Intelligent customer service interaction dialogue method based on semantic understanding

Abstract

The application relates to the technical field of audio processing, in particular to an intelligent customer service interactive dialogue method based on semantic understanding, which determines the syntactic complexity and intention decoupling degree of a user statement in each round of interactive dialogue, combines the number of occurrences of a leading intention in previous interactive dialogue, obtains the user mixed intention separation situation of the leading intention in each round of interactive dialogue, and then obtains the influence weight of each leading intention in each round of interactive dialogue according to the round of each round of interactive dialogue; the dynamic characteristics of the influence weight of each leading intention are determined, the number of occurrences of each leading intention in all rounds of interactive dialogue is combined, the influence weight change trend of each leading intention is obtained, the accuracy of the influence weight of each leading intention and the influence weight change trend can be ensured, so that the influence weight and the influence weight change trend are used as inputs of a network model to train the network model, the network model meets the dynamic change of intention demand in multi-round interactive dialogue, and the accuracy of the network model is ensured.

Description

Technical Field

[0001] This invention relates to the field of audio processing technology, and more specifically to an intelligent customer service interactive dialogue method based on semantic understanding. Background Technology

[0002] Traditional intelligent customer service systems often rely on keyword matching or rule templates for responses. While these methods are effective for simple, single-intent user queries, they reveal significant limitations when faced with complex scenarios. In actual interactions, user queries often contain mixed intents, such as simultaneously inquiring about product features, pricing discounts, and after-sales policies. Furthermore, the wording is flexible and varied, potentially including colloquialisms, omissions, or even semantic ambiguity. Conversations between users and intelligent customer service typically involve multiple rounds, and the historical dialogue information contains the evolutionary logic of user needs and underlying intents.

[0003] Current intelligent customer service systems, in handling multi-turn interactive dialogues with customers, identify intent features from these dialogues and then use them as input to relevant network models, such as natural language understanding (NLP) models, to train the NLP models. Finally, the trained NLP models are used for subsequent user intent recognition. However, current methods typically rely solely on static semantic matching for intent feature identification during multi-turn dialogues, ignoring the dynamic changes in user intent needs throughout the conversation. This results in inaccurate identification of user intent features, often failing to accurately understand user intent. Furthermore, contextual breaks exacerbate the difficulty of semantic understanding, ultimately affecting the accuracy of NLP model training. Summary of the Invention

[0004] To address the technical problem of inaccurate user intent features identified in multi-turn interactive dialogue processing, the present invention aims to provide an intelligent customer service interactive dialogue method based on semantic understanding. The specific technical solution adopted is as follows:

[0005] This invention provides an intelligent customer service interactive dialogue method based on semantic understanding, comprising:

[0006] Determine the syntactic complexity and intent decoupling degree of user statements in each round of interactive dialogue. Combine the number of times the dominant intent of each round of interactive dialogue appears in previous interactive dialogues to obtain the separation of user mixed intents in each round of interactive dialogue.

[0007] Based on the separation of user mixed intents in each round of interactive dialogue, and combined with the round of interactive dialogue, the influence weight of each dominant intent in each round of interactive dialogue is obtained.

[0008] The dynamic characteristics of the influence weights of each dominant intention are determined, and the trend of the influence weight changes of each dominant intention is obtained by combining the number of times each dominant intention appears in all rounds of interactive dialogue. The influence weights of each dominant intention and the trend of the influence weight changes are used as inputs to the network model to train the network model.

[0009] In an exemplary embodiment, after obtaining the influence weights of each dominant intent in each round of interactive dialogue, the semantic understanding-based intelligent customer service interactive dialogue method further includes:

[0010] For a given round of interactive dialogue, determine the percentage of rounds in which the dominant intent of that round of interactive dialogue appears in all subsequent rounds of interactive dialogue.

[0011] Determine the correlation between the intent recognition results of the current round of interactive dialogue and the next round of interactive dialogue;

[0012] The influence weight of the interactive dialogue is adjusted based on the resolution status factor of the current round of dialogue; the resolution status factor is obtained from the proportion of the number of rounds and the correlation, and the resolution status factor is positively correlated with both the proportion of the number of rounds and the correlation.

[0013] In an exemplary embodiment, the process of obtaining the user's mixed intent separation includes:

[0014] For a given round of interactive dialogue, the initial user mixed intent separation performance is obtained based on the syntactic complexity and intent decoupling degree corresponding to that round of interactive dialogue; the initial user mixed intent separation performance is positively correlated with the intent decoupling degree and negatively correlated with the syntactic complexity.

[0015] The initial user mixed intent separation performance is reversed by using the number of times the dominant intent of this round of interactive dialogue appears in all previous rounds of interactive dialogue, thus obtaining the target user mixed intent separation performance of the dominant intent in this round of interactive dialogue.

[0016] In an exemplary embodiment, the step of reverse-correcting the initial user mixed intent separation performance based on the frequency of occurrence of the dominant intent in this round of interactive dialogue in all previous rounds of interactive dialogue includes:

[0017] A correction coefficient is obtained based on the number of times the dominant intent of this round of interactive dialogue appears in all previous rounds of interactive dialogue; the correction coefficient is inversely correlated with the number of appearances.

