a secondary confirmation or quick option is triggered until a response is received from the user. If the user confirms that their intent is accurately understood, the system will continue to recognize the user's task. If the user confirms that their intent is not accurately understood, the system will prompt the user to re-enter the text.

[0009] Preferably, after receiving a response from the user confirming that the intent is understood accurately, S5 uses a BERT+CRF hybrid model to perform structured parsing of the composite intent. The pre-trained BERT model is used to extract deep semantic features of the text, generate the probability distribution of each entity, and based on the output of BERT, the CRF model is combined to perform sequence labeling of the entities, outputting a labeled composite intent. By using grammatical analysis tools to perform grammatical analysis on the text, the grammatical structure of sentences is determined, and semantic roles are labeled to identify the semantic relationships between each component and the predicate. Based on the grammatical and semantic analysis results, the complex intent is broken down into a set of atomic tasks with dependencies. These atomic tasks include single entity query, multi-entity comparison, temporal analysis, data export, visualization generation, conditional reminders, and multi-round correction. Parameters are attached to each atomic task, and the execution order of tasks is dynamically determined through a weighted scoring mechanism. The results of executing atomic tasks in sequence are monitored. If a task fails, the error is recorded and a differential correction algorithm is triggered. The differential correction algorithm locates the error slots, compares the BERT output probability distribution with the CRF annotation results, identifies low-confidence slots, and only requires the user to correct key parameters, rather than re-entering the entire data. User tasks are identified by integrating model parsing and annotation results, syntactic and semantic role annotation results, atomic task sets and their dependencies, task execution order and execution results.

[0010] Preferably, step S6 matches the corresponding financial data interface according to the identified user task, obtains financial data from the called interface, organizes and formats the obtained financial data, generates complete financial query information according to the user task requirements, and feeds the financial query information back to the user.

[0011] The technical effects and advantages of this invention are as follows: (1) Based on semantic analysis, the content is mapped to the financial knowledge graph to identify and correct logical errors. At the same time, feature extraction and intent inference are performed by combining historical dialogue context, which improves the accuracy of semantic understanding.

[0012] (2) Based on the results of user intent analysis, entities related to financial data are specifically identified, which enhances the targeting of entity identification. After generating a dialogue that understands the user intent, it is sent to the interactive interface to ask the user whether the understanding of the intent is correct and wait for the user to confirm further, thus ensuring the accurate grasp of the user intent.

[0013] (3) Based on the use of the intent decomposition model, the intent is split into atomic tasks to identify user tasks, making task identification more accurate. Attached Figure Description

[0014] Figure 1 This is a diagram illustrating the method steps of the present invention. Detailed Implementation

[0015] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. In addition, the forms of the various structures described in the following embodiments are merely illustrative. The intelligent interaction method for financial data based on AI large model dialogue involved in the present invention is not limited to the structures described in the following embodiments. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] like Figure 1 The embodiment shown provides a method for intelligent interaction of financial data based on AI large model dialogue, including: S1: Used to receive voice and text information input by the user, convert the voice information into text, parse the input content, and format the parsed text.

[0017] In this embodiment, S1 acquires the user's voice input through a microphone and obtains the user's text input through an input interface. At the same time, it continuously listens for or waits for user input to capture the original voice signal or text string. When voice input is detected, speech recognition is immediately initiated. The captured raw voice signal is preprocessed, and the speech recognition model is used to convert the preprocessed voice signal into the corresponding text string. When a user provides both voice and text input simultaneously or sequentially, the text converted from voice input is merged with the directly input text, and then the merged text is converted into the corresponding text string. The input text is then segmented using a word segmentation tool and labeled with part-of-speech tags. Key entities are extracted using named entity recognition technology. Dependency parsing or constituent parsing is used to identify sentence structure and semantic roles. The parsing results are then mapped to a predefined format and standardized.

