Method, device and medium for processing intent conflict of customer service dialogue

Abstract

The application discloses a customer service dialogue-oriented intention conflict processing method and device, electronic equipment and a readable storage medium, and belongs to the technical field of intelligent customer service. The method comprises the following steps: when a system obtains current dialogue text, a first candidate intention set is generated based on the current dialogue text; then, a first feature vector set and an intention relationship graph are constructed based on the first candidate intention set; the first candidate intention set is screened based on the first feature vector set, and a second candidate intention set is obtained; then, a second feature vector set is determined from the first feature vector set; and finally, the main intention and the auxiliary intention are determined from the second candidate intention set based on the second feature vector set and the intention relationship graph. The application can accurately identify the main intention of the customer, realize intention resolution under multi-intention conflict, and improve the user experience of the intelligent customer service.

Description

Technical Field

[0001] This application belongs to the field of intelligent customer service technology, specifically relating to an intent conflict handling method for customer service dialogue, an intent conflict handling device for customer service dialogue, an electronic device, and a readable storage medium. Background Technology

[0002] As large language models are increasingly applied in scenarios such as intelligent customer service, work order triage, knowledge-based Q&A, and automated responses, user intent recognition is gradually shifting from rule matching, keyword recognition, and single-label classification to semantic understanding methods based on large language models. Large language models are better at understanding context, ellipsis, cross-sentence reference, and complex requests, making them suitable for handling the colloquial, incomplete, and diverse user expressions in real customer service conversations.

[0003] However, in real-world customer service scenarios, user requests are often not singular intentions, but rather encompass multiple business objectives or processing conditions simultaneously. For example, a single sentence might simultaneously involve issues such as price, installation, logistics, refunds, exchanges, or after-sales progress. These candidate intentions may be parallel, or they may have a sequential order, dependencies, or business conflicts. Therefore, the system needs to further determine the primary and secondary relationships and the processing order.

[0004] Existing methods typically use the top-1 results from the large model directly, or sort by raw confidence scores. Other methods ask the model to provide primary and secondary intents through prompts, or use fixed priority rules for correction. Still other methods rely primarily on the user's current input text, depending only on the large model's raw output or simple sorting. However, in complex, multi-intent scenarios, such as when the user's expression contains omissions, transitions, or complex requests, these methods are prone to inaccurate intent graph identification, resulting in low accuracy in intent conflict resolution and a poor user experience. Summary of the Invention

[0005] To solve the above-mentioned technical problems, this application is implemented as follows: In a first aspect, embodiments of this application provide a method for handling intent conflicts in customer service dialogues, the method comprising: When the current dialogue text is obtained, a first candidate intent set is generated based on the current dialogue text; A first feature vector set and an intent relationship graph are constructed based on the first candidate intent set; the first feature vector set includes each candidate intent in the first candidate intent set. Figure 1 A corresponding feature vector is provided, and the intent relationship graph is used to characterize the relationship between any two candidate intents; The first candidate intent set is filtered based on the first feature vector set to obtain a second candidate intent set, and the second feature vector set is determined from the first feature vector set. The main intent and auxiliary intent are determined from the second candidate intent set based on the second feature vector set and the intent relationship graph.

[0006] In one or more embodiments, generating a first candidate intent set based on the current dialogue text includes: Obtain the historical dialogue text from the most recent N rounds, the last historical idea graph, the confidence level of the last historical idea graph, and the current business stage status; N is a positive integer; Intent identification is performed using the current dialogue text, the historical dialogue text, the last historical idea graph, the confidence level, and the current business stage state to obtain a first candidate intent set including at least one candidate intent.

[0007] In one or more embodiments, constructing a first feature vector set and an intent relationship graph based on the first candidate intent set includes: The associated factors of each first candidate intent in the first candidate intent set are quantified to obtain multiple features corresponding to each first candidate intent; The feature vectors of each first candidate intent are constructed using multiple features of each first candidate intent, resulting in a first feature vector set that includes the feature vectors corresponding to each first candidate intent. Construct an intent relationship graph using each first candidate intent.

[0008] In one or more embodiments, the step of filtering the first candidate intent set based on the first feature vector set to obtain a second candidate intent set, and determining the second feature vector set from the first feature vector set, includes: The first score of the corresponding first candidate intent is calculated using each first feature vector in the first feature vector set; The first candidate intent set is filtered out when the first score is less than the first score threshold, resulting in the second candidate intent set; The first feature vector set is filtered based on the second candidate intent set to obtain the second feature vector set.

