Intelligent multi-round question and answer system based on large model

Abstract

The invention relates to the technical field of artificial intelligence natural language processing, and provides an intelligent multi-round question-answering system based on a large model, comprising a semantic understanding module for generating a semantic understanding state value by extracting a semantic feature vector input by a user and analyzing context association; the context association module is used for extracting semantic features and time sequences of dialogue history based on the state value and generating a context association parameter set; the intention recognition module is used for analyzing the relationship between the user intention and the context according to the context parameter set and outputting an intention recognition parameter set; the answer generation module is used for generating multiple rounds of answer generation values in combination with the knowledge base and the real-time semantic data; the semantic prediction module is used for predicting a semantic trend through a large model, analyzing context change and outputting a semantic prediction value; and the dialogue feedback module is used for performing error analysis according to the semantic predicted value, optimizing answer generation and finally outputting a multi-round dialogue automatic optimization scheme. According to the invention, the accuracy, coherence and adaptive ability of multiple rounds of dialogues can be improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence natural language processing technology, and more specifically, to an intelligent multi-round question-answering system based on a large model. Background Art

[0002] With the continuous development of artificial intelligence technology, intelligent question-answering systems have been widely used in various fields. Existing intelligent question-answering systems are mostly based on rule engines or simple machine learning models, which can achieve a preliminary understanding and response to user questions. These systems identify user intent through preset rules or pattern matching and retrieve corresponding answers from a knowledge base. While this has improved information retrieval efficiency to a certain extent, existing question-answering systems have significant limitations in multi-round dialogue scenarios.

[0003] First, existing question-answering systems are weak in semantic understanding. They can typically only process simple, structured language input and are unable to understand complex natural language text, especially multi-turn conversations involving contextual relationships. This results in the system being unable to accurately grasp the user's true intent, affecting the accuracy and relevance of answers.

[0004] Second, existing question-answering systems have shortcomings in context management. In multi-turn conversations, contextual information is crucial for understanding user intent. However, existing systems often fail to effectively store and utilize conversation history, resulting in information gaps during the conversation and preventing coherent multi-turn dialogue.

[0005] Furthermore, existing question-answering systems lack flexibility in answer generation. They typically rely on pre-set templates or fixed answer libraries and are unable to dynamically generate responses based on real-time semantic data. This makes it difficult for the systems to provide personalized and accurate responses to diverse user questions.

[0006] Furthermore, existing question-answering systems are almost completely lacking in semantic prediction. They are unable to predict and infer the semantic trends of conversations, making it impossible to adjust conversation strategies in advance to better guide the conversation process.

[0007] During the implementation of the embodiments of the present invention, the present applicant discovered at least the following problems or defects in the prior art: existing question-answering systems cannot accurately understand complex natural language text in multi-round dialogue scenarios, cannot effectively manage contextual information, lack a flexible answer generation mechanism, and are unable to perform semantic prediction. These problems result in the existing systems performing less than ideally in multi-round dialogues and failing to meet the high expectations users place on intelligent question-answering systems. Summary of the Invention

[0008] The present invention provides an intelligent multi-round question-answering system based on a large model, comprising: a semantic understanding module, which extracts semantic feature vectors based on natural language text input by a user, analyzes context associations, calculates semantic similarities, integrates them into a semantic feature parameter set, and obtains a semantic understanding state value; a context association module, which extracts semantic features and time series of conversation history based on the semantic understanding state value, generates context association data, screens the optimal context association combination, and obtains a context association parameter set; an intention recognition module, which extracts semantic features of user intention based on the context association parameter set, analyzes the relationship between intention and context, matches intention and context combinations, and obtains an intention recognition parameter set; and an answer generation module, which generates an answer based on the intention recognition parameter set. Extract relevant knowledge from the knowledge base, combine it with real-time semantic data, match knowledge and semantics, generate answers and apply the answers to multi-round dialogues to obtain multi-round answer generation values; the semantic prediction module, based on the multi-round answer generation values, captures the semantic data of the dialogue sampling points, combines the large model to predict and infer semantic changes, analyzes the context changes corresponding to the semantic trends, classifies and organizes the semantic change trends according to the inference results, combines the semantic change information, and performs semantic adjustment and analysis on the classified data to obtain semantic prediction values; the dialogue feedback module, based on the semantic prediction values, uses real-time semantic data and context association data to analyze the error values ​​between semantics and context, adjusts answer generation based on the error values, and obtains an automatic optimization solution for multi-round dialogues.

[0009] Furthermore, the semantic understanding state value includes a semantic feature parameter set, a context association parameter set, and a semantic similarity parameter set; the context association parameter set includes a time series parameter and a semantic feature matching parameter; the intention recognition parameter set includes an intention semantic feature parameter and a context matching parameter; the multi-round answer generation value includes a knowledge matching parameter, a semantic generation parameter, and an answer generation parameter; the semantic prediction value includes a semantic trend analysis parameter and a context relationship parameter; and the multi-round dialogue automatic optimization scheme includes an error analysis parameter and an answer adjustment parameter.

[0010] Furthermore, the semantic understanding module includes: a data extraction submodule that performs vectorization processing of semantic features based on the natural language text input by the user, locates invalid semantics, removes noise data, arranges the extracted semantic features in time series, and generates a semantic feature data set; a similarity calculation submodule that analyzes the similarity between semantic features based on the semantic feature data set, calculates the vector inner product of the semantic features, sorts the similarity values ​​by weight, marks areas with excessive semantic differences, and obtains semantic similarity data; a semantic integration submodule that calls the similarity values ​​for multi-dimensional aggregation based on the semantic similarity data, screens semantic feature differences, classifies them according to the size of similarity, and arranges the semantic features in order to generate a semantic understanding status value.

[0011] Furthermore, the context association module includes: a parameter extraction submodule that identifies the semantic features of the conversation history based on the semantic understanding state value, records the time series and semantic features of the context, standardizes the recorded data, organizes the standardized data and classifies it according to the time series and semantic features, and generates context association data; a context optimization submodule that analyzes the time series and semantic features in the data set according to the context association data, screens parameter combinations with high matching degrees with the semantic state, adjusts the parameter combinations through pattern matching and records the matching results, and generates parameter combination optimization results; a parameter selection submodule retrieves the parameter combination optimization results, determines the time series and semantic feature combination with the best matching degree, adjusts the context association parameters, inputs the control configuration, verifies the stability of the parameter set, and generates the context association parameter set.

[0012] Furthermore, the intention recognition module includes: an intention analysis submodule, based on the context-associated parameter set, collects key data through semantic monitoring, including keywords, semantic features and context time series input by the user, performs time series analysis on the data, eliminates outliers, and partitions the remaining data to obtain intention analysis data; an intention matching submodule analyzes the relationship between intention and context through the intention analysis data, calculates the degree of influence of each intention change on parameter adjustment, determines the optimal matching parameter setting based on the impact score, cyclically adjusts the parameters to capture the optimal combination, and obtains the parameter docking result; an intention integration submodule selects the time series and context combination that matches the current intention from the parameter docking result, performs parameter adjustment tests, optimizes the parameter settings through multiple adjustments and verifications, determines and solidifies the parameters as operating standards, and generates an intention recognition parameter set.