[0018] S2: By performing semantic analysis on the formatted text, mapping the content of the semantic analysis to the financial knowledge graph, identifying and correcting logical errors, and then combining historical dialogue context information, extracting features from the current input text, analyzing and inferring the user's intent.

[0019] In this embodiment, S2 uses an AI big data model to perform semantic understanding on the formatted text, extracts core semantic units, and then obtains the company's financial information and industry standard information to construct a financial knowledge graph. The financial data and rules are mapped to nodes and edges of the knowledge graph, and the semantic units are matched with the nodes in the financial knowledge graph to verify logical relationships and correct any detected logical errors. By extracting key information from historical dialogues, natural language processing techniques are used to generate contextual feature vectors. Then, features are extracted from the current input text and combined with the historical dialogue feature vectors to form context-enhanced feature representations. A classification model is used to classify these context-enhanced feature representations, resulting in classification labels including query, analysis, and prediction categories. Based on the classification results, corresponding intent responses are generated.

[0020] It's important to clarify that a large AI model is used to semantically analyze the text, extracting core semantic units such as "net profit = 10 million yuan" and "asset-liability ratio = 60%". Semantic relationships are identified, such as the "growth" relationship in "net profit increased by 20%". Through company financial data, rules, and logic, the financial data and rules are mapped to nodes and edges in a knowledge graph. Nodes represent financial indicators (e.g., "net profit", "asset-liability ratio") and industry benchmarks (e.g., "industry average net profit margin"). Edges represent logical relationships (e.g., "net profit margin = net profit / operating revenue") and constraints (e.g., "asset-liability ratio should be lower than 70%)". The extracted semantic units are then matched with nodes in the knowledge graph. For example, "net profit = 10 million yuan". The algorithm matches the "net profit" node in the knowledge graph and verifies whether the semantic unit conforms to the logical relationships in the knowledge graph. For example, it checks whether "net profit = 10 million yuan" is consistent with "operating revenue = 50 million yuan" and "net profit margin = 20%". Based on the rules and constraints in the knowledge graph, it detects contradictions or anomalies in the semantic unit. For example, if "net profit margin = 20%" but "operating revenue = 50 million yuan" and "net profit = 8 million yuan" (800 / 5000 = 16%), there is a contradiction. If "asset-liability ratio = 80%", but the industry benchmark is "asset-liability ratio should be lower than 70%", it is marked as an anomaly. Then, it corrects the detected logical errors. For example, it corrects "net profit = 8 million yuan" to "net profit = 10 million yuan" (matching the node in the knowledge graph). (For example, if the net profit margin is 20%), the user is prompted that "the debt-to-asset ratio of 80% is higher than the industry benchmark of 70%, which may indicate financial risk." Based on the user's historical dialogue records, key information from these dialogues is extracted to generate a contextual feature vector. Features are then extracted from the current input text and combined with the historical dialogue feature vector to form a context-enhanced feature representation. A classification model is used to categorize this representation, resulting in the following tags: query type (users want specific data, such as "What was the net profit in 2023?"), analysis type (users want explanations of data relationships, such as "Why did the net profit margin decline?"), and prediction type (users want trend predictions or scenario simulations, such as "Can net profit grow by 10% next year?"). Based on the classification results, a corresponding response is generated. For example, for query type... Return specific data ("Net profit in 2023 was 10 million yuan"), analysis category Explaining the data relationship ("Net profit margin decreased because costs increased by 15%), predictive data) Provide a forecast ("Based on historical data, there is a 60% probability that net profit will grow by 10% next year").

[0021] S3: Based on the results of user intent analysis, identify entities related to financial data.

[0022] In this embodiment, S3 uses named entity recognition technology, combined with the contextual understanding capability of a large model and a predefined financial entity dictionary, to identify key entities related to financial data from the intent response results based on the intent response results.

[0023] S4: By querying the financial knowledge graph for the entities identified that are related to financial data, and combining them with the historical dialogue context, a dialogue that understands the user's intent is generated and sent to the interactive interface. The user is then asked whether the understanding of the intent is correct and is asked to confirm the user's further understanding.

