AI conversation optimization method and system based on multi-level intent understanding

Abstract

The application is suitable for the technical field of dialogue optimization, and particularly relates to an AI dialogue optimization method and system based on multi-level intention understanding. The method comprises generating multiple subsequent dialogue node sequences based on a dialogue intention state vector, calculating a response achievement probability for each subsequent dialogue node sequence, and forming a candidate dialogue evolution track set. The response achievement probabilities of different candidate dialogue evolution tracks at the same node position are compared, node positions with a probability difference exceeding a preset probability threshold are identified, and a divergence node set is generated. An intervention intensity parameter is determined based on the divergence node set, the response achievement probability of the candidate dialogue evolution track is corrected, and an optimized dialogue response is generated. The application realizes node-level comparison and quantifiable adjustment of the dialogue path, so that track selection no longer depends on a single overall probability, but is dynamically corrected based on intention consistency, thereby improving the stability, consistency and goal achievement efficiency of the dialogue response.

Description

Technical Field

[0001] This invention relates to the field of dialogue optimization technology, and in particular to an AI dialogue optimization method and system based on multi-level intent understanding. Background Technology

[0002] As the application of large language models in dialogue systems continues to deepen, generating response text based on semantic understanding has become the mainstream technical approach. Existing dialogue systems typically achieve automatic question answering or intelligent customer service by semantically encoding the current dialogue context and directly generating the next round of response content. However, user intent has dynamic characteristics during multi-round dialogues, especially in the case of topic shifts, target adjustments, and emotional fluctuations. Single semantic matching or overall probability ranking is difficult to accurately reflect the actual dialogue direction.

[0003] To address this issue, existing technologies typically generate several candidate responses directly based on the current context, sort them according to overall semantic similarity or model output probability, and select the response with the highest probability or score as the final output. While this approach is feasible, it lacks sufficient granularity, failing to pinpoint "at which node the divergence occurs," and it also lacks a process for assessing the intensity of the divergence. This results in a somewhat rigid response process, operating in a zero-to-one state with excessively coarse granularity. Therefore, the technical problem this invention aims to solve is how to identify key divergence nodes in the candidate dialogue evolution trajectory and calculate the intent deviation degree based on these nodes, thereby finer-tuning and correcting the response content and reducing granularity. Summary of the Invention

[0004] The purpose of this invention is to provide an AI dialogue optimization method and system based on multi-level intent understanding to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: An AI dialogue optimization method based on multi-level intent understanding, the method comprising: Extract topic intent features, target intent features, and emotional intent features from multi-turn dialogue text, and fuse them to generate a dialogue intent state vector; Multiple subsequent dialogue node sequences are generated based on the dialogue intent state vector, and the response achievement probability is calculated for each subsequent dialogue node sequence to form a set of candidate dialogue evolution trajectories. Compare the response achievement probabilities of different candidate dialogue evolution trajectories at the same node position, identify node positions where the probability difference exceeds a preset probability threshold, and generate a set of divergent nodes; Based on the set of divergence nodes, calculate the difference value between the current dialogue intent state vector and the intent of each candidate dialogue evolution trajectory at the corresponding node position, and generate an intervention intensity parameter based on the intent difference value. The response achievement probability of the candidate dialogue evolution trajectory is corrected according to the intervention intensity parameter. The candidate dialogue evolution trajectory with the highest corrected probability is selected as the target dialogue evolution trajectory, and an optimized dialogue response is generated based on the target dialogue evolution trajectory.

[0006] As a further aspect of the present invention, the step of extracting topic intent features, target intent features, and emotional intent features from multi-turn dialogue text and fusing them to generate a dialogue intent state vector includes: The text of multi-turn dialogues is marked with turn number and user speech is distinguished from system speech. Extract keywords from user speech, count the frequency of each keyword in different rounds, and generate a topic intent vector; Identify target expressions in user speech, extract the target type and its location, and generate a target intent vector; Calculate the emotional polarity and emotional intensity values ​​of user utterances to generate an emotional intent vector; The topic intent vector, target intent vector, and emotional intent vector are concatenated and merged in a fixed dimensional order to generate a dialogue intent state vector.

[0007] As a further aspect of the present invention, the step of generating multiple subsequent dialogue node sequences based on the dialogue intent state vector, and calculating the response achievement probability for each subsequent dialogue node sequence to form a candidate dialogue evolution trajectory set includes: Multiple drafts of subsequent dialogue nodes are constructed based on the dialogue intent state vector. Each draft of a subsequent dialogue node contains topic information and guiding statements. Extract the topic information from the drafts of each subsequent dialogue node and number and mark the topic information. Node connections are constructed based on the logical order of the topic information, forming multiple sequences of subsequent dialogue nodes; Calculate the intent transfer value between adjacent nodes in each subsequent dialogue node sequence; The probability of achieving a response is calculated for each subsequent dialogue node sequence based on the intent transfer value and the target intent matching value, and the node sequence and the probability of achieving a response are combined to generate a candidate dialogue evolution trajectory.

[0008] As a further aspect of the present invention, the step of comparing the response achievement probabilities of different candidate dialogue evolution trajectories at the same node position, identifying node positions where the probability difference exceeds a preset probability threshold, and generating a set of divergent nodes includes: Establish node position indexes for the subsequent dialogue node sequences in each candidate dialogue evolution trajectory; Extract the response achievement probability of each candidate dialogue evolution trajectory at the same node position; Calculate the probability difference of different candidate dialogue evolution trajectories at the same node position; Determine whether the probability difference exceeds a preset probability threshold; Nodes whose positions exceed a preset probability threshold are marked as branch nodes, and a set of branch nodes is generated.

[0009] As a further aspect of the present invention, the step of calculating the intention difference value between the current dialogue intention state vector and the intention difference value of each candidate dialogue evolution trajectory at the corresponding node position based on the set of divergence nodes, and generating an intervention intensity parameter based on the intention difference value, includes: Extract the thematic intent component, target intent component, and emotional intent component corresponding to the divergence node; Extract the corresponding components from the dialogue intent state vector and construct the current intent component matrix; Extract the intent components corresponding to divergence nodes from the candidate dialogue evolution trajectory, and construct the trajectory intent component matrix; Calculate the difference between the current intent component matrix and the trajectory intent component matrix, and normalize the intent difference. Intervention intensity parameters are generated based on the normalized difference values.

