Method and system for dynamic correction of multi-turn dialogue intention based on large language model

Abstract

The application discloses a multi-round dialogue intention dynamic correction method and system based on a large language model, and belongs to the technical field of dialogue system semantics processing. The method comprises the following steps: generating an independent intention list of a current round dialogue sentence of a target user according to a current round dialogue sentence, and determining a historical state vector of each independent intention of the current round dialogue sentence of the target user according to historical multi-round dialogue sentences; determining each independent intention with intention conversion according to the historical state vector of each independent intention and a current state vector of each independent intention; updating an initial context vector of each converted intention, updating the independent intention list of the current round dialogue sentence, and outputting a reply result of the current round dialogue sentence of the target user according to the updated independent intention list. Through the implementation of the application, the problem that the dynamic conversion of the user intention cannot be accurately recognized in the complex multi-round dialogue and the response stability of the unconverted intention is ensured can be solved.

Description

Technical Field

[0001] This application belongs to the field of semantic processing technology for dialogue systems, specifically a method and system for dynamic correction of multi-turn dialogue intent based on a large language model. Background Technology

[0002] In today's era of artificial intelligence, multi-turn dialogue systems have become a core pillar of human-computer interaction, directly impacting the reliability and user satisfaction of applications such as intelligent assistants, customer service robots, and virtual companions. These systems help people efficiently complete information retrieval, task planning, and decision support by understanding and responding to the user's continuous expressions, significantly enhancing the convenience and interactive experience of daily life.

[0003] However, existing multi-turn dialogue methods often show significant shortcomings when faced with the diversity and variability of user input. This is because these methods tend to treat the entire dialogue as a single, holistic process. When a user expresses multiple simultaneous ideas, the system struggles to effectively distinguish and independently analyze these ideas. This can compromise the stability of the entire dialogue process if a user's intention changes later in the conversation. For example, in an intelligent customer service scenario, a user might simultaneously inquire about product prices, return policies, and accessory recommendations. If, in a subsequent exchange, the user suddenly shifts their focus from return policies to accessory details, existing dialogue systems often need to re-examine all intents. This not only consumes additional computing resources but can also lead to delays or errors in responses to other unchanged intents, thus affecting the overall coherence and efficiency of the dialogue.

[0004] Therefore, accurately identifying the dynamic shifts in user intent during complex multi-turn dialogues while ensuring the stability of responses to unshifted intents is a problem that urgently needs to be solved. Summary of the Invention

[0005] To address the above issues, this application provides a method and system for dynamic correction of multi-turn dialogue intent based on a large language model. This method solves the problem of accurately identifying the dynamic changes in user intent in complex multi-turn dialogues while ensuring the stability of responses with unchanged intents.

[0006] To achieve the above objectives, the technical solution adopted in this application is a multi-turn dialogue intent dynamic correction method based on a large language model, including: Obtain the target user's historical multi-turn dialogue statements and the current round of dialogue statements; A list of independent intents for the target user's current round of dialogue is generated based on the current round of dialogue statements, and the historical state vectors of each independent intent for the target user's current round of dialogue statements are determined based on historical multi-round dialogue statements; wherein, the list of independent intents includes several independent intents, a unique identifier for each independent intent, an initial context vector for each independent intent, and a current state vector for each independent intent; Based on the historical state vector and the current state vector of each independent intention, determine the independent intentions that will change for each existing intention, and mark the independent intentions that will change for each existing intention as change intentions in the list of independent intentions. For each change intention, the semantic vector of the current change intention is determined based on the current round of dialogue statements, and the initial context vector is updated based on the semantic vector to obtain the updated context vector of the current change intention. The list of independent intents for the current round of dialogue is updated based on the updated context vector of each transition intent, resulting in the updated list of independent intents. Output the response results for the target user's current round of dialogue based on the updated list of independent intents.

[0007] Furthermore, the step of generating an independent intent list of the target user's current round of dialogue statements based on the current round of dialogue statements includes: The intent segmentation algorithm is used to segment the intent of the current round of dialogue, generating several independent intent fragments of the current round of dialogue. For each independent intent fragment, the independent intent fragment is input into a preset intent classification model so that the preset intent classification model outputs the independent intent of the current independent intent fragment; For each independent intent, the independent intent fragment corresponding to the independent intent is embedded and encoded to obtain the initial context vector of the current independent intent; For each independent intent, obtain the previous round dialogue statement of the independent intent fragment corresponding to the current independent intent, and generate the current state vector of the current independent intent based on the previous round dialogue statement and the independent intent fragment. Each independent intent is assigned a unique identifier, and the independent intents are sorted according to dependency syntax rules to obtain a list of independent intents for the target user's current round of dialogue.

[0008] Further, the step of determining the independent intentions that have undergone intention transformation based on the historical state vectors and current state vectors of each independent intention, and marking the independent intentions that have undergone intention transformation as transformation intentions in the list of independent intentions, includes: For each independent intent, calculate the cosine similarity between the current state vector and the historical state vector of the current independent intent to obtain the cosine similarity result of the current independent intent; Compare the cosine similarity result of the current independent intent with the preset cosine similarity threshold; If the cosine similarity result of the current independent intent is less than the preset cosine similarity threshold, the current independent intent is marked as a change intent; If the cosine similarity result of the current independent intent is greater than or equal to the preset cosine similarity threshold, the current independent intent will not be marked.

[0009] Furthermore, after obtaining the updated list of independent intents, it also includes: For each change intention, determine the total number of interaction steps in the historical multi-turn dialogue statements based on the current change intention; Based on the total number of interaction steps of the current change intention and the cosine similarity result, determine the progress metric of the current change intention; Based on the progress metric of each transformation intention, the transformation intentions are prioritized to obtain the first priority response sequence corresponding to each transformation intention. The remaining independent intentions in the updated list of independent intentions, excluding each transformation intention, are prioritized to obtain the second priority response sequence. The first priority response sequence and the second priority response sequence are concatenated to obtain the final priority response sequence. The step of outputting the response result of the target user's current round of dialogue based on the updated list of independent intents includes: Output the response results for the target user's current round of dialogue statements based on the final priority response sequence and the updated list of independent intents.

[0010] Furthermore, the step of outputting the response result of the target user's current round of dialogue statements based on the final priority response sequence and the updated list of independent intents includes: Based on the updated list of independent intents, generate response sub-results for each independent intent; The response sub-results corresponding to each independent intent are input into the intent conflict detection model, so that the intent conflict detection model extracts the semantic vector of each response sub-result, compares the semantic vectors of each response sub-result, and outputs several semantic conflict sets; wherein, the semantic conflict set includes several independent intent response sub-results that identify semantic conflicts. For each set of semantic conflicts, semantic vector fusion is performed on each independent intent response sub-result that identifies semantic conflicts to generate the conflict-free response sub-result corresponding to the current set of semantic conflicts. Based on the final priority response sequence, the independent intent response sub-results that do not conflict and the deconflict response sub-results are concatenated to output the response result of the target user's current round of dialogue.