[0018] The initial user mixed intent separation behavior is corrected based on the correction coefficient.

[0019] In an exemplary embodiment, the process of obtaining the influence weights includes:

[0020] The influence weight of the dominant intent in this round of interactive dialogue is obtained based on the performance of the target user's mixed intent separation in this round of interactive dialogue and the round number of this round of interactive dialogue; the influence weight is positively correlated with both the performance of the target user's mixed intent separation and the round number of this round of interactive dialogue.

[0021] In an exemplary embodiment, the process of obtaining the dynamic features affecting the weights includes:

[0022] Obtain the slope of each influence weight in the curve showing the change in the influence weight of the dominant intent;

[0023] The mean of the slope is calculated as a dynamic feature of the influencing weight.

[0024] In an exemplary embodiment, the process of obtaining the trend of influence weight changes includes:

[0025] Based on the occurrence frequency of the dynamic features and the dominant intent in all rounds of interactive dialogue, the trend of the influence weight of the dominant intent is obtained; the trend of the influence weight of the dominant intent is positively correlated with the occurrence frequency of the dynamic features and the dominant intent in all rounds of interactive dialogue.

[0026] In an exemplary embodiment, the process of obtaining the syntactic complexity includes:

[0027] Determine the dependency tree of the user statement;

[0028] Obtain the depth of the dependency tree, the total number of dependency edges, and the dependency distance of each dependency edge;

[0029] The syntactic complexity of the user statement is obtained based on the depth of the dependency tree, the total number of dependency edges, and the degree of fluctuation in dependency distance; the syntactic complexity is positively correlated with the depth of the dependency tree, the total number of dependency edges, and the degree of fluctuation in dependency distance.

[0030] In an exemplary embodiment, the process of obtaining the degree of intent decoupling includes:

[0031] Determine the probability of each intent of the user's statement;

[0032] The degree of intention decoupling is obtained based on the degree of fluctuation of the probability and the significance of the intention; the degree of intention decoupling is positively correlated with both the degree of fluctuation of the probability and the significance of the intention; the significance of the intention is positively correlated with the prominence of the maximum probability and the gap between the maximum probability and the minimum probability.

[0033] In one exemplary embodiment, the semantic understanding-based intelligent customer service interactive dialogue method further includes:

[0034] The influence weights of each dominant intent and the trends of their changes are integrated into a structured contextual feature vector, which is then used as input to the natural language understanding model to train the model.

[0035] This invention offers the following advantages: Based on the separation of user mixed intentions in each round of interactive dialogue, and considering the number of rounds, this invention obtains the influence weight of each dominant intention in each round of interactive dialogue. It takes into account the dynamic changes in user intention needs throughout multiple rounds of interactive dialogue, making the influence weight of each dominant intention more consistent with dynamic changes, thus ensuring the accuracy of the influence weight in each round of interactive dialogue. Furthermore, the trend of influence weight changes of each dominant intention, obtained by combining the dynamic characteristics of the influence weight, is more consistent with dynamic changes, ensuring that the trend of influence weight changes deeply integrates with the dynamic changes in user needs, improving accuracy. Therefore, when the obtained user intention features are used as input to train the network model, the network model conforms to the dynamic changes in intention needs during multiple rounds of interactive dialogue, ensuring the accuracy of the network model and making intention recognition more aligned with dialogue logic and the user's true needs, thus more accurately identifying the user's true intention. Attached Figure Description

[0036] Figure 1 This is a flowchart of an intelligent customer service interactive dialogue method based on semantic understanding provided in one embodiment of the present invention;

[0037] Figure 2 This is a flowchart of obtaining syntactic complexity according to an embodiment of the present invention;

[0038] Figure 3 This is a schematic diagram of a dependency tree provided in one embodiment of the present invention;

[0039] Figure 4 This is a flowchart illustrating the process of obtaining the degree of intent decoupling according to an embodiment of the present invention;

[0040] Figure 5 This is a flowchart illustrating the process of obtaining the separation of user mixed intents according to an embodiment of the present invention;

[0041] Figure 6 This is a flowchart of the influence weight correction provided in one embodiment of the present invention;

[0042] Figure 7 This is a graph showing the change in the final influence weight provided in one embodiment of the present invention. Detailed Implementation

[0043] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the specific implementation methods, structures, features, and effects of the present invention are described in detail below with reference to the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0044] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. All data and information collected in this application have been obtained with full consent.

[0045] This embodiment provides a semantic understanding-based intelligent customer service interactive dialogue method, which requires acquiring multi-turn interactive dialogues (data type: text string) between the user and the intelligent customer service. It uses standard NLP (Natural Language Processing) techniques to perform intent recognition and entity extraction to obtain surface semantic elements, including entities and intents.

[0046] In semantic understanding-based intelligent customer service systems, entity extraction and intent recognition are core steps in semantic parsing, i.e., Natural Language Understanding (NLU). Current mainstream methods employ a separate pipelined processing of the BERT-BiLSTM-CRF joint model (named entity recognition) and the BERT fine-tuned classifier (intent recognition) to obtain the entity set and intent for the user's input question data. Entity extraction outputs several entity arrays; intent recognition outputs several intents and their corresponding intent probability distributions. It should be understood that the probability of each intent ranges from 0 to 1, and the sum of the probabilities of all intents is 1. Normally, at least two intents will be identified; therefore, this embodiment takes the identification of at least two intents as an example, without considering the special case of only one intent.