[0013] Furthermore, the answer generation module includes: a knowledge extraction submodule, based on the intention recognition parameter set, locates relevant knowledge in the knowledge base, extracts semantic feature values ​​in the knowledge base, continuously records the semantic change rate of the knowledge, extracts multiple key change nodes corresponding to the change rate, sorts the node values ​​in order, and obtains the current knowledge semantic feature value; a semantic matching submodule, based on the current knowledge semantic feature value, analyzes the node change value and real-time semantic data, and calibrates according to a predetermined matching criterion, calls the matching criterion to perform distribution redistribution within the semantic interval, and obtains a semantic matching structure; the answer generation submodule, based on the semantic matching structure, adopts a semantic generation dynamic adjustment method, measures the distribution of semantics between context change nodes, sets the upper and lower limits of the node threshold, applies the threshold to multiple rounds of dialogue, and distributes it to obtain multiple rounds of answer generation values.

[0014] Furthermore, the formula of the semantic generation dynamic adjustment method is as follows: ;in, Indicates the semantically generated value of the matching context change node; Represents the real-time semantic data calculated by the current node; Represents the adjusted semantic feature value of the previous node; Represents the dynamic adjustment weight coefficient, the value range is ; The lower threshold set for the node is used to control the minimum semantic generation amount.

[0015] Furthermore, the semantic prediction module includes: a semantic data capture submodule that generates values ​​based on the multiple rounds of answers, applies a large model algorithm, captures the semantic data of the sampling points, eliminates outliers and corrects errors, stores them in layers by intervals, performs semantic processing, and generates a semantic data set; a context analysis submodule that divides the intervals according to context-related data based on the semantic data set, extracts change trends and fluctuation features, and generates a context and semantic change feature set; a semantic distribution inference submodule that adjusts feature parameters and calibrates trend data based on the context and semantic change feature set, extracts distribution intervals, and performs numerical prediction to obtain semantic prediction values.

[0016] Furthermore, the formula of the large model algorithm is as follows: ;in, Represents the semantic feature vector; Representative The weight of each data point; Representative The original semantic value of the sampling points; Representative Contextual data of each sampling point; Representative Time series value of sampling points; and Represent the weight coefficients of contextual data and time series respectively; Represents the total number of sampling points.

[0017] Furthermore, the dialogue feedback module includes: an error analysis submodule extracts real-time semantic data and context-related data based on the semantic prediction value, analyzes the real-time semantic value and the prediction value, corresponds the semantic difference value with the current context-related data, and generates a semantic context error value; a parameter adjustment submodule sets the adjustment parameters for answer generation based on the semantic context error value, sets the adjustment range for areas with large errors, fine-tunes the low-error areas, screens matching parameter sets by comparing the adjustment effects, and integrates them to generate an answer adjustment parameter set; a feedback control submodule applies the adjustment parameters to each dialogue node based on the answer adjustment parameter set, implements the answer generation operation item by item, synchronously monitors the semantic and context-related data, gradually adjusts the answer generation order of each node, and generates an automatic optimization plan for multiple rounds of dialogue.

[0018] According to the above-mentioned embodiment of the present invention, there are at least the following beneficial effects: the intelligent multi-round question-answering system of the present invention can improve the accuracy and depth of semantic understanding. Through the semantic understanding module, the system can vectorize the natural language text input by the user, extract semantic feature vectors, analyze context associations, and calculate semantic similarity. This multi-dimensional semantic analysis method enables the system to grasp the user's true intentions more accurately, and even in complex multi-round dialogue scenarios, it can effectively understand each question of the user, thereby providing the user with more accurate and relevant information. At the same time, the context association module can effectively manage and utilize dialogue history information, generate context association parameter sets, and ensure that in multi-round dialogues, the system can continuously track and understand changes in context, and achieve a coherent and smooth dialogue experience.

[0019] In addition, the system of the present invention can dynamically generate answers and perform semantic predictions. The answer generation module can extract relevant knowledge from the knowledge base based on the intent recognition parameter set, and match knowledge with semantics in combination with real-time semantic data to generate answers. This dynamic generation mechanism enables the system to flexibly adjust the answer content according to the real-time changes in the conversation and provide personalized answers. The semantic prediction module can capture the semantic data of the conversation sampling points, combine the large model to predict and infer the semantic changes, and analyze the context changes corresponding to the semantic trends. In this way, the system can predict the user's intentions and the direction of the conversation, thereby better guiding the conversation process and improving the efficiency and quality of the conversation. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily apparent by reading the following detailed description with reference to the accompanying drawings, in which several embodiments of the present invention are shown by way of example and not limitation, in which: Figure 1 A schematic diagram of the structure of an intelligent multi-round question-answering system based on a large model provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0021] The principles and spirit of the present invention will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided solely to enable those skilled in the art to better understand and implement the present invention, and are not intended to limit the scope of the present invention in any way. Rather, these embodiments are provided to make the present invention more thorough and complete, and to fully convey the scope of the present invention to those skilled in the art.

[0022] Those skilled in the art will appreciate that the embodiments of the present invention may be implemented as a system, apparatus, device, method, or computer program product. Therefore, the present invention may be implemented in the following forms: entirely in hardware, entirely in software (including firmware, resident software, microcode, etc.), or in a combination of hardware and software.

[0023] It should be noted that any number of elements in the drawings is for illustration only and not for limitation, and any naming is only for distinction and does not have any limiting meaning.

[0024] Reference below Figure 1 , Figure 1 This is a schematic diagram of the structure of an intelligent multi-round question-answering system based on a large model provided by an embodiment of the present invention. Figure 1 As shown, an intelligent multi-round question-answering system based on a large model includes: a semantic understanding module 101, which extracts semantic feature vectors based on the natural language text input by the user, analyzes context associations, calculates semantic similarity, integrates them into a semantic feature parameter set, and obtains a semantic understanding state value; a context association module 102, which extracts semantic features and time series of the conversation history based on the semantic understanding state value, generates context association data, screens the optimal context association combination, and obtains a context association parameter set; an intention recognition module 103, which extracts semantic features of the user's intention based on the context association parameter set, analyzes the relationship between the intention and the context, matches the intention and the context combination, and obtains an intention recognition parameter set; an answer generation module 104, which generates semantic features and time series of the conversation history based on the semantic understanding state value, generates context association data, screens the optimal context association combination, and obtains a context association parameter set; an intention recognition module 103, which extracts semantic features of the user's intention based on the context association parameter set, analyzes the relationship between the intention and the context, matches the intention and the context combination, and obtains an intention recognition parameter set; and an answer generation module 104, which generates semantic features and time series of the conversation history based on the semantic understanding state value, generates context association data, screens the optimal context association combination, and obtains a context association parameter set. , extract relevant knowledge from the knowledge base, combine it with real-time semantic data, match knowledge and semantics, generate answers and apply the answers to multi-round dialogues to obtain multi-round answer generation values; the semantic prediction module 105, based on the multi-round answer generation values, captures the semantic data of the dialogue sampling points, combines the large model to predict and infer semantic changes, analyzes the context changes corresponding to the semantic trends, classifies and organizes the semantic change trends according to the inference results, combines the semantic change information, and performs semantic adjustment and analysis on the classified data to obtain semantic prediction values; the dialogue feedback module 106, based on the semantic prediction values, analyzes the error values ​​of semantics and context through real-time semantic data and context association data, adjusts the answer generation in combination with the error values, and obtains an automatic optimization solution for multi-round dialogues.