[0009] In one or more embodiments, determining the intention graph and auxiliary intention from the second candidate intention set based on the second feature vector set and the intention relationship graph includes: The second score of each second feature vector in the second feature vector set is calculated using the intent relationship graph; The second candidate intent corresponding to the second feature vector with the highest second score is determined as the primary intent. Based on the intent relationship graph, auxiliary intents are determined from other second candidate intents.

[0010] In one or more embodiments, calculating the second score of each second feature vector in the second feature vector set using the intent relationship graph includes: For any target second feature vector in the second feature vector set, obtain the current business type; If the current business type is not a preset type, then the initial weight coefficient is obtained, and the second score of the target second feature vector is calculated using the initial weight coefficient and the intent relationship graph; If the current business type is the preset type, then obtain the initial weight coefficient; The initial weight coefficients are adjusted according to the preset features in the target second feature vector to obtain the target weight coefficients; The second score of the second feature vector of the target is calculated using the target weight coefficient and the intent relationship graph.

[0011] In one or more embodiments, determining the auxiliary intent from other second candidate intents based on the intent relationship graph includes: At least one third candidate intent is determined from other second candidate intents; the second score of the third candidate intent is less than the highest second score, and greater than the second score threshold. The intent relationship diagram is used to determine whether there is a conflict between each third candidate intent and the intent diagram; If a conflict exists, the third candidate intent that does not conflict will be identified as the auxiliary intent.

[0012] Secondly, embodiments of this application provide an intent conflict handling apparatus for customer service dialogues, the apparatus comprising: The generation module is used to generate a first set of candidate intents based on the current dialogue text when the current dialogue text is obtained; The construction module is used to construct a first feature vector set and an intent relationship graph based on the first candidate intent set; the first feature vector set includes each candidate intent in the first candidate intent set. Figure 1 A corresponding feature vector is provided, and the intent relationship graph is used to characterize the relationship between any two candidate intents; The first determining module is used to filter the first candidate intent set based on the first feature vector set to obtain a second candidate intent set, and to determine the second feature vector set from the first feature vector set. The second determining module is used to determine the main intent and auxiliary intent from the second candidate intent set based on the second feature vector set and the intent relationship graph.

[0013] In one or more embodiments, the generation module is specifically used for: Obtain the historical dialogue text from the most recent N rounds, the last historical idea graph, the confidence level of the last historical idea graph, and the current business stage status; N is a positive integer; Intent identification is performed using the current dialogue text, the historical dialogue text, the last historical idea graph, the confidence level, and the current business stage state to obtain a first candidate intent set including at least one candidate intent.

[0014] In one or more embodiments, the building module is specifically used for: The associated factors of each first candidate intent in the first candidate intent set are quantified to obtain multiple features corresponding to each first candidate intent; The feature vectors of each first candidate intent are constructed using multiple features of each first candidate intent, resulting in a first feature vector set that includes the feature vectors corresponding to each first candidate intent. Construct an intent relationship graph using each first candidate intent.

[0015] In one or more embodiments, the first determining module is specifically used for: The first score of the corresponding first candidate intent is calculated using each first feature vector in the first feature vector set; The first candidate intent set is filtered out when the first score is less than the first score threshold, resulting in the second candidate intent set; The first feature vector set is filtered based on the second candidate intent set to obtain the second feature vector set.

[0016] In one or more embodiments, the second determining module: The calculation submodule is used to calculate the second score of each second feature vector in the second feature vector set using the intent relationship graph; The idea graph determination submodule is used to determine the second candidate idea corresponding to the second feature vector with the highest second score as the main idea; The auxiliary intent determination submodule is used to determine the auxiliary intent from other second candidate intents based on the intent relationship graph.

[0017] In one or more embodiments, the idea diagram determining submodule is specifically used for: For any target second feature vector in the second feature vector set, obtain the current business type; If the current business type is not a preset type, then the initial weight coefficient is obtained, and the second score of the target second feature vector is calculated using the initial weight coefficient and the intent relationship graph; If the current business type is the preset type, then obtain the initial weight coefficient; The initial weight coefficients are adjusted according to the preset features in the target second feature vector to obtain the target weight coefficients; The second score of the second feature vector of the target is calculated using the target weight coefficient and the intent relationship graph.