[0011] Furthermore, the construction of the intent conflict detection model includes: Construct an intent conflict detection training set; wherein, the intent conflict detection training set includes: several intent conflict detection training subsets, each intent conflict detection training subset including several intent response text samples and a set of real semantic conflict samples of the current intent conflict detection training subset; Construct an initial intent conflict detection model; The initial intent conflict detection model is iteratively trained using a subset of intent conflict detection training data as input and a set of predicted semantic conflict samples corresponding to the subset of intent conflict detection training data as output, until the initial intent conflict detection model converges, thereby generating the intent conflict detection model. In each iteration, the initial intent conflict detection model semantically encodes each intent response text sample in the intent conflict detection training subset, generating a semantic encoding vector for each intent response text sample. It then calculates the cosine similarity of each semantic encoding vector, calculates the feature cross-degree between each semantic encoding vector, and outputs the feature cross-degree result. Based on the cosine similarity and feature cross-degree results, it outputs a predicted semantic conflict sample set. Finally, it generates the current iteration loss of the initial intent conflict detection model based on the predicted and actual semantic conflict sample sets, and adjusts the parameters of the initial intent conflict detection model using this current iteration loss.

[0012] Furthermore, for each semantic conflict set, semantic vector fusion is performed on each independent intent response sub-result that identifies a semantic conflict to generate a conflict-free response sub-result corresponding to the current semantic conflict set, including: For each set of semantic conflicts, the independent intent response sub-results with semantic conflicts for each identifier are transformed into semantic conflict encoding vectors; One semantic conflict coding vector is randomly selected from each semantic conflict coding vector as the selected semantic conflict coding vector, and the remaining unselected semantic conflict coding vectors are used as calibration semantic conflict coding vectors. For each calibration semantic conflict coding vector, compare the current calibration semantic conflict coding vector with the selected semantic conflict coding vector, and determine the first semantic conflict coding vector segment and the first semantically consistent coding vector segment of the current calibration semantic conflict coding vector and the selected semantic conflict coding vector, as well as the second semantic conflict coding vector segment and the second semantically consistent coding vector segment of the selected semantic conflict coding vector. A first calibration semantic vector is generated based on the first semantically conflicting coded vector segment and the second semantically consistent coded vector segment. A second calibration semantic vector is generated based on the second semantically conflicting coded vector segment and the first semantically consistent coded vector segment. A conflict-free semantic coded vector to be calibrated is generated based on the first semantically conflicting coded vector segment and the second semantically conflicting coded vector segment. The first calibration semantic vector and the second calibration semantic vector are used to calibrate the conflict-free semantic encoding vector to be calibrated, thereby generating a conflict-free semantic encoding vector. The first semantically consistent encoded vector segment and the second semantically consistent encoded vector segment are semantically fused to generate a fused semantic encoded vector. Based on the decontamination semantic encoding vector and the fusion semantic encoding vector, generate the decontamination response sub-result corresponding to the current semantic conflict set.

[0013] Furthermore, before concatenating the independent intent response sub-results that do not conflict with each other and the deconflict response sub-results according to the final priority response sequence to output the response result for the target user's current round of dialogue, the process also includes: Retrieve the previous round of dialogue from the target user's historical multi-turn dialogue statements; The step of concatenating the independent intent response sub-results without conflict and the de-conflict response sub-results according to the final priority response sequence, and outputting the response result of the target user's current round of dialogue, includes: Based on the final priority response sequence, the independent intent response sub-results that do not conflict and the deconflict response sub-results are concatenated to generate the initial response result for the current round of dialogue statements; Based on the initial response and the previous dialogue statement, generate the first semantic coherence result between the initial response and the previous dialogue statement; If the first semantic coherence result is not less than the preset semantic coherence threshold, the response result of the target user's current round of dialogue is output according to the initial response result; If the first semantic coherence result is less than a preset semantic coherence threshold, the semantic coherence fine-tuning operation is repeated until the second semantic coherence result meets the condition that the second semantic coherence result is not less than the preset semantic coherence threshold. Then, the response result of the target user's current round of dialogue is output according to the updated initial response result. The semantic coherence fine-tuning operation includes: inputting the initial response result, the previous round of dialogue, and the first semantic coherence result into the semantic coherence fine-tuning model so that the semantic coherence fine-tuning model outputs the updated initial response result and calculates the second semantic coherence result between the updated initial response result and the previous round of dialogue.

[0014] The construction of the semantic coherence fine-tuning model includes: Construct a training set for a semantic coherence fine-tuning model; wherein the training set for the semantic coherence fine-tuning model includes: several semantic text samples, associated text samples of each semantic text sample, and semantic coherence results between each semantic text sample and associated text samples. An initial semantic coherence fine-tuning model is constructed. The initial semantic coherence fine-tuning model is iteratively trained using the training set of the initial semantic coherence fine-tuning model until it converges, thus generating a semantic coherence fine-tuning model. In each iteration of training, the initial semantic coherence fine-tuning model extracts features from the current semantic text sample and its associated text samples to obtain the first feature of the current semantic text sample and the second feature of its associated text samples. Based on the semantic coherence result between the current semantic text sample and its associated text samples and the second feature, the first feature is fine-tuned to generate an updated first feature. An updated semantic text sample of the current semantic text sample is then generated based on the updated first feature.

[0015] Based on the above method embodiments, this application provides corresponding system embodiments; One embodiment of this application provides a multi-turn dialogue intent dynamic correction system based on a large language model, including: a dialogue acquisition module, a change intent detection module, a change intent update module, and a dialogue output module; The dialogue acquisition module is used to acquire the target user's historical multi-turn dialogue statements and the current round of dialogue statements; The transition intent detection module is used to generate an independent intent list for the target user's current round of dialogue statements based on the current round of dialogue statements, and to determine the historical state vector of each independent intent for the target user's current round of dialogue statements based on historical multi-round dialogue statements; wherein, the independent intent list includes several independent intents, a unique identifier for each independent intent, an initial context vector for each independent intent, and a current state vector for each independent intent; based on the historical state vector and the current state vector of each independent intent, each independent intent with a transition intent is determined, and each independent intent with a transition intent is marked as a transition intent in the independent intent list; The transition intent update module is used to determine the semantic vector of the current transition intent based on the current round of dialogue statements for each transition intent, update the initial context vector based on the semantic vector to obtain the updated context vector of the current transition intent, and update the list of independent intents of the current round of dialogue statements based on the updated context vectors of each transition intent to obtain the updated list of independent intents. The dialogue output module is used to output the response result of the target user's current round of dialogue statements based on the updated list of independent intents.

[0016] The implementation of this application has the following beneficial effects: This application provides a method and system for dynamic correction of multi-turn dialogue intents based on a large language model. The method acquires the target user's historical multi-turn dialogue statements and the current turn dialogue statement. Based on the current turn dialogue statement, it generates an independent intent list for the target user's current turn dialogue statement. Based on the historical multi-turn dialogue statements, it determines the historical state vector of each independent intent in the target user's current turn dialogue statement. Based on the historical state vector and the current state vector of each independent intent, it determines each independent intent with intent change and marks it as a change intent in the independent intent list. The initial context vector of the change intent is updated, and the current turn dialogue intent is... The list of independent intents for each dialogue statement is updated to obtain an updated list of independent intents. By generating a separate list of independent intents for each dialogue statement in a multi-turn conversation and then comparing it with historical dialogue statements to determine the change of intent, it is possible to ensure that when a user expresses multiple different intents, each intent can be analyzed independently. At the same time, it can accurately identify the user's change of intent in the current dialogue compared with historical dialogues. By identifying the change of intent and the list of independent intents, it is possible to ensure the accurate identification of the change of intent while ensuring the stability of the response for the unchange of intent. This results in the final output response accurately addressing the user's dialogue content and improving the accuracy of the response results in multi-turn conversations. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a method for dynamically correcting multi-turn dialogue intent based on a large language model, as provided in an embodiment of this application.