[0025] The intelligent multi-round question-answering system of the present invention implements in-depth semantic analysis of natural language text input by users through a semantic understanding module. The semantic understanding module extracts semantic feature vectors, analyzes context associations, calculates semantic similarity, and finally integrates them into a semantic feature parameter set to obtain a semantic understanding state value. Semantic feature vectors refer to the conversion of semantic information in a text into a numerical vector form that can be processed by a computer, which is used to represent the semantic content of the text. Context association refers to the logical and semantic connection between the current statement and the previous statement in a multi-round dialogue. Semantic similarity is a quantitative indicator that measures the degree of similarity between different semantic feature vectors. Through these steps, the system can accurately grasp the semantic information input by the user and provide a basis for subsequent dialogue processing.

[0026] Specifically, the data extraction submodule in the semantic understanding module is used to process the natural language text input by the user. By vectorizing the semantic features, locating invalid semantics and removing noise data, the extracted semantic features are then arranged in time series to generate a semantic feature dataset. The similarity calculation submodule analyzes the similarity between semantic features based on the semantic feature dataset, calculates the inner product of the vectors of the semantic features, and marks areas with excessive semantic differences by weight sorting to obtain semantic similarity data. The semantic integration submodule further calls the similarity value for multi-dimensional aggregation, screens the differences in semantic features, classifies them by similarity size, and arranges them in order, and finally generates a semantic understanding status value. For example, the semantic feature vector can be generated by a pre-trained language model, the time series parameter can be the timestamp of each sentence in the conversation, and the semantic similarity parameter is a value obtained by calculating the cosine similarity between vectors and other methods.

[0027] Preferably, the data extraction submodule in the semantic understanding module can adopt a deep learning model, such as a pre-trained model of the Transformer architecture, input the natural language text input by the user, and output the corresponding semantic feature vector. In the similarity calculation submodule, when calculating the inner product of the semantic feature vector, a weight factor can be introduced to assign different weights according to the importance of the semantic features to more accurately reflect the semantic similarity. For example, a higher weight is given to keywords or core semantic features. In the semantic integration submodule, a threshold can be set. When the semantic similarity is lower than the threshold, it is considered that the difference between the two semantic features is too large and further analysis is required. Through these refined operating steps and parameter settings, the semantic understanding module can process the natural language text input by the user more accurately, providing a high-quality semantic understanding foundation for the entire intelligent multi-round question-answering system.

[0028] In some embodiments, the semantic understanding state value includes a semantic feature parameter set, a context association parameter set, and a semantic similarity parameter set. The context association parameter set includes a time series parameter and a semantic feature matching parameter. The intention recognition parameter set includes an intention semantic feature parameter and a context matching parameter. The multi-round answer generation value includes a knowledge matching parameter, a semantic generation parameter, and an answer generation parameter. The semantic prediction value includes a semantic trend analysis parameter and a context relationship parameter. The multi-round dialogue automatic optimization scheme includes an error analysis parameter and an answer adjustment parameter.

[0029] It should be noted that the semantic feature parameter set, context association parameter set, semantic similarity parameter set, intent semantic feature parameters, context matching parameters, knowledge matching parameters, semantic generation parameters, answer generation parameters, semantic trend analysis parameters, context relationship parameters, error analysis parameters, and answer adjustment parameters correspond to the outputs and inputs of different modules in the system, ensuring data interaction and collaborative operation between modules. For example, the semantic feature parameter set is the output of the semantic understanding module, used to describe the semantic features of the user input text; the context association parameter set is the output of the context association module, used to represent the semantic and time series information in the conversation history. By defining and using these parameter sets, the system can achieve full process optimization from semantic understanding to answer generation.

[0030] Specifically, the semantic feature parameter set refers to the collection of semantic features extracted from user input text. These features can include keywords, phrases, semantic vectors, and other features, and are used to describe the core semantic content of the text. The context association parameter set includes time series parameters and semantic feature matching parameters. The time series parameters record the chronological order of each statement in the conversation, while the semantic feature matching parameters indicate the semantic similarity between the current statement and previous statements. The intention semantic feature parameters refer to the semantic features of the user's intention and are used to determine the user's intention category in the intention identification module. The context matching parameters measure the degree of correlation between the user's intention and the context. The knowledge matching parameters, semantic generation parameters, and answer generation parameters are used in the knowledge extraction, semantic matching, and answer generation steps of the answer generation module, respectively, to ensure that the generated answers are closely related to the user's intention and context. The semantic trend analysis parameters and context relationship parameters are used in the semantic prediction module to analyze the semantic change trends and dynamic relationships of the context in the conversation. The error analysis parameters and answer adjustment parameters are used in the conversation feedback module to adjust the answer generation strategy and optimize the conversation effect by analyzing the error between the semantic prediction value and the actual semantic value. The specific settings of these parameters can be adjusted according to the dialogue scenario and user needs. For example, the semantic similarity parameter can be set by calculating the cosine similarity between semantic vectors, and the error analysis parameter can be set by calculating the mean square error between the predicted value and the actual value.

[0031] Preferably, to generate the semantic feature parameter set, a pre-trained language model, such as BERT or the GPT series, can be used to encode the user input text and extract semantic vectors as semantic feature parameters. Within the context-related parameter set, the time series parameter can be specifically set to the timestamp of each statement in the conversation, and the semantic feature matching parameter can be determined by calculating the similarity between the semantic vectors of the current statement and previous statements. To extract the intent semantic feature parameters, a classification algorithm, such as a support vector machine or a convolutional neural network in deep learning, can be used to classify the semantic features of the user input and determine the user's intent category. Context matching parameters can be set by calculating the similarity between the user's intent and the contextual semantic features, for example, by calculating the dot product or cosine similarity between the intent semantic feature vector and the contextual semantic feature vector. For knowledge matching parameters, the knowledge base can be searched for the knowledge items most relevant to the user's intent and their semantic features extracted as knowledge matching parameters. Semantic generation parameters and answer generation parameters can be optimized by defining generation model parameters, such as the hidden layer dimension and learning rate. Semantic trend analysis parameters can be used to predict future semantic trends by analyzing semantic change trends in conversation history, such as through time series analysis. Context relationship parameters can be set by constructing a context relationship graph and analyzing the logical and semantic connections between contexts. Error analysis parameters can be used to evaluate prediction accuracy by calculating the difference between semantic predictions and actual semantic values, such as mean squared error or absolute error. Answer adjustment parameters can be used to adjust the parameters of the answer generation model based on the results of the error analysis parameters, such as adjusting the weights or biases of the generation model, to optimize answer quality. Through these specific parameter settings and optimization methods, the system can more effectively handle multi-turn conversations and improve conversation accuracy and coherence.