[0018] In one or more embodiments, the auxiliary intent determination submodule is specifically used for: At least one third candidate intent is determined from other second candidate intents; the second score of the third candidate intent is less than the highest second score, and greater than the second score threshold. The intent relationship diagram is used to determine whether there is a conflict between each third candidate intent and the intent diagram; If a conflict exists, the third candidate intent that does not conflict will be identified as the auxiliary intent.

[0019] Thirdly, embodiments of this application provide an electronic device including a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the method described in the first aspect.

[0020] Fourthly, embodiments of this application provide a readable storage medium on which a program or instructions are stored, which, when executed by a processor, implement the steps of the method described in the first aspect.

[0021] In this embodiment, when the system obtains the current dialogue text, it generates a first candidate intent set based on the current dialogue text. Then, it constructs a first feature vector set and an intent relationship graph based on the first candidate intent set. The first candidate intent set is then filtered based on the first feature vector set to obtain a second candidate intent set. The second feature vector set is then determined from the first feature vector set. Based on the second feature vector set and the intent relationship graph, the main intent and auxiliary intent are determined from the second candidate intent set. In this way, a unified, multi-dimensional feature vector set is constructed for each candidate intent. Beyond the feature vectors, an intent relationship graph is further constructed, and the conflict edge weights and support edge weights between candidate intents are determined, describing the pairwise relationships and global competition structure between candidate intents. The feature vectors are then weighted and scored using the intent relationship graph and weight coefficients corresponding to different business types. Candidate intents are filtered based on the scores, and finally, the main intent and auxiliary intent are determined from the filtered candidate intents based on the scores. This accurately identifies the customer's main intent, achieves intent resolution under multiple intent conflicts, and improves the user experience of intelligent customer service. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the steps of an embodiment of an intent conflict resolution method for customer service dialogue according to this application. Figure 2 This is an example diagram of the intent relationship of this application; Figure 3 This is a structural block diagram of an embodiment of an intent conflict handling device for customer service dialogue according to this application. Detailed Implementation

[0023] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0024] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0025] The intent conflict handling for customer service dialogue provided in this application will be described in detail below with reference to the accompanying drawings, through specific embodiments and application scenarios.

[0026] As large language models are increasingly applied in scenarios such as intelligent customer service, work order triage, knowledge-based Q&A, and automated responses, user intent recognition is gradually shifting from rule matching, keyword recognition, and single-label classification to semantic understanding methods based on large language models. Large language models are better at understanding context, ellipsis, cross-sentence reference, and complex requests, making them suitable for handling the colloquial, incomplete, and diverse user expressions in real customer service conversations.

[0027] However, in real-world customer service scenarios, user requests are often not singular intentions, but rather encompass multiple business objectives or processing conditions simultaneously. For example, a single sentence might simultaneously involve issues such as price, installation, logistics, refunds, exchanges, or after-sales progress. These candidate intentions may be parallel, or they may have a sequential order, dependencies, or business conflicts. Therefore, the system needs to further determine the primary and secondary relationships and the processing order.

[0028] Existing methods typically use the top-1 results from the large model directly, or sort by raw confidence scores. Other methods ask the model to provide primary and secondary intents through prompts, or use fixed priority rules for correction. Still other methods rely primarily on the user's current input text, depending only on the large model's raw output or simple sorting. However, in complex, multi-intent scenarios, such as when the user's expression contains omissions, transitions, or complex requests, these methods are prone to inaccurate intent graph identification, resulting in low accuracy in intent conflict resolution and a poor user experience.

[0029] Reference Figure 1 This document illustrates a flowchart of an embodiment of an intent conflict resolution method for customer service dialogue, applicable to a customer service dialogue intent conflict resolution system. Users can input text into the system, which then performs intent recognition on the input text, determines the main intent and auxiliary intent, and responds to the user's input text based on the main intent and auxiliary intent. Specifically, the method may include the following steps: Step 101: When the current dialogue text is obtained, a first candidate intent set is generated based on the current dialogue text.

[0030] During a conversation between a user and customer service, when the system receives the current conversation text sent by the user, it can determine at least one candidate intent of the user from the conversation text. By integrating these candidate intents, a candidate intent set can be obtained (for ease of description, it is denoted as the "first candidate intent set"). Among them, the candidate intents are used to represent the user's intent that may exist in the conversation text.