[0010] As a further aspect of the present invention, the step of correcting the response achievement probability of the candidate dialogue evolution trajectory according to the intervention intensity parameter, selecting the candidate dialogue evolution trajectory with the highest corrected probability as the target dialogue evolution trajectory, and generating an optimized dialogue response based on the target dialogue evolution trajectory includes: The response achievement probability of candidate dialogue evolution trajectories is weighted and corrected based on the intervention intensity parameter; Calculate the corrected probability of achieving the response; Select the candidate dialogue evolution trajectory with the highest probability of achieving the corrected response as the target dialogue evolution trajectory; Generate an optimized dialogue response based on the sequence of subsequent dialogue nodes in the target dialogue evolution trajectory; Output the optimized dialogue response.

[0011] The present invention also provides an AI dialogue optimization system based on multi-level intent understanding, the system comprising: The intent state analysis module is used to extract topic intent features, target intent features, and emotional intent features from multi-turn dialogue text, and fuse them to generate a dialogue intent state vector; The dialogue trajectory generation module is used to generate multiple subsequent dialogue node sequences based on the dialogue intent state vector, and calculate the response achievement probability for each subsequent dialogue node sequence to form a candidate dialogue evolution trajectory set. The divergence node statistics module is used to compare the response achievement probabilities of different candidate dialogue evolution trajectories at the same node position, identify node positions where the probability difference exceeds a preset probability threshold, and generate a set of divergence nodes. The intervention parameter generation module is used to calculate the intention difference value between the current dialogue intention state vector and the intention difference value of each candidate dialogue evolution trajectory at the corresponding node position based on the set of divergence nodes, and generate intervention intensity parameters based on the intention difference value. The target trajectory generation module is used to correct the response achievement probability of the candidate dialogue evolution trajectory according to the intervention intensity parameter, select the candidate dialogue evolution trajectory with the highest corrected probability as the target dialogue evolution trajectory, and generate an optimized dialogue response based on the target dialogue evolution trajectory.

[0012] As a further aspect of the present invention, the intent state analysis module includes: The text tagging unit is used to tag the turn of a multi-turn dialogue text and distinguish between user speech and system speech; The topic analysis unit is used to extract keywords from user discourse, count the frequency of each keyword in different rounds, and generate topic intent vectors. The target analysis unit is used to identify target expressions in user speech, extract the target type and its location, and generate a target intent vector. The sentiment analysis unit is used to calculate the sentiment polarity and sentiment intensity values ​​of user utterances and generate sentiment intent vectors. The vector fusion unit is used to concatenate and fuse the topic intent vector, target intent vector, and emotional intent vector in a fixed dimensional order to generate a dialogue intent state vector.

[0013] As a further embodiment of the present invention, the dialogue trajectory generation module includes: The draft generation unit is used to construct multiple drafts of subsequent dialogue nodes based on the dialogue intent state vector. Each draft of a subsequent dialogue node contains topic information and guiding statements. The marking execution unit is used to extract topic information from the drafts of each subsequent dialogue node and to number and mark the topic information. The node sequence generation unit is used to construct node connection relationships based on the logical order between topic information, forming multiple subsequent dialogue node sequences. The transfer value calculation unit is used to calculate the intent transfer value between adjacent nodes in each subsequent dialogue node sequence; The probability generation insertion unit is used to calculate the response achievement probability of each subsequent dialogue node sequence based on the intent transfer value and the target intent matching value, and to combine the node sequence and the response achievement probability to generate the candidate dialogue evolution trajectory.

[0014] As a further aspect of the present invention, the divergence node statistics module includes: The index creation unit is used to establish node position indexes for the sequence of subsequent dialogue nodes in each candidate dialogue evolution trajectory; The probability extraction unit is used to extract the response achievement probability of each candidate dialogue evolution trajectory at the same node position; The probability difference calculation unit is used to calculate the probability difference of different candidate dialogue evolution trajectories at the same node position; The comparison and judgment unit is used to determine whether the probability difference exceeds a preset probability threshold. The marking statistics unit is used to mark the node positions that exceed a preset probability threshold as branch nodes and generate a set of branch nodes.

[0015] Compared with the prior art, the beneficial effects of the present invention are: This invention constructs a multi-level intent state vector, structurally fusing topic intent, target intent, and emotional intent. After generating multiple candidate dialogue evolution trajectories, it compares the response achievement probabilities at the same node positions, identifying a set of divergent nodes where the probability difference exceeds a threshold. Based on the intent components corresponding to the divergent nodes, it calculates the difference between the current intent state and each candidate trajectory, generating an intervention intensity parameter to weight and correct the response achievement probabilities of the candidate trajectories. This invention achieves node-level comparison and quantifiable adjustment of dialogue paths, making trajectory selection no longer dependent on a single overall probability, but based on intent... Figure 1 The consistency of the dialogue response is dynamically adjusted to improve its stability, consistency, and the efficiency of goal achievement. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention.

[0017] Figure 1 A flowchart illustrating the AI ​​dialogue optimization method based on multi-level intent understanding provided in this embodiment of the invention.

[0018] Figure 2 This is a block diagram illustrating the structural composition of an AI dialogue optimization system based on multi-level intent understanding, provided in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0020] Figure 1 The flowchart illustrates an AI dialogue optimization method based on multi-level intent understanding provided in this embodiment of the invention. The method includes: Step S100: Extract topic intent features, target intent features, and emotional intent features from the multi-turn dialogue text, and fuse them to generate a dialogue intent state vector; This process extracts multi-level intent features from multi-turn dialogue text between the user and the system, and fuses them to generate a dialogue intent state vector. Specifically, the multi-turn dialogue text is segmented by turn to distinguish between user utterances and system utterances, and intent analysis is performed only on user utterances. For user utterances, topic intent features are extracted, such as identifying topics like "price," "performance," "discount," and "comparison" through keyword statistics and topic dictionary matching, and calculating the frequency of each topic word in all turns and its most recent occurrence weight to form a topic intent vector T=(t1,t2,…,tn). Simultaneously, target intent features in user expressions are identified, such as phrases like "want to buy," "learn more," and "compare models," and a target intent vector G=(g1,g2,…,gm) is generated based on a preset target type set. Furthermore, an emotion dictionary is used to score the emotion polarity and calculate the intensity of user utterances, generating an emotion intent vector E=(e_pos,e_neg,e_intensity). Finally, T, G, and E are concatenated or weighted according to preset weights to form a unified dialogue intent state vector V=(T,G,E), which is used to represent the comprehensive intent state of the current dialogue.