[0032] In some embodiments, the semantic understanding module includes: a data extraction submodule that performs vectorization processing of semantic features based on the natural language text input by the user, locates invalid semantics, removes noise data, arranges the extracted semantic features in time series, and generates a semantic feature data set; a similarity calculation submodule that analyzes the similarity between semantic features based on the semantic feature data set, calculates the vector inner product of the semantic features, sorts the similarity values ​​by weight, marks areas with excessive semantic differences, and obtains semantic similarity data; a semantic integration submodule that calls the similarity values ​​for multi-dimensional aggregation based on the semantic similarity data, screens semantic feature differences, classifies them according to the similarity size, and arranges the semantic features in order to generate a semantic understanding status value.

[0033] It should be noted that the semantic understanding module processes the natural language text input by the user, extracts semantic feature vectors, analyzes contextual associations, calculates semantic similarity, and ultimately generates a semantic understanding status value. Semantic feature vectors convert the semantic information in the text into a computer-processable numerical vector form, representing the semantic content of the text. Contextual association refers to the logical and semantic connection between the current statement and previous statements in multiple rounds of dialogue. Semantic similarity is a quantitative indicator that measures the degree of similarity between different semantic feature vectors. Through these steps, the system can accurately grasp the semantic information of the user input, providing a foundation for subsequent dialogue processing.

[0034] Specifically, the semantic understanding module includes three submodules: a data extraction submodule, a similarity calculation submodule, and a semantic integration submodule. The data extraction submodule is used to process the natural language text input by the user, perform vectorization processing of semantic features, locate invalid semantics and remove noise data, and then arrange the extracted semantic features in time series to generate a semantic feature dataset. The similarity calculation submodule analyzes the similarity between semantic features based on the semantic feature dataset, calculates the inner product of the semantic feature vectors, and marks areas with excessive semantic differences by weight sorting to obtain semantic similarity data. The semantic integration submodule further uses similarity values ​​for multi-dimensional aggregation, screens semantic feature differences, classifies and arranges semantic features in order by similarity, and finally generates a semantic understanding status value. For example, the semantic feature vector can be generated by a pre-trained language model, the time series parameter can be the timestamp of each sentence in the conversation, and the semantic similarity parameter is a value obtained by calculating the cosine similarity between vectors and other methods.

[0035] Preferably, the data extraction submodule can adopt a deep learning model, such as a pre-trained model of the Transformer architecture, input the natural language text input by the user, and output the corresponding semantic feature vector. In the similarity calculation submodule, when calculating the inner product of the semantic feature vector, a weight factor can be introduced to assign different weights according to the importance of the semantic features to more accurately reflect the semantic similarity. For example, a higher weight is given to keywords or core semantic features. In the semantic integration submodule, a threshold can be set. When the semantic similarity is lower than the threshold, it is considered that the difference between the two semantic features is too large and further analysis is required. Through these refined operating steps and parameter settings, the semantic understanding module can process the natural language text input by the user more accurately, providing a high-quality semantic understanding foundation for the entire intelligent multi-round question-answering system.

[0036] In some embodiments, the context association module includes: a parameter extraction submodule that identifies the semantic features of the conversation history based on the semantic understanding state value, records the time series and semantic features of the context, standardizes the recorded data, organizes the standardized data and classifies it according to the time series and semantic features, and generates context association data; a context optimization submodule that analyzes the time series and semantic features in the data set according to the context association data, screens parameter combinations with high matching degrees with the semantic state, adjusts the parameter combinations through pattern matching and records the matching results, and generates parameter combination optimization results; a parameter selection submodule retrieves the parameter combination optimization results, determines the time series and semantic feature combination with the best matching degree, adjusts the context association parameters, inputs the control configuration, verifies the stability of the parameter set, and generates the context association parameter set.

[0037] It's important to note that the context association module is a key component of intelligent multi-turn question-answering systems for processing and optimizing conversation history information. Its primary function is to extract semantic features and time series information from the conversation history based on semantic understanding state values, generate context association data, and select the optimal context association combination to ultimately obtain a context association parameter set. This context association parameter set includes time series parameters and semantic feature matching parameters. These parameters describe the semantic and temporal information in the conversation history, ensuring that the system maintains contextual coherence and consistency across multiple rounds of conversation.

[0038] Specifically, the context association module consists of three submodules: a parameter extraction submodule, a context optimization submodule, and a parameter selection submodule. The parameter extraction submodule, based on the semantic understanding state value, identifies semantic features in the conversation history, records the context's time series and semantic features, and normalizes this data, organizing it into data categorized by time series and semantic features to generate context association data. The context optimization submodule, based on the context association data, analyzes the relationship between time series and semantic features in the dataset, selects parameter combinations that highly match the semantic state, performs pattern matching, records the matching results, adjusts the parameter combinations, and generates optimized parameter combination results. The parameter selection submodule retrieves the optimized parameter combination results, determines the optimal matching time series and semantic feature combination, adjusts the context association parameters, inputs control configurations, verifies the stability of the parameter set, and ultimately generates the context association parameter set. The time series parameters can be the timestamps of each statement in the conversation, and the semantic feature matching parameters can be values ​​calculated by calculating the semantic similarity between the current statement and previous statements.

[0039] Preferably, the parameter extraction submodule can use a time series analysis method to timestamp each statement in the conversation history and extract a semantic feature vector. The context optimization submodule can perform cluster analysis on the context-related data through a clustering algorithm, such as K-means or DBSCAN, to screen out the parameter combination that best matches the current semantic state. The parameter selection submodule can verify the stability and effectiveness of the parameter combination through a cross-validation method. For example, multiple different time windows can be set, and the semantic feature matching degree within each time window can be calculated separately, and the parameters corresponding to the time window with the highest matching degree can be selected as the optimal parameters. Through these refined operating steps and parameter settings, the context association module can more effectively process the conversation history information and ensure the coherence and consistency of multiple rounds of conversations.

[0040] In some embodiments, the intention recognition module includes: an intention analysis submodule, based on the context-associated parameter set, collects key data through semantic monitoring, including keywords, semantic features and context time series input by the user, performs time series analysis on the data, eliminates outliers, and partitions the remaining data to obtain intention analysis data; an intention matching submodule analyzes the relationship between intention and context through the intention analysis data, calculates the degree of influence of each intention change on parameter adjustment, determines the optimal matching parameter setting based on the impact score, cyclically adjusts the parameters to capture the optimal combination, and obtains the parameter docking result; an intention integration submodule selects the time series and context combination that matches the current intention from the parameter docking result, performs parameter adjustment tests, optimizes the parameter settings through multiple adjustments and verifications, determines and solidifies the parameters as operating standards, and generates an intention recognition parameter set.