[0031] In this embodiment of the application, generating a first candidate intent set based on the current dialogue text includes: Obtain the historical dialogue text from the most recent N rounds, the last historical idea graph, the confidence level of the last historical idea graph, and the current business stage status; N is a positive integer; Intent identification is performed using the current dialogue text, the historical dialogue text, the last historical idea graph, the confidence level, and the current business stage state to obtain a first candidate intent set including at least one candidate intent.

[0032] Specifically, the system can obtain the historical dialogue text of the most recent N rounds (N is a positive integer, for example, the historical dialogue text of the most recent 5 rounds), the last historical idea graph, the confidence level of the historical idea graph, and the current business stage status.

[0033] Among them, the most recent N rounds of historical dialogue text are the saved or encoded original dialogue content of the most recent N rounds, where N is a preset parameter that can be set according to actual needs; the last historical main graph is a structured summary of the processing result of the last round of dialogue, which complements the most recent N rounds of historical dialogue text; the current business stage status is used to characterize the current stage status of the business, such as "pre-sales", "in-sales", "after-sales", etc. The current business stage status can be determined primarily based on the reliable status of back-end business systems such as orders, payments, logistics, and after-sales work orders. User selection, the current page of the user's dialogue with customer service, and the user's statement can be used to assist in the determination. When the status of multiple data sources conflict, the current business stage status can be corrected according to the reliability and timestamp of the data source. If the current business stage status cannot be determined and cannot be corrected, then the current business stage status can be set to "unknown".

[0034] After obtaining the above information, intent recognition can be performed using this information and the current dialogue text to obtain at least one candidate intent. Integrating the candidate intents yields the first candidate intent set.

[0035] In the process of intent recognition, a large language model and a pre-trained intent recognition model can be used to perform semantic parsing on the current dialogue text, and then intent recognition can be performed on the semantic parsing to generate a first set of candidate intents. And, the semantic score corresponding to each candidate intent. For example, the first candidate intent set is {buy, return}, and the corresponding semantic scores are {0.83, 0.57}.

[0036] By using multiple pieces of information, such as the current dialogue text, historical dialogue text, the last historical idea graph, confidence level, and the current business stage status, the accuracy of intent recognition can be greatly improved, providing a guarantee for subsequent determination of idea graphs and auxiliary intents.

[0037] It should be noted that, in addition to using large language models and pre-trained intent recognition models for intent recognition, other methods or models can also be used for intent recognition. In practical applications, the specific method or model used for intent recognition can be set according to actual needs, and this is not limited in the application embodiments.

[0038] Furthermore, when outputting candidate intents, one of the following methods can be used: direct output from the model, output from a multi-label classification model, or output from a rule engine integrated with the model. Alternatively, any combination of the above methods can be used. In practical applications, the specific method of outputting candidate intents can also be set according to actual needs, and this application embodiment does not impose any restrictions on this.

[0039] Step 102: Construct a first feature vector set and an intent relationship graph based on the first candidate intent set; the first feature vector set includes each candidate intent in the first candidate intent set. Figure 1 A corresponding feature vector is provided, and the intent relationship graph is used to characterize the relationship between any two candidate intents.

[0040] After obtaining the first candidate intent set, a feature vector corresponding to each candidate intent in the first candidate intent set can be constructed. By integrating multiple feature vectors, a feature vector set (denoted as the "first feature vector set") can be obtained.

[0041] Feature vectors describe whether a single intent is valid. However, when multiple candidate intents exist simultaneously, relying solely on the features of a single candidate intent is insufficient to determine the relationships between them. These relationships can include conflicting and supporting relationships. For example, "purchase intention" and "return request" typically conflict in business direction, meaning the candidate intents for "purchase intention" and "return request" are in conflict. "Refund progress query" usually depends on "refund request," meaning the candidate intents for "refund progress query" and "return request" are supporting. "Event consultation" often co-occurs with "purchase consultation," meaning the candidate intents for "event consultation" and "purchase consultation" are also supporting. Therefore, after obtaining the first set of feature vectors, an intent relationship graph can be constructed using the candidate intents from the first set of candidate intents.

[0042] In this embodiment of the application, the step of constructing a first feature vector set and an intent relationship graph based on the first candidate intent set includes: The associated factors of each first candidate intent in the first candidate intent set are quantified to obtain multiple features corresponding to each first candidate intent; The feature vectors of each first candidate intent are constructed using multiple features of each first candidate intent, resulting in a first feature vector set that includes the feature vectors corresponding to each first candidate intent. Construct an intent relationship graph using each first candidate intent.