[0021] Step S200: Generate multiple subsequent dialogue node sequences based on the dialogue intent state vector, and calculate the response achievement probability for each subsequent dialogue node sequence to form a candidate dialogue evolution trajectory set; After obtaining the dialogue intent state vector V, multiple subsequent dialogue node sequences are generated based on this vector, and the response achievement probability is calculated for each subsequent dialogue node sequence, forming a set of candidate dialogue evolution trajectories. Specifically, the top k topic categories are selected according to the weight sorting in the topic intent vector T, and a corresponding dialogue node draft is generated for each topic. Each node draft contains a topic identifier and guiding statements. Then, according to the preset topic transfer rules, multiple node drafts are combined in logical order to form several subsequent dialogue node sequences. For example, sequences such as S1=[performance introduction, price description, purchase confirmation] and S2=[promotional activities, limited-time reminder, purchase confirmation] can be formed. For each node sequence, the intent transfer similarity between adjacent nodes is calculated, and combined with the matching value between the endpoint node and the target intent vector G, the response achievement probability P of the sequence is calculated. This probability can be multiplied, and finally, each node sequence and its corresponding probability are combined to form a set of candidate dialogue evolution trajectories.

[0022] Step S300: Compare the response achievement probabilities of different candidate dialogue evolution trajectories at the same node position, identify the node positions where the probability difference exceeds a preset probability threshold, and generate a set of divergent nodes; After forming a set of candidate dialogue evolution trajectories, the response achievement probabilities of different candidate dialogue evolution trajectories at the same node positions are compared to identify the set of divergent nodes. Specifically, all candidate dialogue evolution trajectories are aligned so that each node sequence is indexed and marked at the same position, such as position 1, position 2, and position 3. Then, the cumulative response achievement probability value or local transition probability value corresponding to different trajectories at the same node position is extracted, and the probability difference ΔP between each trajectory is calculated. For example, at node position 2, the probability of trajectory S1 is 0.62, and the probability of trajectory S2 is 0.41, then ΔP = 0.21. If ΔP exceeds a preset probability threshold (e.g., 0.15), then the node position is marked as a divergent node. All node positions are traversed to identify all positions that meet the conditions, forming a set of divergent nodes D = {d1, d2, ...}. This set is used to identify key node positions where the dialogue development direction diverges significantly.

[0023] Step S400: Based on the set of divergence nodes, calculate the difference value between the current dialogue intent state vector and the intent of each candidate dialogue evolution trajectory at the corresponding node position, and generate an intervention intensity parameter based on the intent difference value; After obtaining the set of divergence nodes D, the intention difference values ​​between the current dialogue intention state vector V and each candidate dialogue evolution trajectory at the corresponding node positions are calculated based on the set of divergence nodes, and an intervention intensity parameter is generated based on the difference values. Specifically, for each divergence node position di, the intention vector Vi_track of the corresponding node at that position, as well as the topic, goal, and emotion components corresponding to the current dialogue intention state vector V_current, are extracted. The difference value Diff_i between V_current and Vi_track is calculated using a vector difference calculation method (e.g., cosine distance or Euclidean distance). The difference values ​​of all divergence nodes are normalized to obtain a standardized difference sequence. The intervention intensity parameter K is generated based on the magnitude of the difference values, for example, K = average(Diff_i) × adjustment coefficient. The intervention intensity parameter reflects the degree of deviation between the current true intention state and the predicted intention of the candidate path.

[0024] Step S500: Correct the response achievement probability of the candidate dialogue evolution trajectory according to the intervention intensity parameter, select the candidate dialogue evolution trajectory with the highest probability after correction as the target dialogue evolution trajectory, and generate an optimized dialogue response based on the target dialogue evolution trajectory; After obtaining the intervention intensity parameter K, the response achievement probabilities of the candidate dialogue evolution trajectories are corrected, and the candidate dialogue evolution trajectory with the highest corrected probability is selected as the target dialogue evolution trajectory. Specifically, the original response achievement probability P_i of each candidate trajectory is weighted and corrected with the intervention intensity parameter K, for example, by calculating the corrected probability P_i' = P_i × (1 - K × S_i), where S_i is the deviation coefficient derived from the intent difference value. After calculating the corrected probabilities of all trajectories, the trajectory with the largest P_i' is selected as the target dialogue evolution trajectory. Finally, the final optimized dialogue response is generated based on the topic information and guiding statements corresponding to each node in the target trajectory and output to the user terminal, realizing adaptive dialogue optimization based on multi-level intent understanding.

[0025] Regarding step S100, the step of extracting topic intent features, target intent features, and emotional intent features from the multi-turn dialogue text, and fusing them to generate a dialogue intent state vector, includes: The text of multi-turn dialogues is marked with turn number and user speech is distinguished from system speech. Extract keywords from user speech, count the frequency of each keyword in different rounds, and generate a topic intent vector; Identify target expressions in user speech, extract the target type and its location, and generate a target intent vector; Calculate the emotional polarity and emotional intensity values ​​of user utterances to generate an emotional intent vector; The topic intent vector, target intent vector, and emotional intent vector are concatenated and merged in a fixed dimensional order to generate a dialogue intent state vector.

[0026] In one embodiment of the technical solution of this invention, a specific process for generating dialogue intent state vectors is provided. This process actually consists of four steps: the generation of topic intent vectors, target intent vectors, and emotional intent vectors, plus a fusion process. Specifically, the acquired multi-turn dialogue text is first processed in a structured manner. Specifically, the complete dialogue is divided into turns according to chronological order, with each "user speech + system response" constituting a dialogue turn, and each turn is assigned a unique turn number i. Subsequently, the text is differentiated by role based on the source identifier of the speech, and user speech and system speech are stored in different data structures respectively. This step only performs intent analysis on user speech; therefore, system speech is marked during the data preprocessing stage but does not participate in vector calculation. Simultaneously, the user speech undergoes text cleaning, including removing meaningless symbols, standardizing capitalization, and standardizing numerical expressions. After the above processing, a well-structured set of turn-marked user speech, U={U1,U2,…,Un}, is obtained, laying the foundation for subsequent extraction of topic intent features, target intent features, and emotional intent features.

[0027] After round labeling, keywords are extracted from the text in the user discourse set U, and a topic intent vector is generated. Specifically, each user discourse is split into a set of terms using word segmentation, and stop words are removed while retaining keywords with actual semantic meaning. Keywords are matched against a predefined topic dictionary, which includes several predefined topic categories, such as "price," "performance," "after-sales," and "discounts." For each category, the frequency fi of its corresponding keywords in each round is calculated, and a time weight factor wi (e.g., higher weight for recent rounds) is introduced to calculate a weighted frequency value (summing the product of weight and frequency). The weighted frequencies of each topic category are arranged in a fixed order to form a topic intent vector T=(F1,F2,…,Fk), which is then normalized so that each component is between 0 and 1, representing the intensity distribution of the user's current topic focus.