[0047] For each round of a multi-turn interactive dialogue, including user statements (such as questions raised by the user) and the corresponding answers from the intelligent customer service, the process involves: for each user statement, obtaining the intents within the statement and their probabilities, and identifying the intent with the highest probability as the dominant intent for that round. This yields the dominant intent for each round of the interactive dialogue. It should be understood that the dominant intent of one round may appear in subsequent rounds of the interactive dialogue.

[0048] like Figure 1 As shown in the figure, the intelligent customer service interactive dialogue method based on semantic understanding provided in this embodiment includes the following steps:

[0049] Step S1: Determine the syntactic complexity and intent decoupling degree of user statements in each round of interactive dialogue. Combine the number of times the dominant intent of each round of interactive dialogue appears in previous interactive dialogues to obtain the separation of user mixed intents in each round of interactive dialogue.

[0050] Step S2: Based on the separation of user mixed intents in each round of interactive dialogue, and combined with the round of interactive dialogue, obtain the influence weight of each dominant intent in each round of interactive dialogue.

[0051] Step S3: Determine the dynamic characteristics of the influence weight of each dominant intention, and combine the occurrence frequency of each dominant intention in all rounds of interactive dialogue to obtain the trend of the influence weight change of each dominant intention.

[0052] The following detailed explanation of each step, in conjunction with the accompanying drawings, is provided.

[0053] Step S1: Determine the syntactic complexity and intent decoupling degree of user statements in each round of interactive dialogue. Combine the number of times the dominant intent of each round of interactive dialogue appears in previous interactive dialogues to obtain the separation of user mixed intents in each round of interactive dialogue.

[0054] The more complex the sentence structure of a user's statement and the more significant the nesting of clauses, the greater the understanding load on the intelligent customer service's interactive dialogue. In other words, the more complex the sentence structure of a user's statement, the more difficult it is to process the interactive dialogue.

[0055] For any round of interactive dialogue, the syntactic complexity of the user's statements in that round of dialogue is obtained. In an exemplary embodiment, the syntactic complexity of the user's statements is determined based on a dependency parsing algorithm, such as... Figure 2 As shown, a specific process for obtaining syntactic complexity is given:

[0056] Step S11: Determine the dependency tree of the user statement.

[0057] Determine the dependency tree of the user's statement. The basic principle of dependency parsing is to analyze the dependency relationships between words in a sentence and construct a dependency tree to represent these relationships. In this tree, each word in the sentence is a node, and the dependency relationships are the edges between nodes; the root node usually corresponds to the core word, and other words are connected to the core word through dependency relationships. In the dependency tree, based on the dependency tree structure, the distribution of dependency edges, and the distance between two words with a dependency relationship in the linear order of the sentence, the number of words between them is used as the dependency distance of the dependency edge. For example, the dependency tree of the sentence "Please help me check the progress of the unresolved return application and urge customer service to process it as soon as possible" is as follows: Figure 3 As shown.

[0058] Step S12: Obtain the depth of the dependency tree, the total number of dependency edges, and the dependency distance of each dependency edge.

[0059] Retrieve the depth of the dependency tree, the total number of dependency edges, and the dependency distance of each dependency edge. The depth of the dependency tree represents the length of the longest path from the root node to the farthest leaf node.

[0060] Step S13: Based on the depth of the dependency tree, the total number of dependency edges, and the degree of fluctuation in dependency distance, obtain the syntactic complexity of the user statement.

[0061] To obtain the degree of fluctuation of dependency distances in the dependency tree, the standard deviation is used to characterize the degree of fluctuation, that is, to obtain the standard deviation of the dependency distance of each dependency edge in the dependency tree.

[0062] The greater the depth of the dependency tree, the more complex the hierarchical structure of the user statement, and the higher the syntactic complexity of the user statement. The total number of dependency edges in the dependency tree represents the number of dependency relationships between all words in the user statement. The more complex the nested clauses in the user statement, the more dependency edges there will be, and the higher the syntactic complexity of the user statement. The greater the fluctuation of dependency distance in the dependency tree, that is, the larger the standard deviation of dependency distance, the more unstable the number of words between two words with dependency relationships in the user statement, and the higher the syntactic complexity of the user statement. Therefore, the syntactic complexity of a user statement is positively correlated with the depth of the dependency tree, the total number of dependency edges, and the standard deviation of dependency distance. In an exemplary embodiment, a specific quantification method for syntactic complexity is given below:

[0063] ;

[0064] in, Indicates the first The syntactic complexity of user statements in a round-robin interactive dialogue; Indicates the first The depth of the dependency tree of user statements in a round-robin interactive dialogue; Indicates the first The total number of dependency edges for user statements in a round-robin interactive dialogue; Indicates the first The standard deviation of dependency distances for user statements in a round-robin interactive dialogue. In the dependency tree, Reflecting the The dependency structure of user statements in a round-robin interactive dialogue. This represents the normalization function, which can be the tanh function in this case.