[0018] Figure 2 This is a schematic diagram of the structure of a multi-turn dialogue intent dynamic correction system based on a large language model, provided as an embodiment of this application. Detailed Implementation

[0019] To enable those skilled in the art to better understand the technical solution, the present application will be described in detail below with reference to the embodiments. The description in this section is only exemplary and explanatory, and should not be used to limit the scope of protection of the present application in any way.

[0020] To enable dialogue systems to accurately detect the changed and unchanged intentions in user dialogues, thereby making the dialogue system more accurate in semantic processing and providing responses that meet user needs, this application provides a multi-turn dialogue intention dynamic correction method based on a large language model.

[0021] like Figure 1 The image shown is an embodiment of a multi-turn dialogue intent dynamic correction method based on a large language model provided in this application, comprising: Step S1: Obtain the target user's historical multi-turn dialogue statements and the current round of dialogue statements; Step S2: Generate an independent intent list for the target user's current round of dialogue based on the current round of dialogue statements, and determine the historical state vector of each independent intent for the target user's current round of dialogue statements based on historical multi-round dialogue statements; wherein, the independent intent list includes several independent intents, a unique identifier for each independent intent, an initial context vector for each independent intent, and a current state vector for each independent intent; Step S3: Based on the historical state vector of each independent intention and the current state vector of each independent intention, determine the independent intentions that have changed, and mark the independent intentions that have changed as change intentions in the list of independent intentions. Step S4: For each change intention, determine the semantic vector of the current change intention based on the current round of dialogue statements, and update the initial context vector based on the semantic vector to obtain the updated context vector of the current change intention. Step S5: Update the list of independent intents for the current round of dialogue statements based on the updated context vector of each transition intent to obtain the updated list of independent intents; Step S6: Output the response results of the target user's current round of dialogue statements based on the updated list of independent intents.

[0022] For step S1, when the dialogue system performs semantic processing on the user's dialogue, it first needs to obtain the target user's historical multi-turn dialogue statements and the dialogue statement of the current turn. The dialogue statement of the current turn is the key focus of the dialogue system. Based on the target user's current turn dialogue statement, the dialogue system needs to determine the desired dialogue response. Simultaneously, it needs to use historical multi-turn dialogue statements to refine the current turn dialogue statement and produce a more suitable response. It should be noted that historical multi-turn dialogue statements include the dialogue statement from the first statement entered by the target user in the dialogue system's dialog box to the dialogue statement of the turn preceding the current turn, but do not include the current turn dialogue statement.

[0023] For step S2, parse the target user's current round of dialogue statements to determine the individual intents present in the current round of dialogue statements, and generate an independent intent list based on each individual intent.

[0024] In a preferred embodiment, generating an independent intent list of the target user's current round of dialogue statements based on the current round of dialogue statements includes: segmenting the current round of dialogue statements based on an intent separation algorithm to generate several independent intent fragments of the current round of dialogue statements; for each independent intent fragment, inputting the independent intent fragment into a preset intent classification model so that the preset intent classification model outputs the independent intent of the current independent intent fragment; for each independent intent, embedding and encoding the independent intent fragment corresponding to the independent intent to obtain an initial context vector of the current independent intent; for each independent intent, obtaining the previous round of dialogue statements of the independent intent fragment corresponding to the current independent intent, and generating a current state vector of the current independent intent based on the previous round of dialogue statements and the independent intent fragments; assigning a unique identifier to each independent intent, and sorting each independent intent based on dependency syntax rules to obtain an independent intent list of the target user's current round of dialogue statements.

[0025] Specifically, an intent segmentation algorithm based on dependency syntax rules is used to segment the intent of the current round of dialogue, resulting in multiple independent intent fragments. It should be noted that the intent segmentation algorithm based on dependency syntax rules used in this application uses syntactic root nodes, parallel relationships, and conjunctions as segmentation criteria, primarily for splitting multi-intent statements. For example, if the current round of dialogue is "I want to book a flight and check hotel information," the intent segmentation algorithm can identify the two intents "want to book a flight" and "check hotel information," resulting in two independent intent fragments: "want to book a flight" and "check hotel information." These two independent intent fragments are then input into a preset intent classification model, which outputs the two independent intents "book a flight" and "check hotel information" based on the input fragments. The preset intent separation model is built on a classification model. The construction of the preset intent separation model includes the construction of a training dataset and model training. When constructing the training dataset, multiple independent intent fragment samples are obtained from big data, and each independent intent fragment sample is labeled with the independent intent corresponding to each independent intent fragment sample. When training the model, an initial classification model is first constructed, and supervised training is performed on the initial classification model using the constructed training dataset. When the initial classification model converges, the converged initial classification model is used as the preset intent separation model.

[0026] After separating the independent intents of the current round of dialogue statements, it is necessary to embed and encode the independent intent fragments corresponding to the independent intents. In this application, the embedding and encoding adopts BERT-based model encoding, and the resulting initial context vector only represents the semantics of the current independent intent, which is a static and invariant semantic encoding. For example, the initial context vector of "booking a flight ticket" can be encoded into an initial context vector [0.3, 0.5, 0.7] based on the BERT model. Obtain the previous round of dialogue statements corresponding to the current independent intent fragments. Generate the current state vector of the current independent intent based on the previous round of dialogue statements and independent intent fragments. The obtained current state vector is associated with the current independent intent and the previous round of dialogue statements, and is used to represent the user's latest attitude, changes, and emphasis on the independent intent in this round. For example, if the obtained current state vector is [0], it means that the independent intent of the current dialogue round has not changed in the previous round of dialogue. Conversely, if the current state vector is [1], it means that the independent intent of the current dialogue round has changed in the previous round of dialogue. After determining the independent intents, their initial context vectors, and current state vectors, each independent intent is assigned a unique identifier (identity ID). Based on dependency syntax rules, the independent intents are sorted according to the root node position to generate a tree-like list of independent intents for the current round of dialogue statements.

[0027] The process involves parsing the target user's historical multi-turn dialogue statements and determining the historical state vector of each independent intent in the current turn of the dialogue. For example, it involves determining the intent in the current turn of the dialogue. Figure 1 "Booking a flight" and Italy Figure 2 After "checking hotels", the current meaning is retrieved from historical multi-turn dialogue statements. Figure 1 Harmony Figure 2 The frequency of occurrence and successful processing of statements in historical multi-turn dialogues, thereby obtaining the meaning. Figure 1 Harmony Figure 2 The historical state vector in a multi-turn historical dialogue. (If the desired result is obtained...) Figure 1 The historical state vector [2,1,0,1] indicates that booking a flight ticket occurred twice in the historical dialogue: once successfully, zero times unsuccessfully, and once pending. Pending processing typically occurs when the user modifies the dialogue content after the system hasn't responded to their question. After completing the above processing, a list of independent intents for the target user's current dialogue round and the historical state vectors of each associated independent intent can be obtained. By determining the independent intents before processing the historical dialogue, the computational power requirement can be reduced.