[0041] It's important to note that the intent recognition module is a key component of intelligent multi-round question-answering systems for determining user intent. Its function is to extract the semantic features of user intent based on a context-related parameter set, analyze the relationship between intent and context, match the intent and context combinations, and ultimately generate an intent recognition parameter set. The intent recognition parameter set includes intent semantic feature parameters and context matching parameters. These parameters describe the semantic features of user intent and the degree of match between intent and context, ensuring the system accurately understands the user's intent and generates appropriate responses.

[0042] Specifically, the intent recognition module consists of three submodules: intent analysis, intent matching, and intent integration. The intent analysis submodule, based on a context-related parameter set, collects key data through semantic monitoring, including user-entered keywords, semantic features, and contextual time series. It then performs time series analysis on the data, removes outliers, and partitions the remaining data to generate intent analysis data. The intent matching submodule analyzes the relationship between intent and context using the intent analysis data, calculates the impact of each intent change on parameter adjustment, determines the optimal matching parameter settings based on the impact score, and iteratively adjusts the parameters to capture the optimal combination, generating parameter matching results. The intent integration submodule selects time series and context combinations that match the current intent from the parameter matching results, conducts parameter adjustment experiments, optimizes the parameter settings through multiple adjustments and verifications, and finalizes and solidifies the parameters as operational standards, generating the intent recognition parameter set. Intent semantic feature parameters refer to the semantic features of user intent, which can be extracted using a pretrained language model. Context matching parameters refer to the degree of match between intent and context, and can be determined by calculating the similarity between the intent semantic feature vector and the context semantic feature vector.

[0043] Preferably, the intent analysis submodule can employ a deep learning model, such as a recurrent neural network (RNN) or its variant, a long short-term memory network (LSTM). It takes as input the user's natural language input and a set of context-related parameters, and outputs intent analysis data. The intent matching submodule can determine the degree of match between the intent and the context by calculating the cosine similarity between the intent semantic feature vector and the context semantic feature vector. For example, a threshold can be set above which the similarity is considered a good match between the intent and the context. The intent integration submodule can optimize parameter settings through multiple experiments and verifications to ensure the accuracy and stability of intent recognition. For example, cross-validation can be used to test different parameter combinations and select the optimal parameter combination as the final intent recognition parameter set. Through these refined steps and parameter settings, the intent recognition module can more accurately identify the user's intent, providing reliable input for the answer generation module.

[0044] In some embodiments, the answer generation module includes: a knowledge extraction submodule based on the intention recognition parameter set, locates relevant knowledge in the knowledge base, extracts semantic feature values ​​in the knowledge base, continuously records the semantic change rate of the knowledge, extracts multiple key change nodes corresponding to the change rate, sorts the node values ​​in order, and obtains the current knowledge semantic feature value; a semantic matching submodule based on the current knowledge semantic feature value, analyzes the node change value and real-time semantic data, and calibrates according to a predetermined matching criterion, calls the matching criterion to perform distribution redistribution within the semantic interval, and obtains a semantic matching structure; the answer generation submodule is based on the semantic matching structure, adopts a semantic generation dynamic adjustment method, measures the distribution of semantics between context change nodes, sets the upper and lower limits of the node threshold, applies the threshold to multiple rounds of dialogue, and distributes it to obtain multiple rounds of answer generation values.

[0045] It's important to note that the answer generation module is a key component of the intelligent multi-turn question-answering system for generating answers. Its primary function is to extract relevant knowledge from the knowledge base based on the intent recognition parameter set, match knowledge with semantics using real-time semantic data, generate answers, and apply these answers to multi-turn conversations, ultimately generating a multi-turn answer generation value. This multi-turn answer generation value includes knowledge matching parameters, semantic generation parameters, and answer generation parameters. These parameters describe the quality and relevance of the generated answers, ensuring the system provides accurate and coherent responses.

[0046] Specifically, the answer generation module includes three submodules: a knowledge extraction submodule, a semantic matching submodule, and an answer generation submodule. The knowledge extraction submodule locates relevant knowledge in the knowledge base based on the intent recognition parameter set, extracts semantic feature values ​​from the knowledge base, records the semantic change rate of the knowledge, extracts multiple key change nodes in the change rate, sorts the node values ​​in order, and obtains the current knowledge semantic feature value. The semantic matching submodule analyzes the node change value and real-time semantic data based on the current knowledge semantic feature value, calibrates according to the predetermined matching criteria, calls the matching criteria to perform distribution redistribution within the semantic interval, and obtains the semantic matching structure. The answer generation submodule is based on the semantic matching structure and adopts a dynamic adjustment method for semantic generation to measure the distribution of semantics between context change nodes, set the upper and lower limits of the node threshold, apply the threshold to multiple rounds of dialogue, and distribute it to obtain multiple rounds of answer generation values. The knowledge matching parameter refers to the degree of match between the knowledge in the knowledge base and the user's intention, which can be determined by calculating the similarity between the knowledge semantic feature vector and the intention semantic feature vector; the semantic generation parameter refers to the semantic features when generating answers, which can be generated by a pre-trained language model; the answer generation parameter refers to the specific content and format of the generated answers, which can be determined by template generation or free generation.

[0047] Preferably, the knowledge extraction submodule can adopt a deep learning model, such as a pre-trained model of the Transformer architecture, input an intent recognition parameter set, and extract the semantic feature values ​​of relevant knowledge from the knowledge base. The semantic matching submodule can determine the best matching knowledge node by calculating the similarity between the knowledge semantic feature value and the real-time semantic data. For example, the similarity can be calculated using methods such as cosine similarity or Euclidean distance. The answer generation submodule can adopt a sequence generation model, such as the GPT series, input a semantic matching structure and context information, and generate a natural language answer. In the semantic generation dynamic adjustment method, a dynamic adjustment weight coefficient can be set to dynamically adjust the answer generation process according to the semantic features of the node according to the context change. For example, the weight coefficient can be set to a value range of 0 to 1, and the weight can be dynamically adjusted according to the change in context to generate an answer that is more in line with the user's intention. Through these refined operation steps and parameter settings, the answer generation module can more effectively generate high-quality answers and improve the interactive performance of the system.

[0048] In some embodiments, the formula of the semantic generation dynamic adjustment method is as follows: ;in, Indicates the semantically generated value of the matching context change node; Represents the real-time semantic data calculated by the current node; Represents the adjusted semantic feature value of the previous node; Represents the dynamic adjustment weight coefficient, the value range is ; The lower threshold set for the node is used to control the minimum semantic generation amount.