[0065] In this embodiment, the current round of interactive dialogue is set as the xth round of interactive dialogue. Alternatively, the xth round of interactive dialogue can be regarded as the last round of interactive dialogue in a multi-round interactive dialogue.

[0066] In processing user statements, there may be errors in the intelligent customer service's recognition of user input intent during interactive dialogues. Therefore, the intent recognition result reflects this intent recognition bias. Specifically, the intent result includes several intents of the user statement and the probability of each intent, with each intent representing a specific user goal. The higher the probability of a particular intent and the greater the difference in probability relative to other intents, the more accurate the intent recognition result and the higher the degree of intent decoupling. Therefore, if... Figure 4 As shown, the following is a specific process for obtaining the degree of decoupling of the user's intent:

[0067] Step S14: Determine the probability of each intent of the user's statement;

[0068] Step S15: Based on the degree of probability fluctuation and the significance of intent, obtain the degree of intent decoupling.

[0069] The degree of fluctuation in the probability of obtaining each intent of the user's statement is represented by the standard deviation, which is the standard deviation of the probability of obtaining each intent in the user's statement. The larger the standard deviation, the more chaotic the distribution of intent probabilities, the clearer the identification of the user's intent, the weaker the interference of other intents on the dominant intent, and the higher the degree of intent decoupling. The degree of intent decoupling is positively correlated with the standard deviation of the probability.

[0070] The intent salience of user statements is obtained by analyzing the distribution of the intent with the highest probability (i.e., the dominant intent) among all intents. A larger difference between the highest and lowest probabilities indicates a clearer understanding of the user statement's intent, weaker interference from other intents on the dominant intent, and thus higher intent salience and a greater degree of intent decoupling. Intent salience is positively correlated with the difference between the highest and lowest probabilities.

[0071] The salience of the highest probability is determined. Salience represents the prominence of the highest probability relative to other probabilities. Since the sum of all probabilities is 1, the higher the value of the highest probability, the higher its salience. Therefore, the value of the highest probability is used as its salience here. The higher the salience of the highest probability, the clearer the user's intent is identified, the weaker the interference of other intents on the dominant intent, and the higher the intent salience and the higher the degree of intent decoupling. Intent salience is positively correlated with the salience of the highest probability, i.e., with the highest probability.

[0072] In one exemplary embodiment, a specific method for quantifying the degree of intended decoupling is given below:

[0073] ;

[0074] in, Indicates the first The degree of decoupling between the intent of user statements in round-robin interactive dialogue. Indicates the first The standard deviation of the probability of the intent of user statements in a round-robin interactive dialogue; They represent the first The maximum and minimum probabilities among all the intentions of the user's statements in a round-robin interactive dialogue.

[0075] In the continuous multi-round interactive dialogue between users and intelligent customer service, for any given round of dialogue, combined with the presentation of intent in previous rounds of dialogue, if a certain dominant intent appears repeatedly in multiple rounds of dialogue, it may mean that the user's problem has not been resolved. In this case, the separation of the user's mixed intent becomes worse, affecting the separation effect of the user's mixed intent in the current round of dialogue.

[0076] In one exemplary embodiment, such as Figure 5 As shown, the following is a specific process for obtaining information on the separation of user mixed intents:

[0077] Step S16: For a certain round of interactive dialogue, based on the syntactic complexity and intent decoupling degree of the corresponding round of interactive dialogue, obtain the initial user mixed intent separation performance of the corresponding round of interactive dialogue.

[0078] For the Round of interactive dialogue, the first The higher the syntactic complexity of a round-based interactive dialogue, the more difficult it is to achieve mixed intent separation. The worse the initial mixed user intent separation in a round of interactive dialogue, the better the performance. The higher the degree of decoupling of intents in a round-based interactive dialogue, the easier it is to achieve separation of mixed intents. The better the initial user intent separation in a round-based interactive dialogue, the better the performance. Therefore, the initial user intent separation performance is positively correlated with the degree of intent decoupling and inversely correlated with syntactic complexity. In an exemplary embodiment, the following gives the first... A specific way to quantify the separation of initial user mixed intents in a round-robin interactive dialogue:

[0079] ;

[0080] in, Indicates the first The performance of separating the initial mixed user intent in the corresponding round of interactive dialogue. The performance regarding the separation of initial mixed user intent in round-based interactive dialogues only applies to the first round. The performance of user mixed intent separation in round-robin interactive dialogue.

[0081] Step S17: Use the number of times the dominant intent of this round of interactive dialogue appears in all previous rounds of interactive dialogue to reverse correct the initial user mixed intent separation performance, and obtain the target user mixed intent separation performance of the dominant intent in this round of interactive dialogue.

[0082] Determine the first The preceding rounds of interactive dialogue include: Round 1, Round 2, and so on up to Round 3. A round of interactive dialogue. Confirm the first... The number of times the dominant intent of a round of interactive dialogue appears in all previous rounds of interactive dialogue, since each round of interactive dialogue either appears once in the first round. The dominant intent of the first round of interactive dialogue is either not to appear in the second round. The dominant intent of the first round of interactive dialogue, therefore, the second round The number of times the dominant intent in a round of interactive dialogue appears in all previous rounds of interactive dialogue can be understood as: the number of times the dominant intent appears in all previous rounds of interactive dialogue. In all previous rounds of interactive dialogue, the first round appeared The number of rounds of interactive dialogue that represent the dominant intent of the dialogue.