[0028] For step S3, in a preferred embodiment, determining each independent intent with an intent change based on the historical state vector and the current state vector of each independent intent, and marking each independent intent with an intent change as a change intent in the independent intent list, includes: for each independent intent, calculating the cosine similarity between the current state vector and the historical state vector of the current independent intent to obtain the cosine similarity result of the current independent intent; comparing the cosine similarity result of the current independent intent with a preset cosine similarity threshold; if the cosine similarity result of the current independent intent is less than the preset cosine similarity threshold, marking the current independent intent as a change intent; if the cosine similarity result of the current independent intent is greater than or equal to the preset cosine similarity threshold, not marking the current independent intent.

[0029] Specifically, assuming the current state vector of the independent intent "book a flight ticket" is [1] and the historical state vector is [2,1,0,1], the dimensions of the current state vector and the historical state vector are first unified, that is, the current state vector and the historical state vector are mapped, and the current state vector [1] is mapped to a 4-dimensional vector, resulting in [1,1,1,1]. The current state vector and the historical state vector are normalized, resulting in the normalized result of the current state vector [0.5, 0.5, 0.5, 0.5] and the normalized result of the historical state vector [0.816, 0.408, 0, 0.408]. The cosine similarity result is calculated to be 0.166 using the following cosine calculation formula.

[0030]

[0031] in, This represents the cosine similarity result; the closer it is to 1, the more similar the two vectors are. This is the current state vector; This is the historical state vector.

[0032] After calculating the cosine similarity result, it is compared with a preset cosine similarity threshold. If it is less than the preset threshold, the current independent intent is marked as a change intent; if it is greater than or equal to the preset threshold, the current independent intent is not marked and is determined to be a non-change intent. Change intents and independent intents are then marked in the independent intent list. This completes the identification of change intents in the current round of dialogue, while not intervening in non-change intents. The preset cosine similarity threshold is determined by collecting multi-turn dialogue samples from specific domains and conducting statistical analysis of positive and negative intent change samples. In the travel domain example above, a large number of real multi-turn dialogue intent change samples from flight and hotel scenarios are selected, and the cosine similarity of the vectors for actual intent change and non-change intent is calculated separately. The boundary value of the two data distributions is then used as the preset cosine similarity threshold.

[0033] For step S4, for each change intention, determine the semantic vector of the current change intention based on the current round of dialogue statements, and update the initial context vector of the independent intention corresponding to each change intention based on the semantic vector of the current change intention to obtain the updated context vector of the current change intention.

[0034] For step S5, the list of independent intents for the current round of dialogue is updated according to the updated context vector of each transition intent, resulting in the updated list of independent intents.

[0035] For step S6, the dialogue system can generate an accurate response to the user's current dialogue statement based on the updated list of independent intents.

[0036] To avoid conflicts between intents and issues with contextual coherence in the output response, this application proposes a preferred implementation method.

[0037] In a preferred embodiment, after obtaining the updated list of independent intents, the method further includes: for each change intent, determining the total number of interaction steps in the historical multi-turn dialogue statements based on the historical multi-turn dialogue statements; determining the progress metric of the current change intent based on the total number of interaction steps and the cosine similarity result; prioritizing each change intent based on its progress metric to obtain a first priority response sequence for each change intent; prioritizing the remaining independent intents in the updated list of independent intents (excluding the change intents) to obtain a second priority response sequence; and concatenating the first and second priority response sequences to obtain a final priority response sequence. The step of outputting the response result of the target user's current turn dialogue statement based on the updated list of independent intents includes: outputting the response result of the target user's current turn dialogue statement based on the final priority response sequence and the updated list of independent intents.

[0038] Specifically, after obtaining the updated list of independent intents, the total number of interaction steps for each change intent in the updated list is determined based on the target user's historical multi-turn dialogue statements. For example, if the change intent "book a flight" appears 4 times in the historical multi-turn dialogue statements, then the total number of interaction steps for the change intent "book a flight" in the historical multi-turn dialogue is determined to be 4. It should be noted that the number of interaction steps is determined based on the number of turns in the historical multi-turn dialogue. If the same intent appears multiple times in the same turn, the number of interaction steps is not calculated based on the number of times it appears, but only on the number of turns. After determining the total number of interaction steps for the current change intent in the historical multi-turn dialogue statements, the progress metric of the current change intent is determined based on the total number of interaction steps and the cosine similarity result. The formula for calculating the progress metric is as follows:

[0039]

[0040] in, For progress metrics; This is the current state vector; This is a vector of historical states; This is the cosine similarity result.

[0041] Calculating the progress metric of the current intention to change allows for the quantification of the degree and urgency of the intention change. It transforms the dynamic change of user intention from an abstract semantic change into a concrete numerical value of 0 to 1, accurately measuring the magnitude of the intention change, such as the change from "searching for hotels" to "booking hotels." It also accurately measures the urgency of the interaction between the intentions before and after the change; for example, intentions repeatedly mentioned by the user have higher urgency. The calculated progress metric of each intention to change can serve as the core basis for prioritizing intentions. During priority ranking, each intention to change is prioritized based on its progress metric, resulting in a first-priority response sequence for each intention to change. The remaining independent intentions in the updated list of independent intentions (excluding the intentions to change) are then prioritized to obtain a second-priority response sequence. The first and second priority response sequences are then concatenated to obtain the final priority response sequence. In this final priority response sequence, the first-priority response sequence corresponding to the intention to change will be responded to first by the system, and the intention to change with a higher progress metric in the first-priority response sequence will also be responded to first. This priority ranking ensures that the dialogue system prioritizes responding to the intentions that users are more concerned about and those with more significant changes, avoiding response delays or the neglect of key intentions due to indiscriminate processing.

[0042] In a preferred embodiment, the step of outputting the response result of the target user's current round of dialogue statement based on the final priority response sequence and the updated list of independent intents includes: generating response sub-results corresponding to each independent intent based on the updated list of independent intents; inputting the response sub-results corresponding to each independent intent into an intent conflict detection model, so that the intent conflict detection model extracts the semantic vectors of each response sub-result and compares the semantic vectors of each response sub-results to output several semantic conflict sets; wherein, the semantic conflict set includes several independent intent response sub-results that are identified as having semantic conflicts; for each semantic conflict set, performing semantic vector fusion on each independent intent response sub-result that is identified as having semantic conflicts to generate a conflict-free response sub-result corresponding to the current semantic conflict set; and concatenating each independent intent response sub-result that does not have a conflict and each conflict-free response sub-result according to the final priority response sequence to output the response result of the target user's current round of dialogue statement.