[0028] This process identifies target expressions in user utterances and generates target intent vectors. Specifically, a target expression rule base is constructed, including typical expressions representing purchase intention, comparison intention, consultation intention, and hesitation / wait-and-see intention, such as "I want to buy," "Can you make it cheaper?", "Compare," and "Think about it." Rule matching is performed on the user utterance set U to identify utterance fragments that conform to the target expression rules, and their round position pi is recorded. Based on the matching results, the corresponding target types are counted, and weights are assigned based on their position, with higher weights for later rounds. After weighted statistical analysis of each target type, a target intent vector G=(g1,g2,…,gm) is generated, where each dimension represents the strength of a certain target type. This vector reflects the user's specific behavioral tendencies at the current stage of the conversation.

[0029] After generating the topic intent vector and target intent vector, sentiment analysis is performed on user utterances to generate a sentiment intent vector. Specifically, a sentiment lexicon is established, including positive sentiment words, negative sentiment words, and intensity modifiers. The sentiment lexicon is matched word-by-word through the user utterances, calculating the frequency of positive sentiment words (npos) and the frequency of negative sentiment words (nneg), and weighting the sentiment score based on the intensity modifiers. The sentiment polarity value Epol can be calculated as (npos-nneg) / (npos+nneg+1), and the sentiment intensity value Eint can be calculated as the proportion of the total number of sentiment words to the text length. For multi-turn dialogues, higher weights can be assigned to the most recent turns. The final generated sentiment intent vector is E=(Epol,Eint), where Epol represents the sentiment direction and Eint represents the sentiment intensity, used to characterize the user's emotional state.

[0030] After obtaining the topic intent vector T, target intent vector G, and emotional intent vector E, the three types of vectors are concatenated and fused to generate a dialogue intent state vector V. Specifically, T, G, and E are first standardized to ensure that each vector component is within a uniform numerical range. Then, they are concatenated in a fixed dimensional order, for example, V=(T1,…,Tk,G1,…,Gm,Epol,Eint). To avoid any one type of feature excessively influencing the overall vector, fusion weights α, β, and γ can be set to weight T, G, and E respectively, i.e., V=(αT,βG,γE). This dialogue intent state vector V, as a comprehensive expression of the current dialogue, includes both the topic focus structure and the target inclination and emotional state, providing a unified input basis for subsequently generating candidate dialogue evolution trajectories.

[0031] Regarding step S200, the step of generating multiple subsequent dialogue node sequences based on the dialogue intent state vector, and calculating the response achievement probability for each subsequent dialogue node sequence to form a candidate dialogue evolution trajectory set includes: Multiple drafts of subsequent dialogue nodes are constructed based on the dialogue intent state vector. Each draft of a subsequent dialogue node contains topic information and guiding statements. Extract the topic information from the drafts of each subsequent dialogue node and number and mark the topic information. Node connections are constructed based on the logical order of the topic information, forming multiple sequences of subsequent dialogue nodes; Calculate the intent transfer value between adjacent nodes in each subsequent dialogue node sequence; The probability of achieving a response is calculated for each subsequent dialogue node sequence based on the intent transfer value and the target intent matching value, and the node sequence and the probability of achieving a response are combined to generate a candidate dialogue evolution trajectory.

[0032] In one example of the technical solution of this invention, the generation process of the candidate dialogue evolution trajectory is described, and multiple drafts of subsequent dialogue nodes are constructed based on the obtained dialogue intent state vector V. Specifically, the dialogue intent state vector V is first input into a preset dialogue strategy template library. This template library is structured and stored according to different topic categories and target types. For example, when the "price" component in the topic intent vector is high and the "purchase intention" component in the target intent vector is high, the "price guidance + transaction confirmation" template is called first. The system sorts the components in V according to their numerical values, selects the top N topic categories as the priority development direction, and generates corresponding guidance statements for each topic, such as "Do you need to know about the current preferential policies?" or "Do you want to compare the parameters of different models?". Each topic information and its corresponding guidance statement constitute a draft of a subsequent dialogue node, denoted as Ni=(Ti,Si), where Ti represents the topic information number and Si represents the guidance statement text. By changing the tone or order of the guidance, multiple different node drafts can be formed, providing a set of candidate nodes for the subsequent construction of branch sequences.

[0033] After generating multiple drafts of subsequent dialogue nodes, the topic information in each draft is extracted and numbered. Specifically, the topic information Ti in each draft Ni is matched with a preset topic category library to confirm its category, and a unique number is assigned to each category, such as T1 for price, T2 for performance, and T3 for after-sales service. Drafts of nodes with the same topic category but different guiding statements are assigned the same topic number but different node sequence numbers, such as T1-1 and T1-2. Through this numbering method, all draft nodes are constructed into a structured node set N={N1,N2,…,Nk}, and the topic number, target association type, and sentiment tendency value corresponding to each node are recorded. This numbering and marking process ensures traceability and consistency in the subsequent logical sequence construction and probability calculation, avoiding semantic confusion of nodes.

[0034] The system constructs node connections based on the logical order of topic information, forming multiple sequences of subsequent dialogue nodes. Specifically, it predefines the logical order rules of topics, such as "needs confirmation, parameter description, price communication, and transaction guidance" as the logic of a regular sales process. When the topic number of a draft node meets the logical predecessor condition, it can establish a connection with subsequent topic nodes. For example, node T2 (performance-related) can serve as the predecessor of node T1 (price-related). The system establishes directed edges in the node set N according to the above logical rules, forming a directed graph structure G=(N,E). Subsequently, by traversing all paths in the graph structure that meet the initial conditions, it generates multiple sequences of subsequent dialogue nodes Pj=(Nj1,Nj2,…,Njn). Each node sequence represents a possible dialogue development path, and its length can be determined according to a preset round limit, such as being limited to 3 to 5 nodes, thus forming multiple candidate branch structures.

[0035] After forming multiple subsequent dialogue node sequences, the intent transfer value between adjacent nodes in each sequence is calculated. Specifically, for adjacent nodes Nji and Nj(i+1) in sequence Pj, the topic vector component, target vector component, and emotion vector component corresponding to their topic numbers are extracted to construct node intent vectors Vi and Vi+1. The transfer value between the two vectors is calculated, which can be transformed using cosine similarity or Euclidean distance, for example, the transfer value Mji=cos(Vi,Vi+1). This transfer value reflects the degree of intent coherence when guiding from the previous node to the next node; the closer the value is to 1, the smoother the logical connection. The transfer values ​​of the entire node sequence are multiplied or averaged to obtain the overall sequence transfer coherence index Mj, which is used to represent the coherence and acceptability of the dialogue path at the intent level.