[0043] Specifically, for each candidate intent (denoted as "first candidate intent") in the first candidate intent set, its related factors can be quantified to form a feature representation that can be used for subsequent calculation and comparison, thereby obtaining multiple features corresponding to each first candidate intent. Among them, the related factors are those that can reflect the probability of the candidate intent being true, its business importance, and the degree of matching with the current dialogue scenario.

[0044] Furthermore, the types of characteristics can be divided into two categories: objective characteristics and individual characteristics.

[0045] Objective features are used to ensure the objectivity and computational stability of intent evaluation results. Their values ​​are typically derived from general model judgment results, historical data statistics, or context matching results. Individual features are used to reflect the processing differences under different business scenarios, making intent judgment results more in line with actual business needs. Their values ​​are typically derived from business rules, business priority settings, and user behavior information.

[0046] Among these, objective features can include "current semantic support", "historical prior support" and "contextual continuity".

[0047] Current-round semantic support: This feature characterizes the degree of semantic matching between the current user-input dialogue text and a candidate intent. It can be calculated using a generalized large language model, a pre-trained intent recognition model, or other semantic recognition models, and its value can be normalized to 0 to 1. A higher value indicates that the current dialogue text better supports the candidate intent. Specifically, a generalized large language model can be used to determine the degree of semantic association between the current dialogue text and the candidate intent.

[0048] Historical Priority: This feature characterizes the prevalence of candidate intents in historical data corresponding to the current business scenario and stage, ranging from 0 to 1. Historical Priority is calculated based on the proportion of intent categories that have been labeled, confirmed, or indirectly confirmed by business processing results in historical dialogues, rather than solely based on the occurrence rate of a single keyword. The same intent can correspond to multiple different expressions, and the same keyword may also correspond to different intents. For example, in the historical dialogues of a pre-sales consultation scenario, the "purchase" intent has a high confirmation rate, thus possessing a high historical priority in that scenario. This feature enables the system to make more stable judgments when the semantic scores of multiple candidate intents are similar, by combining the historical intent distribution of the corresponding scenario.

[0049] Contextual continuity: This represents the consistency between the candidate intent and the preceding main intent graph, and its value can be normalized to 0 to 1. The higher the consistency, the closer the candidate intent is to the preceding main intent. By introducing contextual continuity, the intent drift problem caused by incomplete expression in single sentences in multi-turn dialogues can be reduced.

[0050] Personality traits can include "business priorities" and "behavioral support".

[0051] Business Priority: This characteristic represents the importance of candidate intents in actual business operations. It can be preset according to business rules. The higher the importance of the intent, the higher the corresponding business priority value. For example, for high-risk or highly sensitive intents such as complaints or malfunctions, a higher business priority can be set to ensure that the system prioritizes identifying and processing such intents when judging multiple intents.

[0052] Behavioral Support: This feature characterizes whether a user's actions within the current dialogue text or related business process support a candidate intent. It can be calculated based on user behavior such as clicking, browsing, placing orders, submitting work orders, uploading vouchers, selecting menus, and staying on certain pages. For example, if a user spends a considerable amount of time on a product details page and clicks the "Buy Now" button multiple times, their behavior strongly supports the intent to "buy."

[0053] Thus, by quantifying the related factors of candidate intentions to obtain "objective features" and "personal features," a representation of the candidate intentions can be constructed based on these objective and personal features. For each candidate intention... i Construct a unified feature vector:

[0054] in, Used to describe candidate intent Its features include current-round semantic support, contextual continuity, historical prior knowledge, business priority, and behavioral support. Indicates the semantic support of the current round; Indicates the continuity of context, used to reflect the continuity relationship between the candidate intent and the preceding intent graph; Indicates business priority; Indicates the level of support for the behavior; This indicates the historical prior knowledge of the candidate intent. For example, if the candidate intent is "to buy". Fi The feature vector representing "purchase" .

[0055] It should be noted that, in addition to five-dimensional features, feature vectors can also include other features, such as slot completeness, user historical profile matching degree, knowledge base hit confidence, order status matching degree, etc., which means nine-dimensional features can be used. In practical applications, the specific dimensions of features in the feature vector can be set according to actual needs, and this application embodiment does not limit this.

[0056] Furthermore, after obtaining the first feature vector, the conflict or support relationships between each first candidate intent can be calculated to construct an intent relationship graph.