[0083] No. The more times the dominant intent in a round of interactive dialogue appears in all previous rounds of interactive dialogue, the more likely it is to be the dominant intent in the first round. If the dominant intent in a round of interactive dialogue persists throughout all previous rounds, it reflects that the intelligent customer service is inadequate or unable to resolve the user's problem. Therefore, the worse the separation of the target user's mixed intents, the better the performance. Thus, taking the first round as an example... The dominant intent of the first round of interactive dialogue is reverse-corrected by the number of times it appears in all previous rounds of interactive dialogue. The performance of the initial user mixed intent separation in the first round of interactive dialogue yields the results of the first round. The dominant intent of the round-robin dialogue in the first round The performance of the target user's mixed intent in the round-robin interactive dialogue.

[0084] The specific implementation process of reverse correction is as follows: according to the first The dominant intent of the first round of interactive dialogue is calculated by counting the number of times it appears in all previous rounds of interactive dialogue, resulting in a correction coefficient; the correction coefficient is inversely correlated with the number of occurrences. Then, based on the correction coefficient, the first round of interactive dialogue is corrected. The initial user mixed intent separation behavior in a round-based interactive dialogue. In an exemplary embodiment, the following is given: The dominant intent of the round-robin dialogue in the first round The calculation method for the separation of mixed intents of target users in round-robin interactive dialogues:

[0085] ;

[0086] in, Indicates the first The dominant intent of the round-robin dialogue in the first round The performance of the target user's mixed intent in the round-robin interactive dialogue. Indicates the first The number of times the dominant intent in a round of interactive dialogue appears in all previous rounds of interactive dialogue; This represents an exponential function with the natural constant as its base. This represents the correction factor.

[0087] With the first Taking the dominant intent of the first round of interactive dialogue as an example, the first round The dominant intent of the round-robin interactive dialogue, besides the first round Besides appearing in one round of interactive dialogue, it may also appear in other rounds of interactive dialogue, so we can obtain the first... The dominant intent in each round of interactive dialogue is determined by the number of times it appears in all previous rounds of interactive dialogue, thus yielding the [number of rounds]. The dominant intent of each round of interactive dialogue is separated from the mixed intent of the target user corresponding to each round of interactive dialogue.

[0088] With the first The dominant intent of the round-robin dialogue is aimed at the first Round of interactive dialogue, with the first Taking the calculation method of the target user's mixed intent separation in a round-robin interactive dialogue as an example:

[0089] ;

[0090] in, Indicates the first The dominant intent of the round-robin dialogue in the first round The performance of the target user's mixed intent separation in the first round of interactive dialogue, i.e., the... The dominant intent of the round-robin dialogue is aimed at the first Round of interactive dialogue, with the first The performance of the separation of mixed intents of the target user in the round-robin interactive dialogue; Indicates the first The dominant intent of the round-robin dialogue in the first round The number of times it appears in all previous rounds of interactive dialogue.

[0091] Step S2: Based on the separation of user mixed intents in each round of interactive dialogue, and combined with the round of interactive dialogue, obtain the influence weight of each dominant intent in each round of interactive dialogue.

[0092] For each round of interactive dialogue, by determining the dominant intent of each round of interactive dialogue, the questions asked by the user throughout the interactive dialogue are reflected. In this way, the intent recognition results of subsequent interactive dialogues are evaluated, and the different degrees of influence of different existing dominant intents on the dominant intent of subsequent interactive dialogues are quantified. In this way, the influence weight of each dominant intent is determined, so as to more accurately reflect the actual state of the interactive dialogue.

[0093] In a continuous multi-turn interactive dialogue, the dominant intent (question) raised by a user before a particular turn changes over time (turns). Generally, the more recent the dominant intent, the greater its impact on that turn, while the earlier the dominant intent, the smaller its impact. This decaying effect over time is called "timeliness." Correspondingly, for the first... Round of interactive dialogue to obtain the first The number of rounds of interactive dialogue, and based on the number of rounds... The number of rounds of interactive dialogue yields a round coefficient, the 1st round... The larger the round of interactive dialogue, the more... The larger the round coefficient of the interactive dialogue, the more... The dominant intent of the round-robin dialogue in the first round The greater the weight of each round of interactive dialogue. The dominant intent of the round-robin dialogue in the first round The better the separation of target user mixed intents in the first round of interactive dialogue, the more meaningful it is to acquire their dominant intent. The dominant intent of the round-robin dialogue in the first round The greater the impact weight of each round of interactive dialogue, the more positively correlated the impact weight is with the separation of the target user's mixed intent and the number of rounds.