[0043] Specifically, based on the updated list of independent intents, a corresponding response sub-result is generated for each independent intent. For example, regarding the intent... Figure 1 The "Book Flight" function generates a sub-result with a "Link to Book Flight" option; for those interested... Figure 2 The "Search Hotels" function generates sub-results for "Hotel Search Results Near Your Flight Destination". After obtaining these sub-results, they are input into an intent conflict detection model. This model extracts the semantic vectors of each sub-result, compares them, and determines whether semantic conflicts exist between them. If semantic conflicts exist, it outputs a set of semantic conflicts. For example, given the intent... Figure 1 The first sub-result for the "Book Flight" request is "Link to Book Flight," meaning... Figure 2The response sub-result 2 for "Search hotels" is "Search results for hotels near the flight destination". The response sub-result 3 for "Book hotels" is "Search results for hotels near the flight destination, prices of each hotel and room types of each hotel". We set a cosine similarity greater than 0.8 as semantically high similarity and semantic conflict. The response sub-results 1-3 are input into the intent conflict detection model. The model performs semantic encoding on each sub-result using the BERT encoding layer, generating the semantic vector V1 for response sub-result 1: [0.23, -0.15, 0.41, 0.09, -0.32, 0.17, 0.55, -0.08, 0.29, -0.11]. The vector features include semantic features of "flight ticket," "booking," and "link," with the numerical distribution reflecting the core semantics of "flight booking tool." The semantic vector V2 for response sub-result 2 is generated: [-0.12, 0.37, 0.05, 0.28, -0.19, 0.44, -0.21, 0.33, 0.15, ... [0.07], Vector features: Integrating the semantics of "flight destination", "hotel" and "search results", it not only relates to the flight scenario, but also highlights the basic information attributes of the hotel query; Generate the semantic vector V3 of response sub-result 3: [-0.11, 0.39, 0.04, 0.27, -0.20, 0.46, -0.22, 0.35, 0.16, 0.08], Vector features: Highly similar to the vector of response sub-result 2 (cosine similarity ≈ 0.98), with only slight deviations in the values ​​corresponding to the dimensions of "price" and "room type", reflecting the extension and refinement of semantics. By calculating the pairwise cosine similarity between V1, V2, and V3, we can obtain that the cosine similarity between V1 and V2 is -0.0121, the cosine similarity between V1 and V3 is -0.0181, and the cosine similarity between V2 and V3 is 1.1475 (normalized to 1.0). That is, the similarity between V1 and V2, and between V1 and V3 is close to 0 and negative, indicating that "booking flights" and "checking hotels / booking hotels" have no semantic conflict. The similarity between V2 and V3 is ≈1.0, indicating that "checking hotels" and "booking hotels" have a high semantic overlap and belong to the inclusion conflict. At this time, the intent conflict detection model outputs a semantic conflict set, which includes response sub-result 2 and response sub-result 3, and marks the text in response sub-result 2 that conflicts with response sub-result 3, and the text in response sub-result 3 that conflicts with response sub-result 2.

[0044] In a preferred embodiment, the construction of the intent conflict detection model includes: constructing an intent conflict detection training set; wherein the intent conflict detection training set includes: several intent conflict detection training subsets, each intent conflict detection training subset including several intent response text samples and a set of real semantic conflict samples of the current intent conflict detection training subset; constructing an initial intent conflict detection model; using the intent conflict detection training subsets as input to the initial intent conflict detection model and the predicted semantic conflict sample set corresponding to the intent conflict detection training subsets as output, iteratively training the initial intent conflict detection model until the initial intent conflict detection model converges, thereby generating the intent conflict detection model; wherein, in each iteration process... In the initial intent conflict detection model, semantic encoding is performed on each intent response text sample in the intent conflict detection training subset to generate a semantic encoding vector for each intent response text sample. The cosine similarity of each semantic encoding vector is calculated to obtain the cosine similarity result of each semantic encoding vector. The feature cross-degree between each semantic encoding vector is calculated and the feature cross-degree result of each semantic encoding vector is output. Based on the cosine similarity result and the feature cross-degree result of each semantic encoding vector, a predicted semantic conflict sample set is output. The current iteration loss of the initial intent conflict detection model is generated based on the predicted semantic conflict sample set and the real semantic conflict sample set. The parameters of the initial intent conflict detection model are adjusted based on the current iteration loss of the initial intent conflict detection model.

[0045] Specifically, the initial intent conflict detection model is constructed based on the BERT semantic fusion feature classification model. This model includes a BERT pre-trained encoding layer, a semantic feature fusion layer, a cosine similarity calculation layer, a feature cross-operation layer, and a fully connected classification output layer. When training the initial intent conflict detection model, a subset of intent conflict detection training data is used as input, and the set of predicted semantic conflict samples corresponding to that subset is used as output. Supervised learning training is then performed on the initial intent conflict detection model. In each round of training, the BERT pre-trained encoding layer and semantic feature fusion layer of the initial intent conflict detection model semantically encode each intent response text sample in the intent conflict detection training subset, generating a semantic encoding vector for each intent response text sample. The cosine similarity calculation layer calculates the cosine similarity of each semantic encoding vector to obtain the cosine similarity result of each semantic encoding vector. The feature cross operation layer calculates the feature cross degree between each semantic encoding vector and outputs the feature cross degree result of each semantic encoding vector. The fully connected classification output layer outputs the predicted semantic conflict sample set based on the cosine similarity result and the feature cross degree result of each semantic encoding vector.

[0046] In a preferred embodiment, the step of performing semantic vector fusion on each independent intent response sub-result that identifies semantic conflict for each semantic conflict set to generate a conflict-free response sub-result corresponding to the current semantic conflict set includes: for each semantic conflict set, converting each independent intent response sub-result that identifies semantic conflict into a semantic conflict encoding vector; randomly selecting one semantic conflict encoding vector from each semantic conflict encoding vector as the selected semantic conflict encoding vector, and using the remaining unselected semantic conflict encoding vectors as calibration semantic conflict encoding vectors; for each calibration semantic conflict encoding vector, comparing the current calibration semantic conflict encoding vector with the selected semantic conflict encoding vector, determining a first semantic conflict encoding vector segment and a first semantic consistency encoding vector segment between the current calibration semantic conflict encoding vector and the selected semantic conflict encoding vector, and selecting the semantic conflict encoding vector. The first semantic conflict encoded vector segment and the second semantic consistency encoded vector segment are used to generate a first calibration semantic vector based on the first semantic conflict encoded vector segment and the second semantic consistency encoded vector segment. A second calibration semantic vector is generated based on the second semantic conflict encoded vector segment and the first semantic consistency encoded vector segment. A conflict-free semantic encoded vector to be calibrated is generated based on the first semantic conflict encoded vector segment and the second semantic conflict encoded vector segment. The conflict-free semantic encoded vector to be calibrated is calibrated using the first calibration semantic vector and the second calibration semantic vector to generate a conflict-free semantic encoded vector. The first semantic consistency encoded vector segment and the second semantic consistency encoded vector segment are semantically fused to generate a fused semantic encoded vector. A conflict-free response sub-result corresponding to the current semantic conflict set is generated based on the conflict-free semantic encoded vector and the fused semantic encoded vector.