[0036] After obtaining the transition coherence index Mj for each node sequence, the response achievement probability is calculated by combining it with the target intent matching value. Specifically, the target intent vector G is extracted from the dialogue intent state vector V and matched with the target component of the terminal node of the node sequence to calculate the target intent matching value Qj, for example, Qj=cos(G,Gj_end). Then, a response achievement probability function is constructed, which can be defined as Pj=Mj×Qj, where Mj represents the sequence coherence and Qj represents the target matching degree. Pj is normalized so that the sum of the probabilities of all sequences is 1. Finally, each node sequence Pj is combined with its corresponding response achievement probability Pj to form the candidate dialogue evolution trajectory Aj.

[0037] Regarding step S300, the step of comparing the response achievement probabilities of different candidate dialogue evolution trajectories at the same node position, identifying node positions where the probability difference exceeds a preset probability threshold, and generating a set of divergent nodes includes: Establish node position indexes for the subsequent dialogue node sequences in each candidate dialogue evolution trajectory; Extract the response achievement probability of each candidate dialogue evolution trajectory at the same node position; Calculate the probability difference of different candidate dialogue evolution trajectories at the same node position; Determine whether the probability difference exceeds a preset probability threshold; Nodes whose positions exceed a preset probability threshold are marked as branch nodes, and a set of branch nodes is generated.

[0038] In one example of the technical solution of this invention, the generation process of divergent nodes is described. Let the set of candidate dialogue evolution trajectories be A = {A1, A2, ..., Am}. Each trajectory Aj contains a subsequent dialogue node sequence Pj = (Nj1, Nj2, ..., Njn) and its overall response achievement probability Pj. To achieve structural alignment between different trajectories, all node sequences are first standardized in length. When the length of a sequence is less than the preset maximum number of rounds L, a termination placeholder node N0 is added to its end. Subsequently, each node sequence is assigned a unified node position index i = 1, 2, ..., L according to the generation order, where i represents the i-th node position in the dialogue development. Through this node position index, different candidate dialogue evolution trajectories can be structurally aligned, making Nji in Aj and Nki in Ak comparable at the same position index i.

[0039] After establishing the node position index, the response achievement probability of each candidate dialogue evolution trajectory is extracted at the same node position. Specifically, although the overall response achievement probability Pj has been calculated for each trajectory, this embodiment further decomposes the trajectory into positions. For the i-th node position in trajectory Aj, the stage response achievement probability of the subsequence formed from the starting node to that node position is calculated, denoted as Pj(i). This probability can be obtained by accumulating the intent transition values ​​of the first i nodes and combining them with the target intent stage matching value.

[0040] After obtaining the probability matrix P, the probability difference between different candidate dialogue evolution trajectories at the same node position is calculated. Specifically, for each node position index i, the set of stage response achievement probabilities for all trajectories at that position is extracted as {P1(i), P2(i), ..., Pm(i)}. Then, the difference between the maximum and minimum values ​​in this set is calculated, denoted as ΔPi = max(Pj(i)) - min(Pj(i)). This difference represents the degree of differentiation in effects between different dialogue development paths at node position i. Variance or standard deviation can also be calculated to measure distribution dispersion, but in this embodiment, the range form is preferred to highlight the largest difference. If the value of ΔPi is large, it indicates a significant difference in the subsequent development potential between different trajectories at that node position, representing a key turning point in the differentiation of dialogue strategies.

[0041] After obtaining the probability difference ΔPi at each node location, it is determined whether the probability difference exceeds a preset probability threshold. Specifically, a probability threshold θ is preset, for example, θ=0.25. When ΔPi≥θ, it is considered that there is a significant divergence at that node location. The threshold θ can be determined based on historical dialogue data statistics, for example, by statistically analyzing the average probability difference between high-conversion and low-conversion dialogues at key nodes, and taking its mean or percentile value as the threshold benchmark. By comparing ΔPi with θ one by one, node locations with obvious path divergence significance in the dialogue development process can be screened out. This judgment process has clear numerical boundaries and does not rely on subjective judgment, thereby ensuring the objectivity and repeatability of divergence identification.

[0042] For node positions that satisfy the condition ΔPi≥θ, they are marked as branching nodes, and a branching node set D is generated. Specifically, let the set of node position indices that satisfy the condition be {ik}, then the corresponding branching node set is represented as D={Nik|ΔPik≥θ}. Simultaneously, the probability distribution of stage response achievement corresponding to each trajectory at each branching node is recorded for subsequent calculation of intent deviation. The branching node set D essentially represents the key decision points most likely to produce different outcomes during the dialogue evolution process; these nodes are the core locations for subsequent intervention intensity calculation and trajectory correction.

[0043] Regarding step S400, the step of calculating the intention difference value between the current dialogue intention state vector and the intention difference value of each candidate dialogue evolution trajectory at the corresponding node position based on the set of divergence nodes, and generating the intervention intensity parameter based on the intention difference value, includes: Extract the thematic intent component, target intent component, and emotional intent component corresponding to the divergence node; Extract the corresponding components from the dialogue intent state vector and construct the current intent component matrix; Extract the intent components corresponding to divergence nodes from the candidate dialogue evolution trajectory, and construct the trajectory intent component matrix; Calculate the difference between the current intent component matrix and the trajectory intent component matrix, and normalize the intent difference. Intervention intensity parameters are generated based on the normalized difference values.

[0044] After determining the differences in intent among the evolutionary trajectories of each candidate dialogue, intervention parameters need to be generated based on the calculated differences in intent, as follows: For the set D of divergence nodes generated in the previous steps, the corresponding topic intent component, target intent component, and emotional intent component are extracted for each divergence node. Specifically, let the position index of the divergence node be i, and the corresponding node be N(i). First, the intent expression content on which the node was generated is traced back, and keyword matching and target expression recognition are performed on the node text, mapping them to the topic intent space and the target intent space, respectively. For example, when the divergence node is "Do I need to know the package price?", the topic intent component corresponds to the "price consultation" dimension, and the target intent component corresponds to the "purchase decision promotion" dimension. At the same time, the emotional polarity value and intensity value are calculated based on the emotional tendency words of the node's statement to form the emotional intent component. The three components are represented as Ti, Gi, and Ei, respectively, all in vector form, and are consistent with the dimensions of the generated dialogue intent state vector.