[0094] According to the The dominant intent of the round-robin dialogue in the first round The performance of the target user's mixed intent separation in the first round of interactive dialogue, and the performance of the second round of interactive dialogue. The round coefficient of the interactive dialogue is obtained. The dominant intent of the round-robin dialogue in the first round The influence weight of each round of interactive dialogue. A specific quantification method is given below:

[0095] ;

[0096] in, Indicates the first The dominant intent of the round-robin dialogue in the first round The influence weight of round-robin interactive dialogue Indicates the first The round of the interactive dialogue, where x represents the current round of the interactive dialogue, or the last round of the interactive dialogue. Indicates the first The round coefficient of a round-based interactive dialogue is essentially a factor of the number of rounds. The number of rounds in the interactive dialogue is normalized. Earlier rounds of interactive dialogue, that is... The smaller the value, the greater the degree to which the dominant intention of the interactive dialogue decays over time in the current round of interactive dialogue. In other words, the more distant the dominant intention is from the current round of interactive dialogue, the smaller its influence weight. Thus, we obtain the influence weight of the dominant intention in each round of interactive dialogue in each round of interactive dialogue.

[0097] However, the resolution status of the dominant intent in each round of interactive dialogue also affects the influence weight of the dominant intent. For example, a dominant intent raised a long time ago but not yet resolved may be more important than a recently resolved dominant intent. The logic for adjusting the influence weight of a single dominant intent is: increase the influence weight of unresolved dominant intents and decrease the influence weight of resolved dominant intents.

[0098] The more frequently a dominant intent in a given round of interaction appears in subsequent rounds, the more likely that intent remains unresolved, and its influence weight needs to be increased. In the transition between adjacent rounds of interaction, if the dominant intent of a given round does not appear in the next round, the intelligent customer service can preliminarily determine that the dominant intent has been resolved, thereby reducing the proportion of that round's influence weight in the overall influence calculation.

[0099] Therefore, after obtaining the influence weights of each dominant intent in each round of interactive dialogue, such as Figure 6 As shown, the semantic understanding-based intelligent customer service interactive dialogue method also includes a process for correcting the influencing weights, as detailed below:

[0100] Step S21: For a given round of interactive dialogue, determine the percentage of interactive dialogues in which the dominant intent of that round of interactive dialogue appears in all subsequent rounds of interactive dialogue.

[0101] For the Round of interactive dialogue to determine the first The dominant intent of the round-robin dialogue in the first round The percentage of interactive dialogues that appear in all subsequent rounds of interactive dialogue. First, the percentage of interactive dialogues in the first round is determined. The dominant intent of the round-robin dialogue in the first round The number of times the dialogue appears in all subsequent rounds of dialogue, i.e., the number of rounds of dialogue that appear in all subsequent rounds of dialogue. Then, obtain the number of rounds of dialogue. Calculate the total number of rounds in all rounds of interactive dialogue after the first round. The dominant intent of the round-robin dialogue in the first round The ratio of the number of interactive dialogues that occur in all subsequent rounds of interactive dialogue to the total number of rounds is used as the basis for determining the number of rounds of interactive dialogue. The dominant intent of the round-robin dialogue in the first round The percentage of interactive dialogues that appear in all subsequent rounds of interactive dialogue. A higher percentage indicates a higher number of rounds. The dominant intent of the round-robin dialogue in the first round The more significantly a problem remains unresolved in subsequent rounds of interactive dialogue, the more its influence weight needs to be positively adjusted.

[0102] Step S22: Determine the correlation between the intent recognition results of the current round of interactive dialogue and the next round of interactive dialogue.

[0103] Get the The intent recognition results of the first round of interactive dialogue, and the next round of interactive dialogue (i.e., the first round of dialogue) The intent recognition results of the first round of interactive dialogue are then obtained. The correlation between these two intent recognition results is then calculated. The greater the correlation, the higher the likelihood of a higher ... The more coherent the intent of one round of interactive dialogue with the next round of interactive dialogue, the less the dominant intent is resolved, and the more the influence weight needs to be positively adjusted; the smaller the correlation, that is, the greater the difference between the two intent recognition results, the more it indicates a break in intent logic, and the greater the degree of downward adjustment of influence weight.

[0104] In an exemplary embodiment, since the intent recognition result consists of a discrete number of intents, the correlation can be obtained in the following ways: obtain the chi-square test result of the two intent recognition results and normalize it as the correlation between the two intent recognition results; or, obtain the Cramer V coefficient of the two intent recognition results and normalize it as the correlation between the two intent recognition results.

[0105] Step S23: Adjust the influence weight of this round of interactive dialogue based on the resolution status factor of this round of interactive dialogue.

[0106] Based on the obtained round-count ratio and correlation, the solution state factor is obtained. The above analysis shows that the solution state factor is positively correlated with both the round-count ratio and the correlation. For the... Round of interactive dialogue, according to the first The dominant intent of the round-robin dialogue in the first round The percentage of interactive dialogues that appear in all subsequent rounds of interactive dialogue, and the percentage of the number of interactive dialogues in the first round. The correlation between the intent recognition results of the first round of interactive dialogue and the next round of interactive dialogue is obtained. The dominant intent of the round-robin dialogue in the first round The solution state factor for round-robin interactive dialogue is calculated using the following formula:

[0107] ;

[0108] in, Indicates the first The dominant intent of the round-robin dialogue in the first round The resolution of state factors in round-robin interactive dialogue. Indicates the first The dominant intent of the round-robin dialogue in the first round The number of rounds of interactive dialogue that occur in all subsequent rounds of interactive dialogue. Indicates the first The total number of rounds in all rounds of interactive dialogue after the first round of dialogue. Indicates the first The dominant intent of the round-robin dialogue in the first round The percentage of interactive dialogue rounds that appear in all interactive dialogue rounds after the first round of interactive dialogue; Indicates the first The correlation between the results of the first round of interactive dialogue and the intent recognition results of the next round of interactive dialogue.