[0049] It's important to note that the dynamic adjustment of semantic generation is a key technology used in the answer generation module to optimize the answer generation process. This method dynamically adjusts the semantic generation value to ensure that answers can flexibly adapt to changes in context, thereby generating answers that better align with user intent and the conversational scenario. The core of this method is to generate high-quality answers by calculating and adjusting dynamic weight coefficients, combining the real-time semantic data of the current node with the semantic feature values ​​of the previous node. This method can effectively improve the coherence and relevance of answers, especially in multi-round conversations, and can better adapt to the dynamic changes in the conversation.

[0050] Specifically, the dynamic adjustment method for semantic generation involves several key parameters: the semantic generation value of the node matching the context change, the real-time semantic data calculated for the current node, the adjusted semantic feature value of the previous node, the dynamic adjustment weight coefficient, and the lower threshold set for the node. The semantic generation value of the node matching the context change refers to the semantic feature value of the answer generated at each key node in the conversation. The real-time semantic data calculated for the current node refers to the semantic feature value calculated based on user input and context information in the current conversation node. The adjusted semantic feature value of the previous node refers to the adjusted semantic feature value of the previous conversation node, which is used for reference and comparison. The dynamic adjustment weight coefficient is a value between 0 and 1 that controls the degree of integration between the semantic data of the current node and the semantic feature value of the previous node. The lower threshold set for the node controls the minimum amount of semantic generation to ensure that the generated answer has a certain level of semantic richness. These parameters work together to generate the final semantic generation value through a specific calculation method.

[0051] Preferably, the method for dynamically adjusting semantic generation can be implemented through the following steps: First, determine the real-time semantic data of the current node. This can be obtained by encoding user input and contextual information using a pre-trained language model. Second, obtain the adjusted semantic feature value of the previous node. This can be achieved by saving the semantic state of the previous node. Then, set a dynamic adjustment weight coefficient, which can be dynamically adjusted based on the fluency and coherence requirements of the conversation. For example, at conversation turning points, the weight can be appropriately reduced to rely more on the semantic data of the current node; at conversation continuation points, the weight can be appropriately increased to maintain conversation coherence. Finally, based on the lower threshold set for the node, calculate the semantic generation value for the node matching the context change. The specific calculation method can be described as follows: the semantic generation value is equal to the dynamic adjustment weight coefficient multiplied by the real-time semantic data of the current node, plus (1 minus the dynamic adjustment weight coefficient) multiplied by the semantic feature value of the previous node, plus the lower threshold set for the node. This method can flexibly adjust the generation of answers based on real-time changes in the conversation, ensuring that the answers both meet the current semantic requirements and maintain the overall coherence of the conversation.

[0052] In some embodiments, the semantic prediction module includes: a semantic data capture submodule that generates values ​​based on the multiple rounds of answers, applies a large model algorithm, captures the semantic data of the sampling points, eliminates outliers and corrects errors, stores them in layers by intervals, performs semantic processing, and generates a semantic data set; a context analysis submodule that divides the intervals according to context-related data based on the semantic data set, extracts change trends and fluctuation features, and generates a context and semantic change feature set; a semantic distribution inference submodule that adjusts feature parameters and calibrates trend data based on the context and semantic change feature set, extracts distribution intervals, and performs numerical predictions to obtain semantic prediction values.

[0053] It's important to note that the semantic prediction module is a key component of intelligent multi-turn question-answering systems for predicting future semantic trends and contextual changes in conversations. Its primary function is to capture semantic data from conversation sampling points based on multi-turn answer generation values, combine it with a large model to predict and infer semantic changes, and analyze contextual changes corresponding to semantic trends. This enables the system to predict user intent and conversational direction, thereby better guiding the conversation process and improving conversation efficiency and quality. The semantic prediction value includes semantic trend analysis parameters and contextual relationship parameters, which describe the predicted semantic change trends and the dynamic relationship between the context.

[0054] Specifically, the semantic prediction module consists of three submodules: a semantic data capture submodule, a context analysis submodule, and a semantic distribution inference submodule. The semantic data capture submodule generates values ​​based on multiple rounds of responses and applies a large model algorithm to capture semantic data at sampling points. It removes outliers and corrects errors, stores data in layers by interval, performs semantic processing, and generates a semantic dataset. The context analysis submodule divides the semantic dataset into intervals based on context-related data, extracts change trends and fluctuation features, and generates a context and semantic change feature set. The semantic distribution inference submodule adjusts feature parameters and calibrates trend data based on the context and semantic change feature set, extracts distribution intervals, and performs numerical predictions to obtain semantic prediction values. Semantic trend analysis parameters refer to the trend characteristics of semantic changes in conversations and can be obtained through time series analysis methods. Context relationship parameters refer to the logical and semantic connections between contexts and can be determined by constructing a context relationship graph.

[0055] The semantic data capture submodule can preferably utilize a deep learning model, such as a pre-trained model using the Transformer architecture, taking as input the values ​​generated by multiple rounds of responses and outputting semantic data at the sampling points. When capturing semantic data, a sampling frequency can be set, for example, once per round of conversation, to ensure data timeliness and accuracy. The context analysis submodule can analyze the changing trends and fluctuation characteristics of the semantic dataset using time series analysis methods, such as moving averages or exponential smoothing. The semantic distribution inference submodule can predict semantic change trends using regression analysis or deep learning prediction models, such as recurrent neural networks (RNNs) or long short-term memory networks (LSTMs). For example, an LSTM model can be trained using the context and a set of semantic change features as input to predict the semantic feature values ​​at the next time point. Through these refined steps and parameter settings, the semantic prediction module can more accurately predict the future direction of the conversation, providing the system with proactive conversation management capabilities.

[0056] In some embodiments, the formula of the large model algorithm is as follows: ;in, Represents the semantic feature vector; Representative The weight of each data point; Representative The original semantic value of the sampling points; Representative Contextual data of each sampling point; Representative Time series value of sampling points; and Represent the weight coefficients of contextual data and time series respectively; Represents the total number of sampling points.

[0057] It's important to note that the large model algorithm plays a core role in the semantic prediction module, processing and analyzing semantic data to predict semantic trends in conversations. This algorithm comprehensively considers the original semantic values ​​of sampling points, contextual data, and time series values ​​to generate semantically-enhanced feature vectors. This approach effectively captures dynamic semantic changes in conversations, providing the system with more accurate semantic predictions, thereby optimizing the conversational flow and improving the user experience.

[0058] Specifically, the large model algorithm involves several key parameters: semanticized feature vectors, data point weights, the original semantic values ​​of sampling points, contextual data, time series values, and weight coefficients for the contextual data and time series. The semanticized feature vectors refer to the semantic features processed by the algorithm and are used to represent the semantic state of each sampling point in the conversation. The data point weights refer to the importance of each sampling point in the calculation and can be adjusted based on the sampling point's temporal position or semantic importance. The original semantic value of a sampling point refers to the semantic information directly obtained in the conversation, the unprocessed raw data. Contextual data refers to contextual information related to the current semantics and is used to enhance semantic understanding. The time series value refers to the timestamp of each sentence in the conversation and is used to indicate the order of the sentences. The weight coefficients for the contextual data and time series are used to adjust the influence of the context and time series in the calculation and can be set according to the specific needs of the conversation.