[0045] After extracting the intent components of the divergence nodes, the corresponding components are extracted from the current dialogue intent state vector to construct the current intent component matrix. Specifically, let the current dialogue intent state vector be V=(T,G,E), where T is the topic intent vector, G is the target intent vector, and E is the emotion intent vector. For the divergence node position i, according to the intent dimension index determined in S41, the corresponding topic dimension value Ti′ is extracted from T, the corresponding target dimension value Gi′ is extracted from G, and the corresponding emotion dimension value Ei′ is extracted from E. The three components are arranged in a fixed order to form a column vector Vi=[Ti′,Gi′,Ei′]ᵀ. If there are multiple divergence nodes, the column vectors corresponding to each node are concatenated column by column to form the current intent component matrix M_current, which has a dimension of 3×k, where k is the number of divergence nodes. This matrix is ​​used to represent the intent expression intensity structure of the current dialogue state at key nodes.

[0046] The intent components corresponding to the divergence nodes are extracted from the candidate dialogue evolution trajectory set A, and a trajectory intent component matrix is ​​constructed. Specifically, for each candidate dialogue evolution trajectory Aj, the node Nji with the same index as the divergence node is located in its node sequence, and the same component extraction method is performed on this node to obtain the topic component Tji, target component Gji, and emotion component Eji at that position of the trajectory. The three are combined into a column vector Vj(i)=[Tji,Gji,Eji]ᵀ. If there are k divergence nodes, the corresponding column vector is extracted for each divergence node position and concatenated to form the trajectory intent component matrix M_j, which also has a dimension of 3×k. In this way, the current dialogue intent component matrix M_current and the trajectory intent component matrices M_j are completely consistent in structure and dimension, and have the conditions for direct numerical comparison.

[0047] After obtaining the current intent component matrix M_current and the trajectory intent component matrix M_j, the difference between the two matrices is calculated and normalized. Specifically, the matrix difference can be calculated by summing the squared differences element by element, i.e., D_j = Σ(M_current - M_j)², where the summation covers all 3×k elements. This difference value D_j reflects the degree of deviation between the current dialogue state and the candidate trajectory's intent at key divergence nodes. Since the component value ranges may differ in different dialogue scenarios, D_j is normalized to avoid the influence of units, for example, by using D_j_norm = D_j / (D_max + ε), where D_max is the maximum difference value in historical samples, and ε is a small constant to prevent division by zero. After normalization, the difference value is controlled between 0 and 1, with a larger value indicating a higher degree of deviation.

[0048] Intervention intensity parameters are generated based on the normalized difference values. Specifically, for each candidate dialogue evolution trajectory Aj, its normalized difference value is D_j_norm. The intervention intensity parameter Ij is set as a monotonic function of the difference value, for example, Ij = α × D_j_norm, where α is an adjustment coefficient used to control the upper limit of the intervention amplitude. When D_j_norm is large, it indicates that the current dialogue state differs significantly from the intent of the trajectory at key nodes, requiring strong intervention for probability correction; when D_j_norm is small, it indicates that the trajectory is highly consistent with the current intent, and intervention can be reduced. Finally, a corresponding intervention intensity parameter set I = {I1, I2, ..., Im} is generated for each candidate trajectory. The intervention intensity parameter set is the input parameter for the subsequent correction process.

[0049] Regarding step S500, the step of correcting the response achievement probability of the candidate dialogue evolution trajectory according to the intervention intensity parameter, selecting the candidate dialogue evolution trajectory with the highest corrected probability as the target dialogue evolution trajectory, and generating an optimized dialogue response based on the target dialogue evolution trajectory includes: The response achievement probability of candidate dialogue evolution trajectories is weighted and corrected based on the intervention intensity parameter; Calculate the corrected probability of achieving the response; Select the candidate dialogue evolution trajectory with the highest probability of achieving the corrected response as the target dialogue evolution trajectory; Generate an optimized dialogue response based on the sequence of subsequent dialogue nodes in the target dialogue evolution trajectory; Output the optimized dialogue response.

[0050] In one embodiment of the technical solution of this invention, the response achievement probability of the candidate dialogue evolution trajectory set is weighted and corrected based on the obtained set of intervention intensity parameters. To ensure the computability and monotonicity of the correction process, an exponential decay correction method or a linear suppression method can be adopted. For example, the correction function can be constructed as: P′=P×(1-β×I), where β is the correction ratio coefficient, with a value ranging from 0 to 1. When I is large, it indicates that the deviation between the current dialogue intention state and the trajectory at the divergence node is relatively high, so the response achievement probability is suppressed more significantly; when I is close to 0, the original probability remains basically unchanged. In this way, the difference in intention at the divergence node is directly converted into a probability adjustment factor, realizing a numerical mapping from "intention deviation" to "trajectory suppression".

[0051] After weighted correction, the set of corrected response achievement probabilities is calculated. To avoid imbalance in the sum of corrected probabilities, normalization can be performed, for example, using proportional normalization to ensure that the sum of corrected probabilities for all candidate dialogue evolution trajectories is 1. This normalization step ensures that the probabilities of each trajectory are within a uniform dimension and maintain their relative magnitudes. If the system sets a minimum probability threshold, trajectories below the threshold can be pruned and removed from the candidate set to reduce subsequent computational complexity. Finally, a stable and comparable set of corrected response achievement probabilities is obtained, providing a clear ranking basis for trajectory selection.

[0052] After obtaining the set of corrected response achievement probabilities, all candidate dialogue evolution trajectories are sorted, and the candidate dialogue evolution trajectory with the highest corrected response achievement probability is selected as the target dialogue evolution trajectory. Specifically, let A_k be the trajectory corresponding to the maximum corrected probability, then A_k is determined as the target dialogue evolution trajectory. This selection process is based on the principle of numerical maximization, avoiding subjective rule intervention and ensuring that the decision-making process is entirely driven by the aforementioned intent difference calculation and probability correction mechanism. Since the corrected probability has comprehensively considered the original response achievement capability and the degree of intent deviation, the selected target trajectory maintains both a high achievement capability and a high degree of consistency with the current dialogue intent state, thereby achieving optimal matching of the dialogue path.