[0109] A dominant intent that has been resolved to a higher degree has a smaller resolution state factor value; an unresolved dominant intent has a larger resolution state factor value.

[0110] Then, based on the solution state factor, the first... The dominant intent of the round-robin dialogue in the first round The influence weight of the round-robin interactive dialogue. In an exemplary embodiment, the correction method is as follows:

[0111] ;

[0112] in, Indicates the first The dominant intent of the round-robin dialogue in the first round The final impact weight of each round of interactive dialogue. This yields the final impact weight of the dominant intent in each round of interactive dialogue.

[0113] Step S3: Determine the dynamic characteristics of the influence weight of each dominant intention, and combine the occurrence frequency of each dominant intention in all rounds of interactive dialogue to obtain the trend of the influence weight change of each dominant intention.

[0114] For any dominant intent, which has a corresponding final influence weight in each round of interactive dialogue, the final influence weight of the dominant intent in each round of interactive dialogue is arranged in chronological order and curve fitted to obtain the curve of the change of the final influence weight of the dominant intent.

[0115] The dynamic characteristics of the final influence weight change curve of the dominant intent are obtained, i.e., the trend of the final influence weight change. When the final influence weight of the dominant intent shows an increasing trend, it reflects that the urgency of the dominant intent or user attention is continuously rising, which may mean that the user's need to resolve the dominant intent has not been met, or that they want to prioritize the processing of the dominant intent. When the final influence weight of the dominant intent shows a decreasing trend, it reflects that the dominant intent may be gradually approaching a resolution, or that the user's attention has shifted to other dominant intents. When the final influence weight of the dominant intent tends to stabilize in multiple rounds of interactive dialogue, it indicates that the resolution process of the dominant intent is in a relatively stagnant state, or that its importance remains stable. Figure 7 As shown, the horizontal axis represents each round of interactive dialogue, and the vertical axis represents the final influence weight. Among them, the dominant intention... Figure 1 and dominant intention Figure 2 If the intent needs in rounds A and B of the interactive dialogue are resolved, the final influence weight gradually decreases, while the dominant intent... Figure 3 The ultimate impact of this factor gradually increases, and its persistent nature remains unresolved.

[0116] In an exemplary embodiment, the process of obtaining the dynamic characteristics of the final influence weight change curve of the dominant intent includes: obtaining the slope of each final influence weight in the final influence weight change curve of the dominant intent, and calculating the mean of the slopes in the final influence weight change curve of the dominant intent as the dynamic characteristics of the final influence weight of the dominant intent. When the mean slope is greater than 0, it indicates that the final influence weight is on an increasing trend, and the greater the mean slope is than 0, the stronger the increasing trend of the final influence weight; when the mean slope is equal to 0, it indicates that the final influence weight is in a stable state; when the mean slope is less than 0, it indicates that the final influence weight is on a decreasing trend, and the less the mean slope is than 0, the stronger the decreasing trend of the final influence weight. Therefore, numerically speaking, the larger the mean slope, the stronger the dynamic characteristics of the final influence weight change curve of the dominant intent, reflecting a stronger user demand for resolving the dominant intent, a stronger trend in the influence weight change of the dominant intent, and a positive correlation between the influence weight change trend and the dynamic characteristics.

[0117] The more times the dominant intent appears in all rounds of interactive dialogue, the more serious the degree to which the dominant intent has not been resolved in the continuous dialogue across all rounds of interactive dialogue. The stronger the trend of the influence weight of the dominant intent, the more positively correlated the trend of the influence weight is with the number of occurrences.

[0118] Based on the dynamic characteristics of the final influence weight change curve of the dominant intention and the number of times the dominant intention appears in all rounds of interactive dialogue, the trend of the influence weight change of the dominant intention is obtained, and a specific calculation process is given below:

[0119] ;

[0120] in, Indicates the first The changing trend of the influence weight of the dominant intent in the round-robin interactive dialogue; Indicates the first The number of times the dominant intent in a round-based interactive dialogue occurs across all round-based interactive dialogues; Indicates to Normalization; Indicates the first The dynamic characteristics of the dominant intent in a round-robin interactive dialogue; express A function used for normalization.

[0121] This yields the final influence weight of the dominant intent in each round of interactive dialogue, as well as the trend of the influence weight changes of each dominant intent.

[0122] The dominant intent of each round of interactive dialogue, its final influence weight in each round, and the trend of the influence weight changes of each dominant intent are used as input to the network model to train the network model. In an exemplary embodiment, the dominant intent of each round of interactive dialogue, its final influence weight in each round, and the trend of the influence weight changes of each dominant intent are integrated into a structured contextual feature vector, which is used as input to the natural language understanding model to train the natural language understanding model.