[0047] Specifically, continuing the example above, response sub-results 2 and 3 are transformed into semantic conflict coding vectors V4: [-0.12, 0.37, 0.05, 0.28, -0.19, 0.44, -0.21, 0.33, 0.15, 0.07] and V5: [-0.11, 0.39, 0.04, 0.27, -0.20, 0.46, -0.22, 0.35, 0.16, 0.08]. V4 is used as the selected semantic conflict coding vector, and V5 is used as the calibration semantic conflict coding vector. By comparing V4 and V5, if the values ​​are close, they are determined to be semantically consistent coding segments; if the values ​​differ significantly, they are determined to be semantically conflict coding segments. The second semantically consistent encoded vector segment [-0.12, 0.37, 0.05, 0.28, -0.19], the second semantically conflicted encoded vector segment [0.44, -0.21, 0.33, 0.15, 0.07], the first semantically consistent encoded vector segment [-0.11, 0.39, 0.04, 0.27, -0.20], and the first semantically conflicted encoded vector segment [0.46, -0.22, 0.35, 0.16, 0.08] can be obtained. A first calibration semantic vector [0.46, -0.22, 0.35, 0.16, 0.08, -0.12, 0.37, 0.05, 0.28, -0.19] is generated based on the first semantically conflicting coded vector segment and the second semantically consistent coded vector segment. A second calibration semantic vector [0.44, -0.21, 0.33, 0.15, 0.07, -0.11, 0.39, 0.04, 0.27, -0.20] is generated based on the second semantically conflicting coded vector segment and the first semantically consistent coded vector segment. A conflict-resolved semantic coded vector to be calibrated [0.46, -0.22, 0.35, 0.16, 0.08, 0.44, -0.21, 0.33, 0.15] is generated based on the first and second semantically conflicting coded vector segments. [0.07]; Using the first calibration semantic vector and the second calibration semantic vector, the uncalibrated conflict-free semantic encoding vector is corrected by weighted averaging along each dimension, resulting in the conflict-free semantic encoding vector [0.45, -0.22, 0.34, 0.16, 0.08, 0.45, -0.21, 0.34, 0.15, 0.07].The first semantically consistent encoded vector segment and the second semantically consistent encoded vector segment are semantically fused to generate a fused semantic encoded vector [-0.12, 0.38, 0.04, 0.27, -0.20]. The conflict-resolved semantic encoded vector and the fused semantic encoded vector are concatenated to obtain the semantic encoded vector of the conflict-resolved response sub-result corresponding to the current semantic conflict set [0.45, -0.22, 0.34, 0.16, 0.08, 0.45, -0.21, 0.34, 0.15, 0.07, -0.12, 0.38, 0.04, 0.27, -0.20]. Based on this semantic encoded vector, the conflict-resolved response sub-result "Hotels near your flight destination have been searched and their prices and room types are displayed" is output.

[0048] Based on the final priority response sequence, the independent intent response sub-results without conflict and the de-conflict response sub-results are concatenated to output the response result for the target user's current round of dialogue. Continuing the above example, the response result for the target user's current opposing dialogue statement is output as: "I have searched for hotels near your flight destination and displayed the prices and room types of each hotel. I have also provided you with a link to book your flight."

[0049] In a preferred embodiment, before concatenating the non-conflicting independent intent response sub-results and the de-conflict response sub-results according to the final priority response sequence to output the response result of the target user's current round of dialogue, the method further includes: obtaining the previous round of dialogue from the target user's historical multi-round dialogue statements; the step of concatenating the non-conflicting independent intent response sub-results and the de-conflict response sub-results according to the final priority response sequence to output the response result of the target user's current round of dialogue includes: concatenating the non-conflicting independent intent response sub-results and the de-conflict response sub-results according to the final priority response sequence to generate an initial response result for the current round of dialogue; and generating a first semantic connection between the initial response result and the previous round of dialogue based on the initial response result and the previous round of dialogue. The semantic coherence result is as follows: if the first semantic coherence result is not less than a preset semantic coherence threshold, the response result of the target user's current round of dialogue is output according to the initial response result; if the first semantic coherence result is less than the preset semantic coherence threshold, the semantic coherence fine-tuning operation is repeated until the second semantic coherence result satisfies the condition that the second semantic coherence result is not less than the preset semantic coherence threshold, and the response result of the target user's current round of dialogue is output according to the updated initial response result; wherein, the semantic coherence fine-tuning operation includes: inputting the initial response result, the previous round of dialogue statement and the first semantic coherence result into the semantic coherence fine-tuning model, so that the semantic coherence fine-tuning model outputs the updated initial response result, and calculating the second semantic coherence result between the updated initial response result and the previous round of dialogue statement.

[0050] Specifically, the previous round's dialogue statement is obtained from the target user's historical multi-turn dialogue statements. Assuming the previous round's dialogue statement was "How's the weather today?", and the initial response to the current round's dialogue statement was "I have searched for hotels near your flight destination and displayed their prices and room types. I have also provided a link to book your flight." By converting the previous round's dialogue statement and the initial response to the current round's dialogue statement into semantic encoding vectors, and then calculating the cosine similarity between these two semantic encoding vectors, the first semantic coherence result can be obtained. The value range of the first semantic coherence result is [-1, 1], with values ​​closer to 1 indicating greater contextual coherence. Because the semantic relevance between "How's the weather today?" and "I've searched for hotels near your flight destination and displayed their prices and room types, and provided a link to book your flight," the calculated first semantic coherence result is set to 0.23, and the preset semantic coherence threshold is 0.8. Since the first semantic coherence result is less than the preset threshold, "I've searched for hotels near your flight destination and displayed their prices and room types, and provided a link to book your flight," "How's the weather today?" and "0.23" are input into the semantic coherence fine-tuning model. The model updates the initial response based on the input content and outputs "How's the weather today?" The system displays a message stating, "The weather is mainly sunny. We have also checked hotel and room rates near your flight destination and provided a booking link." It then calculates the second semantic coherence between "The weather is mainly sunny. We have also checked hotel and room rates near your flight destination and provided a booking link" and "How's the weather today?". If this second semantic coherence result is greater than a preset semantic coherence threshold, the final response, "The weather is mainly sunny. We have also checked hotel and room rates near your flight destination and provided a booking link," is output. If the second semantic coherence result is less than the preset semantic coherence threshold, an updated initial response is generated until the semantic coherence threshold requirement is met. The preset semantic coherence threshold is determined by statistically analyzing the cosine similarity distribution range of multiple rounds of dialogue context in a large number of selected scenarios, combined with manual calibration of dialogue fluency and iterative model optimization. In the above example travel scenario, this solution has been experimentally optimized to a value of 0.8.

[0051] In a preferred embodiment, the construction of the semantic coherence fine-tuning model includes: Construct a training set for a semantic coherence fine-tuning model; wherein the training set for the semantic coherence fine-tuning model includes: several semantic text samples, associated text samples of each semantic text sample, and semantic coherence results between each semantic text sample and associated text samples. An initial semantic coherence fine-tuning model is constructed. The initial semantic coherence fine-tuning model is iteratively trained using the training set of the initial semantic coherence fine-tuning model until it converges, thus generating a semantic coherence fine-tuning model. In each iteration of training, the initial semantic coherence fine-tuning model extracts features from the current semantic text sample and its associated text samples to obtain the first feature of the current semantic text sample and the second feature of its associated text samples. Based on the semantic coherence result between the current semantic text sample and its associated text samples and the second feature, the first feature is fine-tuned to generate an updated first feature. An updated semantic text sample of the current semantic text sample is then generated based on the updated first feature.