[0053] After determining the target dialogue evolution trajectory A_k, an optimized dialogue response is generated based on its subsequent dialogue node sequence. The system first extracts the node's topic information, guiding statements, and contextual connection information, and then adjusts the statement expression based on the emotional intent component in the current dialogue intent state vector. For example, when the emotional intent component indicates that the user is hesitant, a buffer expression is added to the original node's guiding statement, such as "You can learn more about the details first." The optimized node content is then checked for consistency with the historical dialogue context to ensure topic continuity and logical coherence. Finally, an optimized dialogue response text that is consistent with the target trajectory and semantically complete is generated.

[0054] After generating the optimized dialogue response, it is encapsulated according to a preset output format and sent to the user terminal. Specifically, the response text can be converted into a standard character stream data structure, including the current round identifier, timestamp, and session identifier, to ensure the continuity of multi-round dialogue management. Before output, a semantic validity check can be performed to ensure there are no punctuation errors or semantic conflicts. Subsequently, the optimized dialogue response is transmitted to the user display interface through the dialogue interface module for real-time feedback.

[0055] Figure 2 The block diagram of the AI ​​dialogue optimization system based on multi-level intent understanding provided in the embodiments of the present invention, as a preferred embodiment of the technical solution of the present invention, also provides an AI dialogue optimization system 10 based on multi-level intent understanding, the AI ​​dialogue optimization system 10 based on multi-level intent understanding comprising: The intent state analysis module 11 is used to extract topic intent features, target intent features and emotional intent features from multi-turn dialogue text, and fuse them to generate a dialogue intent state vector; The dialogue trajectory generation module 12 is used to generate multiple subsequent dialogue node sequences based on the dialogue intent state vector, and calculate the response achievement probability for each subsequent dialogue node sequence to form a candidate dialogue evolution trajectory set. The divergence node statistics module 13 is used to compare the response achievement probabilities of different candidate dialogue evolution trajectories at the same node position, identify the node positions where the probability difference exceeds a preset probability threshold, and generate a set of divergence nodes. The intervention parameter generation module 14 is used to calculate the intention difference value between the current dialogue intention state vector and the evolution trajectory of each candidate dialogue at the corresponding node position based on the set of divergence nodes, and generate an intervention intensity parameter based on the intention difference value. The target trajectory generation module 15 is used to correct the response achievement probability of the candidate dialogue evolution trajectory according to the intervention intensity parameter, select the candidate dialogue evolution trajectory with the highest corrected probability as the target dialogue evolution trajectory, and generate an optimized dialogue response based on the target dialogue evolution trajectory.

[0056] Furthermore, the intent state analysis module 11 includes: The text tagging unit is used to tag the turn of a multi-turn dialogue text and distinguish between user speech and system speech; The topic analysis unit is used to extract keywords from user discourse, count the frequency of each keyword in different rounds, and generate topic intent vectors. The target analysis unit is used to identify target expressions in user speech, extract the target type and its location, and generate a target intent vector. The sentiment analysis unit is used to calculate the sentiment polarity and sentiment intensity values ​​of user utterances and generate sentiment intent vectors. The vector fusion unit is used to concatenate and fuse the topic intent vector, target intent vector, and emotional intent vector in a fixed dimensional order to generate a dialogue intent state vector.

[0057] Specifically, the dialogue trajectory generation module 12 includes: The draft generation unit is used to construct multiple drafts of subsequent dialogue nodes based on the dialogue intent state vector. Each draft of a subsequent dialogue node contains topic information and guiding statements. The marking execution unit is used to extract topic information from the drafts of each subsequent dialogue node and to number and mark the topic information. The node sequence generation unit is used to construct node connection relationships based on the logical order between topic information, forming multiple subsequent dialogue node sequences. The transfer value calculation unit is used to calculate the intent transfer value between adjacent nodes in each subsequent dialogue node sequence; The probability generation insertion unit is used to calculate the response achievement probability of each subsequent dialogue node sequence based on the intent transfer value and the target intent matching value, and to combine the node sequence and the response achievement probability to generate the candidate dialogue evolution trajectory.

[0058] Furthermore, the divergence node statistics module 13 includes: The index creation unit is used to establish node position indexes for the sequence of subsequent dialogue nodes in each candidate dialogue evolution trajectory; The probability extraction unit is used to extract the response achievement probability of each candidate dialogue evolution trajectory at the same node position; The probability difference calculation unit is used to calculate the probability difference of different candidate dialogue evolution trajectories at the same node position; The comparison and judgment unit is used to determine whether the probability difference exceeds a preset probability threshold. The marking statistics unit is used to mark the node positions that exceed a preset probability threshold as branch nodes and generate a set of branch nodes.

[0059] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An AI dialogue optimization method based on multi-level intent understanding, characterized in that, The method includes: Extract topic intent features, target intent features, and emotional intent features from multi-turn dialogue text, and fuse them to generate a dialogue intent state vector; Multiple subsequent dialogue node sequences are generated based on the dialogue intent state vector, and the response achievement probability is calculated for each subsequent dialogue node sequence to form a set of candidate dialogue evolution trajectories. Compare the response achievement probabilities of different candidate dialogue evolution trajectories at the same node position, identify node positions where the probability difference exceeds a preset probability threshold, and generate a set of divergent nodes; Based on the set of divergence nodes, calculate the difference value between the current dialogue intent state vector and the intent of each candidate dialogue evolution trajectory at the corresponding node position, and generate an intervention intensity parameter based on the intent difference value. The response achievement probability of the candidate dialogue evolution trajectory is corrected according to the intervention intensity parameter. The candidate dialogue evolution trajectory with the highest corrected probability is selected as the target dialogue evolution trajectory, and an optimized dialogue response is generated based on the target dialogue evolution trajectory.

2. The AI ​​dialogue optimization method based on multi-level intent understanding according to claim 1, characterized in that, The steps of extracting topic intent features, target intent features, and emotional intent features from multi-turn dialogue text and fusing them to generate a dialogue intent state vector include: The text of multi-turn dialogues is marked with turn number and user speech is distinguished from system speech. Extract keywords from user speech, count the frequency of each keyword in different rounds, and generate a topic intent vector; Identify target expressions in user speech, extract the target type and its location, and generate a target intent vector; Calculate the emotional polarity and emotional intensity values ​​of user utterances to generate an emotional intent vector; The topic intent vector, target intent vector, and emotional intent vector are concatenated and merged in a fixed dimensional order to generate a dialogue intent state vector.