[0024] In this embodiment, step S4 uses the identified entities as query conditions, links the entity names to a financial knowledge graph, and performs queries based on the relationships between entities. The query results are obtained from the knowledge graph, including a data value, a data range, a descriptive text, and an information link. The system also retrieves and analyzes the historical dialogues preceding the current conversation to understand the relationship between the current query and previous discussions. The knowledge graph query results and contextual information are integrated to form a complete understanding of the user's intent. Based on this integrated intent, a natural and fluent dialogue response is generated. The generated dialogue text is then appropriately formatted. The message is sent to the user's current interactive interface. Immediately after sending the understood dialogue, a confirmation request is sent to explicitly ask the user whether the system's understanding of their intent is accurate. A multimodal confirmation request containing key entities is generated. This request supports text highlighting, voice restatement, and visual entity annotation. The system then enters a waiting state and does not actively perform any further operations. When entering the waiting state, a timeout management mechanism is activated. If no response is received within 90 seconds, a secondary confirmation or quick option is triggered until a response is received from the user. If the user confirms that their intent is accurately understood, the system will continue to perform the user task recognition. If the user confirms that their intent is not accurately understood, the user will be prompted to re-enter the text.

[0025] S5: By receiving confirmation information from the user, the intent is broken down into atomic tasks using the intent decomposition model, and the user's task is identified.

[0026] In this embodiment, after receiving a response from the user confirming that the intention is understood accurately, step S5 uses a BERT+CRF hybrid model to perform structured parsing of the composite intention. The pre-trained BERT model is used to extract deep semantic features of the text, generate the probability distribution of each entity, and based on the output of BERT, the CRF model is combined to perform sequence labeling of the entities, outputting the labeled composite intention. By using grammatical analysis tools to perform grammatical analysis on the text, the grammatical structure of sentences is determined, and semantic roles are labeled to identify the semantic relationships between each component and the predicate. Based on the grammatical and semantic analysis results, the complex intent is broken down into a set of atomic tasks with dependencies. These atomic tasks include single entity query, multi-entity comparison, temporal analysis, data export, visualization generation, conditional reminders, and multi-round correction. Parameters are attached to each atomic task, and the execution order of tasks is dynamically determined through a weighted scoring mechanism. The results of executing atomic tasks in sequence are monitored. If a task fails, the error is recorded and a differential correction algorithm is triggered. The differential correction algorithm locates the error slots, compares the BERT output probability distribution with the CRF annotation results, identifies low-confidence slots, and only requires the user to correct key parameters, rather than re-entering the entire data. User tasks are identified by integrating model parsing and annotation results, syntactic and semantic role annotation results, atomic task sets and their dependencies, task execution order and execution results.