[0059] Preferably, the large model algorithm can be implemented through the following steps: First, collect data from sample points in the conversation, including the original semantic value, contextual data, and time series values. Then, assign a weight to each data point. The weight can be adjusted based on the data point's recency or semantic importance. For example, more recent conversation data can be given a higher weight. Next, set weight coefficients for the contextual data and time series. These coefficients can be adjusted based on the contextual complexity and temporal relevance of the conversation. For example, in conversations with frequent context changes, the weight of the contextual data can be increased. Finally, the semanticized feature vector is calculated algorithmically. The specific calculation method can be described as follows: the semanticized feature vector is equal to the weight of each data point multiplied by the weighted sum of its original semantic value, contextual data, and time series value. This method can comprehensively consider multiple factors in the conversation and generate more accurate semantic prediction results, thereby supporting dynamic adjustment of the conversation.

[0060] In some embodiments, the dialogue feedback module includes: an error analysis submodule extracts real-time semantic data and context-related data based on the semantic prediction value, analyzes the real-time semantic value and the prediction value, corresponds the semantic difference value with the current context-related data, and generates a semantic context error value; a parameter adjustment submodule sets the adjustment parameters for answer generation based on the semantic context error value, sets the adjustment range for areas with large errors, fine-tunes the low-error areas, screens the matching parameter sets by comparing the adjustment effects and integrates them to generate an answer adjustment parameter set; a feedback control submodule applies the adjustment parameters to each dialogue node based on the answer adjustment parameter set, implements the answer generation operation item by item, synchronously monitors the semantic and context-related data, gradually adjusts the answer generation order of each node, and generates an automatic optimization plan for multiple rounds of dialogue.

[0061] It's important to note that the dialogue feedback module is a key component of intelligent multi-turn question-answering systems for optimizing conversational flow and improving answer quality. Its primary function is to analyze the semantic and contextual error values ​​based on semantic predictions, using real-time semantic data and contextual association data. It then adjusts answer generation based on these error values, ultimately resulting in an automatic multi-turn dialogue optimization solution. This module monitors semantic changes and contextual associations in real time during conversations. Through error analysis and parameter adjustment, it dynamically optimizes answer generation strategies, thereby improving conversational accuracy and coherence.

[0062] Specifically, the dialogue feedback module consists of three submodules: an error analysis submodule, a parameter adjustment submodule, and a feedback control submodule. The error analysis submodule extracts real-time semantic data and contextual data based on semantic predictions, analyzes the real-time semantic values ​​and predicted values, and matches the semantic difference with the current contextual data to generate a semantic contextual error value. The parameter adjustment submodule sets adjustment parameters for answer generation based on the semantic contextual error value, setting adjustment ranges for areas with large errors and fine-tuning for areas with low errors. By comparing the adjustment results, matching parameter sets are selected and integrated to generate an answer adjustment parameter set. Based on the answer adjustment parameter set, the feedback control submodule applies the adjustment parameters to each dialogue node, performing answer generation operations item by item. It simultaneously monitors semantic and contextual data, gradually adjusts the answer generation order for each node, and generates an automatic optimization plan for multiple rounds of dialogue. The semantic contextual error value refers to the difference between real-time semantic data and predicted semantic data, and is used to measure prediction accuracy. The answer adjustment parameter set refers to the answer generation parameters adjusted based on the error analysis results to optimize answer quality.

[0063] Preferably, the error analysis submodule can generate a semantic context error value by calculating the mean square error or absolute error between the real-time semantic data and the predicted semantic data. For example, for each dialogue node, the Euclidean distance between its predicted semantic vector and the actual semantic vector is calculated to obtain the error value. The parameter adjustment submodule can dynamically adjust the answer generation parameters according to the size of the error value. For example, when the error value exceeds a preset threshold, the adjustment amplitude of the answer generation is increased; when the error value is lower than the threshold, fine-tuning is performed. The feedback control submodule can apply the adjusted parameters to each dialogue node and dynamically adjust the answer generation strategy by monitoring the changes in semantic and context-related data in real time. For example, a feedback loop can be set to update the answer adjustment parameter set in real time based on the monitoring results to ensure the consistency and accuracy of the dialogue. Through these detailed operation steps and parameter settings, the dialogue feedback module can effectively optimize the multi-round dialogue process and improve the interactive performance of the system.

[0064] The aforementioned embodiments of the present invention have the following beneficial effects: The intelligent multi-turn question-answering system can enhance the semantic understanding capabilities of multi-turn conversations. By combining semantic feature extraction, contextual analysis, and intent recognition, it can accurately capture user intent and generate coherent conversational responses. The system utilizes dynamic semantic prediction and feedback optimization mechanisms to adjust answer strategies in real time, ensuring the natural flow of conversations. Furthermore, through knowledge base matching and semantic trend analysis, it can improve the accuracy and adaptability of responses.

[0065] This system optimizes the interactive experience of multi-round conversations. Through error analysis and parameter adjustment, it continuously improves the quality of answer generation, making conversations more logical and consistent. Combined with the computing power of large models, it can efficiently handle complex semantic changes and enhance the stability of contextual associations. Dynamic adjustment of semantic generation and predictive inference also enhance the system's intelligence, enabling it to adapt to conversational needs in diverse scenarios, ultimately achieving efficient and accurate intelligent question-and-answer interaction.

[0066] Furthermore, the storage medium of the embodiment of the present application stores program instructions that can implement all the above methods, wherein the program instructions can be stored in the above storage medium in the form of a software product, including a number of instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) or a processor to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, or a terminal device such as a computer, server, mobile phone, or tablet.

[0067] The above descriptions merely illustrate some preferred embodiments of the present invention and the underlying technical principles. Those skilled in the art should understand that the scope of the invention encompassed by the embodiments of the present invention is not limited to technical solutions formed by specific combinations of the aforementioned technical features. It also encompasses other technical solutions formed by any combination of the aforementioned technical features or their equivalents, without departing from the aforementioned inventive concept. For example, a technical solution formed by replacing the aforementioned features with (but not limited to) technical features with similar functions disclosed in the embodiments of the present invention.

Claims

1. An intelligent multi-round question-answering system based on a large model, characterized by: The system includes: a semantic understanding module, which extracts semantic feature vectors based on natural language text input by the user, analyzes context associations, calculates semantic similarity, integrates them into a semantic feature parameter set, and obtains a semantic understanding state value; a context association module, which extracts semantic features and time series of conversation history based on the semantic understanding state value, generates context association data, screens the optimal context association combination, and obtains a context association parameter set; an intention recognition module, which extracts semantic features of user intention based on the context association parameter set, analyzes the relationship between intention and context, matches intention and context combinations, and obtains an intention recognition parameter set; and an answer generation module, which extracts relevant knowledge from the knowledge base based on the intention recognition parameter set. , combined with real-time semantic data, matches knowledge and semantics, generates answers and applies the answers to multi-round dialogues to obtain multi-round answer generation values; the semantic prediction module, based on the multi-round answer generation values, captures the semantic data of the dialogue sampling points, combines the large model to predict and infer semantic changes, analyzes the context changes corresponding to the semantic trends, classifies and organizes the semantic change trends according to the inference results, combines the semantic change information, performs semantic adjustment and analysis on the classified data, and obtains the semantic prediction value; the dialogue feedback module, based on the semantic prediction value, uses real-time semantic data and context association data to analyze the error value of semantics and context, adjusts the answer generation in combination with the error value, and obtains an automatic optimization solution for multi-round dialogues.