[0123] During the training process of a Natural Language Understanding (NLP) model, the integration of these structured contextual feature vectors helps the model overcome the limitations of single-round text processing and interpret new user questions from a global perspective of historical interactions. For example, if the final influence weight of an unresolved dominant intent in the past continues to increase, the NLP model can prioritize associating this dominant intent with the potential relationship between the new question and determine whether the user is supplementing information or inquiring about progress. If the final influence weight of a resolved dominant intent significantly decreases, the NLP model can weaken its interference with the current identification, avoiding misassociations. After completing contextual modeling in this way, the NLP model can more accurately identify the true intent of the user's new question—not only understanding the literal meaning but also capturing its position in the complete dialogue logic (such as continuing old needs, proposing new needs, or transitional needs). This process lays a solid foundation for the response strategy of the next round of interactive dialogue (such as prioritizing the answering of unresolved dominant intents and taking over the user's focus), ultimately achieving a coherent and accurate response of intelligent customer service to user needs, improving the overall smoothness and satisfaction of the interaction.

[0124] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0125] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A semantic understanding-based intelligent customer service interaction dialogue method, characterized in that, The method comprises the following steps: determining the syntactic complexity and the intention decoupling degree of the user statement in each round of interactive dialogue, combining the number of occurrences of the dominant intention in the previous interactive dialogue to obtain the user mixed intention separation of the dominant intention in each round of interactive dialogue; According to the user mixed intention separation of each dominant intention in each round of interactive dialogue, and combining the round number of each round of interactive dialogue, the influence weight of each dominant intention in each round of interactive dialogue is obtained; Determine the dynamic characteristics of the influence weight of each dominant intention, and combine the number of occurrences of each dominant intention in all rounds of interactive dialogue to obtain the influence weight change trend of each dominant intention; The influence weight and the influence weight change trend of each dominant intention are used as the input of the network model to train the network model; The network model is a natural language understanding model; The process of obtaining the user mixed intention separation includes: For a certain round of interactive dialogue, the initial user mixed intention separation performance corresponding to the round of interactive dialogue is obtained according to the syntactic complexity and intention decoupling degree corresponding to the round of interactive dialogue; The initial user mixed intention separation performance is positively correlated with the intention decoupling degree and negatively correlated with the syntactic complexity; The number of occurrences of the dominant intention in all previous rounds of interactive dialogue is used to correct the initial user mixed intention separation performance to obtain the target user mixed intention separation performance of the dominant intention in the round of interactive dialogue; The process of obtaining the intention decoupling degree includes: Determine the probability of each intention of the user statement; According to the standard deviation of the probability, and the maximum probability value, and the difference between the maximum probability and the minimum probability, the intention decoupling degree is obtained; The intention decoupling degree is positively correlated with the standard deviation of the probability, the maximum probability value and the difference.

2. The intelligent customer service interaction dialogue method based on semantic understanding according to claim 1, characterized in that, After obtaining the influence weight of each dominant intention in each round of interactive dialogue, the intelligent customer service interactive dialogue method based on semantic understanding further comprises: For a certain round of interactive dialogue, determine the round number proportion of the interactive dialogue in which the dominant intention of the round of interactive dialogue appears in all subsequent rounds of interactive dialogue; Determine the relevance of the intention recognition result of the round of interactive dialogue and the next round of interactive dialogue; According to the solution state factor of the round of interactive dialogue, correct the influence weight of the round of interactive dialogue; The solution state factor is obtained by the round number proportion and the correlation, and the solution state factor is positively correlated with the round number proportion and the correlation.

3. The intelligent customer service interaction dialogue method based on semantic understanding according to claim 1, characterized in that, The number of occurrences of the dominant intention in all previous rounds of interactive dialogue is used to correct the initial user mixed intention separation performance, which includes: According to the number of occurrences of the dominant intention in all previous rounds of interactive dialogue, a correction coefficient is obtained; The correction coefficient is inversely related to the number of occurrences; According to the correction coefficient, the initial user mixed intention separation performance is corrected.

4. The intelligent customer service interaction dialogue method based on semantic understanding according to claim 1, characterized in that, The process of obtaining the influence weight includes: According to the target user mixed intention separation performance of the dominant intention in the round of interactive dialogue and the round of the round of interactive dialogue, an influence weight of the dominant intention in the round of interactive dialogue is obtained; the influence weight is positively correlated with the target user mixed intention separation performance and the round of the round of interactive dialogue.

5. The intelligent customer service interactive dialogue method based on semantic understanding according to claim 1, characterized in that, The process of obtaining the dynamic feature of the influence weight comprises: Obtaining the slope of each influence weight in the change curve of the influence weight of the dominant intention; The average of the slope is calculated as the dynamic feature of the influence weight.

6. The intelligent customer service interaction dialogue method based on semantic understanding according to claim 5, characterized in that, The process of obtaining the influence weight change trend comprises: According to the dynamic feature and the number of occurrences of the dominant intention in all rounds of interactive dialogue, an influence weight change trend of the dominant intention is obtained; the influence weight change trend of the dominant intention is positively correlated with the dynamic feature and the number of occurrences of the dominant intention in all rounds of interactive dialogue.

7. The intelligent customer service interactive dialogue method based on semantic understanding according to claim 1, characterized in that, The intelligent customer service interactive dialogue method based on semantic understanding further comprises: The influence weight and the influence weight change trend of each dominant intention are integrated into a structured context feature vector as an input of a natural language understanding model to train the natural language understanding model.

Images