[0027] It's important to clarify that after using the BERT+CRF hybrid model to perform structured parsing of the user input text, a labeled composite intent is obtained. For example, for the user input "Query XX's net profit in the third quarter of 2023", the model parses it as follows: "XX" is the company entity, "third quarter of 2023" is the time entity, and "net profit" is the financial data entity. These entities and their labels are organized to form an entity-label mapping table for subsequent analysis. Syntactic analysis tools are used to determine the sentence's grammatical structure, such as subject, verb, and object. In the example above, syntactic analysis shows that "query" is the verb, and "XX's net profit in the third quarter of 2023" is the object. Simultaneously, semantic role labeling identifies "XX" as the object-related entity of the action "query", "third quarter of 2023" as the time-limited component, and "net profit" as the specific financial indicator being queried. Combining the syntactic and semantic analysis results with the model's parsing results further clarifies the core content and intent of the user's query. Based on the syntactic and semantic analysis results, the composite intent is broken down into a set of atomic tasks with dependencies. The task "Query XX's net profit for the third quarter of 2023" can be broken down into a single-entity query atomic task, which queries XX's net profit data for the third quarter of 2023. If the user input is more complex, such as "First query the revenue of XX and YY, then compare their revenue figures," it can be broken down into two single-entity query atomic tasks (querying XX's revenue and querying YY's revenue) and one multi-entity comparison atomic task (comparing the revenues of XX and YY). The single-entity query task is explicitly defined as a prerequisite for the multi-entity comparison task. A weighted scoring mechanism dynamically determines the task execution order. Factors such as task urgency and dependencies are considered; for example, urgent queries are prioritized. The execution results are monitored during the execution of the atomic tasks. If the single-entity query task successfully retrieves the revenue data of Tencent and Alibaba, the multi-entity comparison task continues. If the query task fails, error information is recorded, such as data interface call failure or data non-existence. Combining the entity-object mapping table, the results of syntax and semantic analysis, the set of atomic tasks and their dependencies, the task execution order and results, it is determined whether the user task is to perform a single-entity financial data query or a multi-entity comparison analysis. For example, based on the previous analysis, it is clear that the user task is to query Tencent's net profit data for a specific time period.

[0028] S6: Based on the recognition results of the user task, call the corresponding financial data interface to obtain the required financial data, generate financial query information, feed the financial query information back to the user, and complete this interaction.

[0029] In this embodiment, step S6 matches the corresponding financial data interface according to the identified user task, obtains financial data from the called interface, organizes and formats the obtained financial data, generates complete financial query information according to the user task requirements, and feeds the financial query information back to the user.

[0030] It should be noted that the corresponding financial data interface is matched based on the identified user task. For example, to query the net profit of XX in the third quarter of 2023, the company's financial indicator query interface is called, and the parameters "Company Name=XX", "Time=Third Quarter of 2023" and "Financial Indicator=Net Profit" are passed in.

[0031] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0032] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. A method for intelligent interaction of financial data based on AI large-scale model dialogue, characterized in that, include: S1: Used to receive voice and text information input by the user, convert the voice information into text, parse the input content, and format the parsed text. S2: By performing semantic analysis on the formatted text, mapping the content of the semantic analysis to the financial knowledge graph, identifying and correcting logical errors, and then combining historical dialogue context information, extracting features from the current input text, analyzing and inferring the user's intent; S3: Based on the results of user intent analysis, identify entities related to financial data; S4: By querying the financial knowledge graph for the entities identified that are related to financial data, and combining them with the historical dialogue context, a dialogue that understands the user's intent is generated and sent to the interactive interface. The user is then asked whether the understanding of the intent is correct and is asked to confirm the user's further confirmation. S5: By receiving the user's confirmation information, the intent is broken down into atomic tasks using the intent decomposition model to identify the user's task; S6: Based on the recognition results of the user task, call the corresponding financial data interface to obtain the required financial data, generate financial query information, feed the financial query information back to the user, and complete this interaction.

2. The method for intelligent interaction of financial data based on AI large-scale model dialogue according to claim 1, characterized in that, The S1 acquires the user's voice input via a microphone and obtains the user's text input via an input interface. Simultaneously, it continuously listens for or waits for user input to capture the original voice signal or text string. When voice input is detected, speech recognition is immediately initiated. The captured raw voice signal is preprocessed, and the speech recognition model is used to convert the preprocessed voice signal into the corresponding text string. When a user provides both voice and text input simultaneously or sequentially, the text converted from voice input is merged with the directly input text, and then the merged text is converted into the corresponding text string. The input text is then segmented using a word segmentation tool and labeled with part-of-speech tags. Key entities are extracted using named entity recognition technology. Dependency parsing or constituent parsing is used to identify sentence structure and semantic roles. The parsing results are then mapped to a predefined format and standardized.