2. The intelligent multi-round question-answering system based on a large model according to claim 1 is characterized in that: The semantic understanding state value includes a semantic feature parameter set, a context association parameter set, and a semantic similarity parameter set. The context association parameter set includes a time series parameter and a semantic feature matching parameter. The intention recognition parameter set includes an intention semantic feature parameter and a context matching parameter. The multi-round answer generation value includes a knowledge matching parameter, a semantic generation parameter, and an answer generation parameter. The semantic prediction value includes a semantic trend analysis parameter and a context relationship parameter. The multi-round dialogue automatic optimization scheme includes an error analysis parameter and an answer adjustment parameter.

3. The intelligent multi-round question-answering system based on a large model according to claim 1 is characterized in that: The semantic understanding module includes: a data extraction submodule that performs vectorization processing of semantic features based on the natural language text input by the user, locates invalid semantics, removes noise data, arranges the extracted semantic features in time series, and generates a semantic feature data set; a similarity calculation submodule that analyzes the similarity between semantic features based on the semantic feature data set, calculates the vector inner product of the semantic features, sorts the similarity values ​​by weight, marks areas with excessive semantic differences, and obtains semantic similarity data; a semantic integration submodule that calls the similarity values ​​for multi-dimensional aggregation based on the semantic similarity data, screens semantic feature differences, classifies according to the size of similarity, and arranges the semantic features in order to generate a semantic understanding status value.

4. The intelligent multi-round question-answering system based on a large model according to claim 1 is characterized in that: The context association module includes: a parameter extraction submodule that identifies the semantic features of the conversation history based on the semantic understanding state value, records the time series and semantic features of the context, standardizes the recorded data, organizes and classifies the standardized data according to the time series and semantic features, and generates context association data; a context optimization submodule that analyzes the time series and semantic features in the data set based on the context association data, screens parameter combinations with high matching degrees with the semantic state, adjusts the parameter combinations through pattern matching and records the matching results, and generates parameter combination optimization results; a parameter selection submodule that retrieves the parameter combination optimization results, determines the time series and semantic feature combination with the best matching degree, adjusts the context association parameters, inputs control configuration, verifies the stability of the parameter set, and generates a context association parameter set.

5. The intelligent multi-round question-answering system based on a large model according to claim 1 is characterized in that: The intention recognition module includes: an intention analysis submodule that collects key data through semantic monitoring based on the context-related parameter set, including keywords input by the user, semantic features and context time series, performs time series analysis on the data, eliminates outliers, partitions the remaining data, and obtains intention analysis data; an intention matching submodule analyzes the relationship between intention and context through the intention analysis data, calculates the degree of influence of each intention change on parameter adjustment, determines the optimal matching parameter setting based on the impact score, cyclically adjusts the parameters to capture the optimal combination, and obtains the parameter docking result; an intention integration submodule selects the time series and context combination that matches the current intention from the parameter docking result, performs parameter adjustment tests, optimizes the parameter settings through multiple adjustments and verifications, determines and solidifies the parameters as operating standards, and generates an intention recognition parameter set.

6. The intelligent multi-round question-answering system based on a large model according to claim 1 is characterized in that: The answer generation module includes: a knowledge extraction submodule, based on the intention recognition parameter set, locates relevant knowledge in the knowledge base, extracts semantic feature values ​​in the knowledge base, continuously records the semantic change rate of the knowledge, extracts multiple key change nodes corresponding to the change rate, sorts the node values ​​in order, and obtains the current knowledge semantic feature value; a semantic matching submodule, based on the current knowledge semantic feature value, analyzes the node change value and real-time semantic data, and calibrates according to a predetermined matching criterion, calls the matching criterion to perform distribution redistribution within the semantic interval, and obtains a semantic matching structure; the answer generation submodule, based on the semantic matching structure, adopts a semantic generation dynamic adjustment method, measures the distribution of semantics between context change nodes, sets the upper and lower limits of the node threshold, applies the threshold to multiple rounds of dialogue, and distributes it to obtain multiple rounds of answer generation values.

7. The intelligent multi-round question-answering system based on a large model according to claim 6 is characterized in that: The formula of the semantic generation dynamic adjustment method is as follows: ;in, Indicates the semantically generated value of the matching context change node; Represents the real-time semantic data measured by the current node; Represents the adjusted semantic feature value of the previous node; Represents the dynamic adjustment weight coefficient, the value range is ; The lower threshold set for the node is used to control the minimum semantic generation amount.

8. The intelligent multi-round question-answering system based on a large model according to claim 1 is characterized in that: The semantic prediction module includes: a semantic data capture submodule that generates values ​​based on the multiple rounds of answers, applies a large model algorithm, captures the semantic data of sampling points, eliminates outliers and corrects errors, stores them in layers by intervals, performs semantic processing, and generates a semantic data set; a context analysis submodule that divides intervals according to context-related data based on the semantic data set, extracts change trends and fluctuation features, and generates a context and semantic change feature set; a semantic distribution inference submodule that adjusts feature parameters and calibrates trend data based on the context and semantic change feature set, extracts distribution intervals, and performs numerical prediction to obtain a semantic prediction value.

9. The intelligent multi-round question-answering system based on a large model according to claim 8, characterized in that: The formula of the large model algorithm is as follows: ;in, Represents the semantic feature vector; Representative The weight of each data point; Representative The original semantic value of the sampling points; Representative Contextual data of each sampling point; Representative Time series value of sampling points; and Represent the weight coefficients of contextual data and time series respectively; Represents the total number of sampling points.

10. The intelligent multi-round question-answering system based on a large model according to claim 1, characterized in that: The dialogue feedback module includes: an error analysis submodule that extracts real-time semantic data and context-related data based on the semantic prediction value, analyzes the real-time semantic value and the prediction value, corresponds the semantic difference value with the current context-related data, and generates a semantic context error value; a parameter adjustment submodule that sets adjustment parameters for answer generation based on the semantic context error value, sets adjustment amplitudes for areas with large errors, fine-tunes areas with low errors, screens matching parameter sets by comparing adjustment effects, and integrates them to generate an answer adjustment parameter set; a feedback control submodule that applies adjustment parameters to each dialogue node based on the answer adjustment parameter set, implements answer generation operations item by item, synchronously monitors semantic and context-related data, gradually adjusts the answer generation order of each node, and generates an automatic optimization plan for multiple rounds of dialogue.

Images