3. The AI ​​dialogue optimization method based on multi-level intent understanding according to claim 1, characterized in that, The steps of generating multiple subsequent dialogue node sequences based on the dialogue intent state vector, and calculating the response achievement probability for each subsequent dialogue node sequence to form a candidate dialogue evolution trajectory set include: Multiple drafts of subsequent dialogue nodes are constructed based on the dialogue intent state vector. Each draft of a subsequent dialogue node contains topic information and guiding statements. Extract the topic information from the drafts of each subsequent dialogue node and number and mark the topic information. Node connections are constructed based on the logical order of the topic information, forming multiple sequences of subsequent dialogue nodes; Calculate the intent transfer value between adjacent nodes in each subsequent dialogue node sequence; The probability of achieving a response is calculated for each subsequent dialogue node sequence based on the intent transfer value and the target intent matching value, and the node sequence and the probability of achieving a response are combined to generate a candidate dialogue evolution trajectory.

4. The AI ​​dialogue optimization method based on multi-level intent understanding according to claim 1, characterized in that, The steps of comparing the response achievement probabilities of different candidate dialogue evolution trajectories at the same node position, identifying node positions where the probability difference exceeds a preset probability threshold, and generating a set of divergent nodes include: Establish node position indexes for the subsequent dialogue node sequences in each candidate dialogue evolution trajectory; Extract the response achievement probability of each candidate dialogue evolution trajectory at the same node position; Calculate the probability difference of different candidate dialogue evolution trajectories at the same node position; Determine whether the probability difference exceeds a preset probability threshold; Nodes whose positions exceed a preset probability threshold are marked as branch nodes, and a set of branch nodes is generated.

5. The AI ​​dialogue optimization method based on multi-level intent understanding according to claim 1, characterized in that, The step of calculating the intention difference value between the current dialogue intention state vector and the intention difference value of each candidate dialogue evolution trajectory at the corresponding node position based on the set of divergence nodes, and generating the intervention intensity parameter based on the intention difference value, includes: Extract the thematic intent component, target intent component, and emotional intent component corresponding to the divergence node; Extract the corresponding components from the dialogue intent state vector and construct the current intent component matrix; Extract the intent components corresponding to divergence nodes from the candidate dialogue evolution trajectory, and construct the trajectory intent component matrix; Calculate the difference between the current intent component matrix and the trajectory intent component matrix, and normalize the intent difference. Intervention intensity parameters are generated based on the normalized difference values.

6. The AI ​​dialogue optimization method based on multi-level intent understanding according to claim 1, characterized in that, The steps of correcting the response achievement probability of the candidate dialogue evolution trajectory according to the intervention intensity parameter, selecting the candidate dialogue evolution trajectory with the highest corrected probability as the target dialogue evolution trajectory, and generating an optimized dialogue response based on the target dialogue evolution trajectory include: The response achievement probability of candidate dialogue evolution trajectories is weighted and corrected based on the intervention intensity parameter; Calculate the corrected probability of achieving the response; Select the candidate dialogue evolution trajectory with the highest probability of achieving the corrected response as the target dialogue evolution trajectory; Generate an optimized dialogue response based on the sequence of subsequent dialogue nodes in the target dialogue evolution trajectory; Output the optimized dialogue response.

7. An AI dialogue optimization system based on multi-level intent understanding, characterized in that, The system includes: The intent state analysis module is used to extract topic intent features, target intent features, and emotional intent features from multi-turn dialogue text, and fuse them to generate a dialogue intent state vector; The dialogue trajectory generation module is used to generate multiple subsequent dialogue node sequences based on the dialogue intent state vector, and calculate the response achievement probability for each subsequent dialogue node sequence to form a candidate dialogue evolution trajectory set. The divergence node statistics module is used to compare the response achievement probabilities of different candidate dialogue evolution trajectories at the same node position, identify node positions where the probability difference exceeds a preset probability threshold, and generate a set of divergence nodes. The intervention parameter generation module is used to calculate the intention difference value between the current dialogue intention state vector and the intention difference value of each candidate dialogue evolution trajectory at the corresponding node position based on the set of divergence nodes, and generate intervention intensity parameters based on the intention difference value. The target trajectory generation module is used to correct the response achievement probability of the candidate dialogue evolution trajectory according to the intervention intensity parameter, select the candidate dialogue evolution trajectory with the highest corrected probability as the target dialogue evolution trajectory, and generate an optimized dialogue response based on the target dialogue evolution trajectory.

8. The AI ​​dialogue optimization system based on multi-level intent understanding according to claim 7, characterized in that, The intent state analysis module includes: The text tagging unit is used to tag the turn of a multi-turn dialogue text and distinguish between user speech and system speech; The topic analysis unit is used to extract keywords from user discourse, count the frequency of each keyword in different rounds, and generate topic intent vectors. The target analysis unit is used to identify target expressions in user speech, extract the target type and its location, and generate a target intent vector. The sentiment analysis unit is used to calculate the sentiment polarity and sentiment intensity values ​​of user utterances and generate sentiment intent vectors. The vector fusion unit is used to concatenate and fuse the topic intent vector, target intent vector, and emotional intent vector in a fixed dimensional order to generate a dialogue intent state vector.

9. The AI ​​dialogue optimization system based on multi-level intent understanding according to claim 7, characterized in that, The dialogue trajectory generation module includes: The draft generation unit is used to construct multiple drafts of subsequent dialogue nodes based on the dialogue intent state vector. Each draft of a subsequent dialogue node contains topic information and guiding statements. The marking execution unit is used to extract topic information from the drafts of each subsequent dialogue node and to number and mark the topic information. The node sequence generation unit is used to construct node connection relationships based on the logical order between topic information, forming multiple subsequent dialogue node sequences. The transfer value calculation unit is used to calculate the intent transfer value between adjacent nodes in each subsequent dialogue node sequence; The probability generation insertion unit is used to calculate the response achievement probability of each subsequent dialogue node sequence based on the intent transfer value and the target intent matching value, and to combine the node sequence and the response achievement probability to generate the candidate dialogue evolution trajectory.

10. The AI ​​dialogue optimization system based on multi-level intent understanding according to claim 7, characterized in that, The divergence node statistics module includes: The index creation unit is used to establish node position indexes for the sequence of subsequent dialogue nodes in each candidate dialogue evolution trajectory; The probability extraction unit is used to extract the response achievement probability of each candidate dialogue evolution trajectory at the same node position; The probability difference calculation unit is used to calculate the probability difference of different candidate dialogue evolution trajectories at the same node position; The comparison and judgment unit is used to determine whether the probability difference exceeds a preset probability threshold. The marking statistics unit is used to mark the node positions that exceed a preset probability threshold as branch nodes and generate a set of branch nodes.

Images