3. The method for intelligent interaction of financial data based on AI large-scale model dialogue according to claim 2, characterized in that, The S2 uses a large AI model to perform semantic understanding on formatted text, extracts core semantic units, and then obtains the company's financial information and industry standard information to construct a financial knowledge graph. It maps financial data and rules to nodes and edges of the knowledge graph, matches semantic units with nodes in the financial knowledge graph, verifies logical relationships, and corrects detected logical errors. By extracting key information from historical dialogues, natural language processing techniques are used to generate contextual feature vectors. Then, features are extracted from the current input text and combined with the historical dialogue feature vectors to form context-enhanced feature representations. A classification model is used to classify these context-enhanced feature representations, resulting in classification labels including query, analysis, and prediction categories. Based on the classification results, corresponding intent responses are generated.

4. The intelligent interaction method for financial data based on AI large-scale model dialogue according to claim 3, characterized in that, Based on the intent response results, S3 uses named entity recognition technology, combined with the contextual understanding capabilities of a large model and a predefined financial entity dictionary, to identify key entities related to financial data from the intent response results.

5. The intelligent interaction method for financial data based on AI large-scale model dialogue according to claim 4, characterized in that, Step S4 uses the identified entities as query conditions, links the entity names to a financial knowledge graph, and performs queries based on the relationships between entities. The query results are retrieved from the knowledge graph, including a data value, a data range, a descriptive text, and an information link. The system also retrieves and analyzes the historical dialogues preceding the current conversation to understand the relationship between the current query and previous discussions. The knowledge graph query results and contextual information are integrated to form a complete understanding of the user's intent. Based on this integrated intent, a natural and fluent dialogue response is generated. The generated dialogue text is then appropriately formatted and sent to [the relevant platform / organization]. The user interface, after sending a dialogue indicating understanding, immediately sends a confirmation request to explicitly ask the user whether the system's understanding of their intent is accurate. A multimodal confirmation request containing key entities is generated. This request supports text highlighting, voice restatement, and visual entity annotation. The system then enters a waiting state, refraining from further action. A timeout management mechanism is activated during this waiting state. If no response is received within 90 seconds, a secondary confirmation or quick option is triggered until a response is received from the user. If the user confirms accurate understanding of the intent, the system continues to recognize the user's task. If the user confirms inaccurate understanding, the system prompts the user to re-enter the text.

6. The method for intelligent interaction of financial data based on AI large-scale model dialogue according to claim 5, characterized in that, When S5 receives a response from the user confirming that the intention is understood accurately, it uses a BERT+CRF hybrid model to perform structured parsing of the composite intention. It uses a pre-trained BERT model to extract deep semantic features of the text, generates the probability distribution of each entity, and uses the output of BERT combined with the CRF model to perform sequence labeling of entities, outputting a labeled composite intention. By using grammatical analysis tools to perform grammatical analysis on the text, the grammatical structure of sentences is determined, and semantic roles are labeled to identify the semantic relationships between each component and the predicate. Based on the grammatical and semantic analysis results, the complex intent is broken down into a set of atomic tasks with dependencies. These atomic tasks include single entity query, multi-entity comparison, temporal analysis, data export, visualization generation, conditional reminders, and multi-round correction. Parameters are attached to each atomic task, and the execution order of tasks is dynamically determined through a weighted scoring mechanism. The results of executing atomic tasks in sequence are monitored. If a task fails, the error is recorded and a differential correction algorithm is triggered. The differential correction algorithm locates the error slots, compares the BERT output probability distribution with the CRF annotation results, identifies low-confidence slots, and only requires the user to correct key parameters, rather than re-entering the entire data. User tasks are identified by integrating model parsing and annotation results, syntactic and semantic role annotation results, atomic task sets and their dependencies, task execution order and execution results.

7. The intelligent interaction method for financial data based on AI large-scale model dialogue according to claim 6, characterized in that, S6 matches the corresponding financial data interface according to the identified user task, obtains financial data from the called interface, organizes and formats the obtained financial data, generates complete financial query information according to the user task requirements, and feeds the financial query information back to the user.

Images