[0052] Specifically, the aforementioned initial semantic coherence fine-tuning model is a BERT-GPT lightweight sequence generation fine-tuning model, which is a publicly available and mature model in the field of natural language dialogue. Based on the existing model, the initial semantic coherence fine-tuning model is trained by constructing a training set for the semantic coherence fine-tuning model to generate the aforementioned semantic coherence fine-tuning model that conforms to the application of this application.

[0053] After outputting the response to the target user's current round of dialogue, the target user's historical multi-round dialogue statements are updated based on the target user's current round of dialogue statements and response results, so as to be used for semantic correction in the next user dialogue process.

[0054] By performing intent separation, intent shift analysis, response content conflict analysis, and response content contextual coherence analysis on the user's current round of dialogue statements, the dialogue system can accurately identify the user's changed intent, respond to the user's changed intent, and respond to the unchanged intent in the response results. This ensures the logical accuracy and contextual coherence of the response results. The technical solution of this application significantly improves the accuracy of intent correction, response coherence, and overall interaction efficiency of the dialogue system, avoids the response deviation caused by ignoring intent shifts in traditional methods, and achieves a more intelligent multi-turn dialogue experience.

[0055] Based on the above method embodiments, this application provides corresponding system embodiments.

[0056] like Figure 2 As shown, one embodiment of this application provides a multi-turn dialogue intent dynamic correction system based on a large language model, including: a dialogue acquisition module, a change intent detection module, a change intent update module, and a dialogue output module; The dialogue acquisition module is used to acquire the target user's historical multi-turn dialogue statements and the current round of dialogue statements; The transition intent detection module is used to generate an independent intent list for the target user's current round of dialogue statements based on the current round of dialogue statements, and to determine the historical state vector of each independent intent for the target user's current round of dialogue statements based on historical multi-round dialogue statements; wherein, the independent intent list includes several independent intents, a unique identifier for each independent intent, an initial context vector for each independent intent, and a current state vector for each independent intent; based on the historical state vector and the current state vector of each independent intent, each independent intent with a transition intent is determined, and each independent intent with a transition intent is marked as a transition intent in the independent intent list; The transition intent update module is used to determine the semantic vector of the current transition intent based on the current round of dialogue statements for each transition intent, update the initial context vector based on the semantic vector to obtain the updated context vector of the current transition intent, and update the list of independent intents of the current round of dialogue statements based on the updated context vectors of each transition intent to obtain the updated list of independent intents. The dialogue output module is used to output the response result of the target user's current round of dialogue statements based on the updated list of independent intents.

[0057] It should be noted that, in this document, the terms "comprising," "including," and any other variations are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Specific examples have been used in this document to illustrate the principles and implementation methods of the technical solutions of this application. The above examples are only for the purpose of helping to understand the methods and core ideas of this application. The above descriptions are merely preferred embodiments of this application. It should be pointed out that, due to the limitations of written expression and the objective existence of infinite specific structures, those skilled in the art can make several improvements, modifications, or changes without departing from the principles of this application, and can also combine the above technical features in an appropriate manner; these improvements, modifications, changes, or combinations, or the direct application of the concept and technical solutions of this application to other situations without modification, should all be considered within the scope of protection of this application.

Claims

1. A method for dynamic correction of multi-turn dialogue intent based on a large language model, characterized in that, include: Obtain the target user's historical multi-turn dialogue statements and the current round of dialogue statements; A list of independent intents for the target user's current round of dialogue is generated based on the current round of dialogue statements, and the historical state vectors of each independent intent for the target user's current round of dialogue statements are determined based on historical multi-round dialogue statements; wherein, the list of independent intents includes several independent intents, a unique identifier for each independent intent, an initial context vector for each independent intent, and a current state vector for each independent intent; Based on the historical state vector and the current state vector of each independent intention, determine the independent intentions that will change for each existing intention, and mark the independent intentions that will change for each existing intention as change intentions in the list of independent intentions. For each change intention, the semantic vector of the current change intention is determined based on the current round of dialogue statements, and the initial context vector is updated based on the semantic vector to obtain the updated context vector of the current change intention. The list of independent intents for the current round of dialogue is updated based on the updated context vector of each transition intent, resulting in the updated list of independent intents. Output the response results for the target user's current round of dialogue based on the updated list of independent intents.

2. The method for dynamic correction of multi-turn dialogue intent based on a large language model according to claim 1, characterized in that, The step of generating an independent intent list for the target user's current round of dialogue based on the current round of dialogue includes: The intent segmentation algorithm is used to segment the intent of the current round of dialogue, generating several independent intent fragments of the current round of dialogue. For each independent intent fragment, the independent intent fragment is input into a preset intent classification model so that the preset intent classification model outputs the independent intent of the current independent intent fragment; For each independent intent, the independent intent fragment corresponding to the independent intent is embedded and encoded to obtain the initial context vector of the current independent intent; For each independent intent, obtain the previous round dialogue statement of the independent intent fragment corresponding to the current independent intent, and generate the current state vector of the current independent intent based on the previous round dialogue statement and the independent intent fragment. Each independent intent is assigned a unique identifier, and the independent intents are sorted according to dependency syntax rules to obtain a list of independent intents for the target user's current round of dialogue.

3. The method for dynamic correction of multi-turn dialogue intent based on a large language model according to claim 2, characterized in that, The step of determining the independent intentions that have undergone transformation based on the historical state vectors and current state vectors of each independent intention, and marking each independent intention that has undergone transformation as a transformation intention in the list of independent intentions, includes: For each independent intent, calculate the cosine similarity between the current state vector and the historical state vector of the current independent intent to obtain the cosine similarity result of the current independent intent; Compare the cosine similarity result of the current independent intent with the preset cosine similarity threshold; If the cosine similarity result of the current independent intent is less than the preset cosine similarity threshold, the current independent intent is marked as a change intent; If the cosine similarity result of the current independent intent is greater than or equal to the preset cosine similarity threshold, the current independent intent will not be marked.

4. The method for dynamic correction of multi-turn dialogue intent based on a large language model according to claim 3, characterized in that, After obtaining the updated list of independent intents, it also includes: For each change intention, determine the total number of interaction steps in the historical multi-turn dialogue statements based on the current change intention; Based on the total number of interaction steps of the current change intention and the cosine similarity result, determine the progress metric of the current change intention; Based on the progress metric of each transformation intention, the transformation intentions are prioritized to obtain the first priority response sequence corresponding to each transformation intention. The remaining independent intentions in the updated list of independent intentions, excluding each transformation intention, are prioritized to obtain the second priority response sequence. The first priority response sequence and the second priority response sequence are concatenated to obtain the final priority response sequence. The step of outputting the response result of the target user's current round of dialogue based on the updated list of independent intents includes: Output the response results for the target user's current round of dialogue statements based on the final priority response sequence and the updated list of independent intents.

5. The method for dynamic correction of multi-turn dialogue intent based on a large language model according to claim 4, characterized in that, The step of outputting the response results for the target user's current round of dialogue based on the final priority response sequence and the updated list of independent intents includes: Based on the updated list of independent intents, generate response sub-results for each independent intent; The response sub-results corresponding to each independent intent are input into the intent conflict detection model, so that the intent conflict detection model extracts the semantic vector of each response sub-result, compares the semantic vectors of each response sub-result, and outputs several semantic conflict sets; wherein, the semantic conflict set includes several independent intent response sub-results that identify semantic conflicts. For each set of semantic conflicts, semantic vector fusion is performed on each independent intent response sub-result that identifies semantic conflicts to generate the conflict-free response sub-result corresponding to the current set of semantic conflicts. Based on the final priority response sequence, the independent intent response sub-results that do not conflict and the deconflict response sub-results are concatenated to output the response result of the target user's current round of dialogue.

6. The method for dynamic correction of multi-turn dialogue intent based on a large language model according to claim 5, characterized in that, The construction of the intent conflict detection model includes: Construct an intent conflict detection training set; wherein, the intent conflict detection training set includes: several intent conflict detection training subsets, each intent conflict detection training subset including several intent response text samples and a set of real semantic conflict samples of the current intent conflict detection training subset; Construct an initial intent conflict detection model; The initial intent conflict detection model is iteratively trained using a subset of intent conflict detection training data as input and a set of predicted semantic conflict samples corresponding to the subset of intent conflict detection training data as output, until the initial intent conflict detection model converges, thereby generating the intent conflict detection model. In each iteration, the initial intent conflict detection model semantically encodes each intent response text sample in the intent conflict detection training subset, generating a semantic encoding vector for each intent response text sample. It then calculates the cosine similarity of each semantic encoding vector, calculates the feature cross-degree between each semantic encoding vector, and outputs the feature cross-degree result. Based on the cosine similarity and feature cross-degree results, it outputs a predicted semantic conflict sample set. Finally, it generates the current iteration loss of the initial intent conflict detection model based on the predicted and actual semantic conflict sample sets, and adjusts the parameters of the initial intent conflict detection model using this current iteration loss.

7. The method for dynamic correction of multi-turn dialogue intent based on a large language model according to claim 6, characterized in that, For each semantic conflict set, semantic vector fusion is performed on each independent intent response sub-result that identifies a semantic conflict to generate a conflict-free response sub-result corresponding to the current semantic conflict set, including: For each set of semantic conflicts, the independent intent response sub-results with semantic conflicts for each identifier are transformed into semantic conflict encoding vectors; One semantic conflict coding vector is randomly selected from each semantic conflict coding vector as the selected semantic conflict coding vector, and the remaining unselected semantic conflict coding vectors are used as calibration semantic conflict coding vectors. For each calibration semantic conflict coding vector, compare the current calibration semantic conflict coding vector with the selected semantic conflict coding vector, and determine the first semantic conflict coding vector segment and the first semantically consistent coding vector segment of the current calibration semantic conflict coding vector and the selected semantic conflict coding vector, as well as the second semantic conflict coding vector segment and the second semantically consistent coding vector segment of the selected semantic conflict coding vector. A first calibration semantic vector is generated based on the first semantically conflicting coded vector segment and the second semantically consistent coded vector segment. A second calibration semantic vector is generated based on the second semantically conflicting coded vector segment and the first semantically consistent coded vector segment. A conflict-free semantic coded vector to be calibrated is generated based on the first semantically conflicting coded vector segment and the second semantically conflicting coded vector segment. The first calibration semantic vector and the second calibration semantic vector are used to calibrate the conflict-free semantic encoding vector to be calibrated, thereby generating a conflict-free semantic encoding vector. The first semantically consistent encoded vector segment and the second semantically consistent encoded vector segment are semantically fused to generate a fused semantic encoded vector. Based on the decontamination semantic encoding vector and the fusion semantic encoding vector, generate the decontamination response sub-result corresponding to the current semantic conflict set.

8. The method for dynamic correction of multi-turn dialogue intent based on a large language model according to claim 7, characterized in that, Before concatenating the independent intent response sub-results that do not conflict with each other and the deconflict response sub-results according to the final priority response sequence, and outputting the response result for the target user's current round of dialogue, the following steps are also included: Retrieve the previous round of dialogue from the target user's historical multi-turn dialogue statements; The step of concatenating the independent intent response sub-results without conflict and the de-conflict response sub-results according to the final priority response sequence, and outputting the response result of the target user's current round of dialogue, includes: Based on the final priority response sequence, the independent intent response sub-results that do not conflict and the deconflict response sub-results are concatenated to generate the initial response result for the current round of dialogue statements; Based on the initial response and the previous dialogue statement, generate the first semantic coherence result between the initial response and the previous dialogue statement; If the first semantic coherence result is not less than the preset semantic coherence threshold, the response result of the target user's current round of dialogue is output according to the initial response result; If the first semantic coherence result is less than a preset semantic coherence threshold, the semantic coherence fine-tuning operation is repeated until the second semantic coherence result meets the condition that the second semantic coherence result is not less than the preset semantic coherence threshold. Then, the response result of the target user's current round of dialogue is output according to the updated initial response result. The semantic coherence fine-tuning operation includes: inputting the initial response result, the previous round of dialogue, and the first semantic coherence result into the semantic coherence fine-tuning model so that the semantic coherence fine-tuning model outputs the updated initial response result and calculates the second semantic coherence result between the updated initial response result and the previous round of dialogue.

9. The method for dynamic correction of multi-turn dialogue intent based on a large language model according to claim 8, characterized in that, The construction of the semantic coherence fine-tuning model includes: Construct a training set for a semantic coherence fine-tuning model; wherein the training set for the semantic coherence fine-tuning model includes: several semantic text samples, associated text samples of each semantic text sample, and semantic coherence results between each semantic text sample and associated text samples. An initial semantic coherence fine-tuning model is constructed. The initial semantic coherence fine-tuning model is iteratively trained using the training set of the initial semantic coherence fine-tuning model until it converges, thus generating a semantic coherence fine-tuning model. In each iteration of training, the initial semantic coherence fine-tuning model extracts features from the current semantic text sample and its associated text samples to obtain the first feature of the current semantic text sample and the second feature of its associated text samples. Based on the semantic coherence result between the current semantic text sample and its associated text samples and the second feature, the first feature is fine-tuned to generate an updated first feature. An updated semantic text sample of the current semantic text sample is then generated based on the updated first feature.

10. A multi-turn dialogue intent dynamic correction system based on a large language model, characterized in that, The system includes: a dialogue acquisition module, a change of intent detection module, a change of intent update module, and a dialogue output module; The dialogue acquisition module is used to acquire the target user's historical multi-turn dialogue statements and the current round of dialogue statements; The transition intent detection module is used to generate an independent intent list for the target user's current round of dialogue statements based on the current round of dialogue statements, and to determine the historical state vector of each independent intent for the target user's current round of dialogue statements based on historical multi-round dialogue statements; wherein, the independent intent list includes several independent intents, a unique identifier for each independent intent, an initial context vector for each independent intent, and a current state vector for each independent intent; based on the historical state vector and the current state vector of each independent intent, each independent intent with a transition intent is determined, and each independent intent with a transition intent is marked as a transition intent in the independent intent list; The transition intent update module is used to determine the semantic vector of the current transition intent based on the current round of dialogue statements for each transition intent, update the initial context vector based on the semantic vector to obtain the updated context vector of the current transition intent, and update the list of independent intents of the current round of dialogue statements based on the updated context vectors of each transition intent to obtain the updated list of independent intents. The dialogue output module is used to output the response result of the target user's current round of dialogue statements based on the updated list of independent intents.

Images