[0057] In addition to feature vectors, an intent conflict graph is further constructed. .in, V This represents the set of nodes that represent the candidate intents identified in the current text, with each node corresponding to a candidate intent; E Edges representing relationships between candidate intentions describe whether there are relationships such as conflict, dependency, co-occurrence, or mutual support between two intentions. Specifically, for any two candidate intentions... and The conflict edge weights can be calculated separately. With support for border rights .

[0058] 1) Conflict of border rights The degree of conflict between two candidate intentions is expressed as follows:

[0059] in: The degree of competition indicates the extent to which two candidates intend to vie for the position of the main idea in the current dialogue; The mutual exclusion level represents the degree of rejection between two candidate intentions in terms of business logic. Both values ​​are controlled between 0 and 0.5, therefore the conflict edge weight... The range is 0 to 1. When The larger the value, the less suitable the two candidate intents are as parallel primary requirements; when When the value is small, it means that two candidate intentions can be retained simultaneously. For example, the mutual exclusion between "purchase intention" and "return request" is usually high, so the conflict weight between the two is large.

[0060] 2) Support edge weights It is used to represent the support, dependency, or co-occurrence relationship between two candidate intentions, and is calculated as follows:

[0061] in, The degree of coexistence indicates whether two candidate intentions typically coexist in the business chain; Co-occurrence level represents the frequency with which two candidate intentions co-occur in historical customer service conversations. Similarly, and The range is 0 to 0.5, therefore, it supports edge weights. The range is 0 to 1. When The larger the value, the more suitable the two candidate intents are to be retained as related requirements; when... A smaller value indicates a weaker correlation between the two. For example, "refund application" and "refund progress query" usually have a dependency relationship, thus the support edge weight is higher.

[0062] In this way, the intention relationship graph constructed by the edge-level relationship structure can express the relationship between each candidate intention in an explicit way, so that the intention conflict has a computable and interpretable multi-source representation structure before resolution, which provides a guarantee for the subsequent determination of the main intention graph and auxiliary intention.

[0063] It should be noted that intent relationship diagrams can be constructed using explicit diagram structures, and Alternatively, relation matrices, tensors, learnable attention modules, rule tables, or cue word inference results can be used. In practical applications, any method that can model the conflict, dependency, co-occurrence, and process deviation relationships between candidate intents is applicable to the embodiments of this application, and the embodiments of this application do not limit this.

[0064] Step 103: Filter the first candidate intent set based on the first feature vector set to obtain a second candidate intent set, and determine the second feature vector set from the first feature vector set.

[0065] After obtaining the first set of feature vectors, the first candidate intent set can be filtered according to each of the first feature vectors. The first candidate intents with weaker semantic scores in the first candidate intent set are filtered out. In this way, the retained first candidate intents form the second set of candidate intents. The first set of feature vectors is then filtered according to the second set of candidate intents to obtain the second set of feature vectors.

[0066] In this embodiment of the application, the step of filtering the first candidate intent set based on the first feature vector set to obtain a second candidate intent set, and determining the second feature vector set from the first feature vector set, includes: The first score of the corresponding first candidate intent is calculated using each first feature vector in the first feature vector set; The first candidate intent set is filtered out when the first score is less than the first score threshold, resulting in the second candidate intent set; The first feature vector set is filtered based on the second candidate intent set to obtain the second feature vector set.

[0067] Specifically, since the first feature vector includes five features, the sum of the semantic scores of the five features can be calculated to obtain the score of the corresponding first candidate intent (denoted as the "first score"), that is, .

[0068] Then, from the first candidate intent set, candidate intents with a first score not less than a first score threshold are selected, and these candidate intents are integrated to obtain the second candidate intent set. Alternatively, candidate intents in the first candidate intent set with a first score less than the first score threshold are filtered, and the remaining candidate intents are integrated to obtain the second candidate intent set.

[0069] After obtaining the second candidate intent set, feature vectors corresponding to each candidate intent (denoted as "second candidate intent") in the second candidate intent set can be selected from the first feature vector set. These features are then integrated to obtain the second feature vector set.

[0070] In this way, the scores of each candidate intent are calculated by using the semantic scores of each feature in the feature vector, and the candidate intents with weaker intents are filtered out based on the scores, so as to obtain the candidate intents with stronger intents. This reduces the interference of intent recognition and improves the accuracy of subsequent determination of the main intent and auxiliary intent.

[0071] Step 104: Determine the main intent and auxiliary intent from the second candidate intent set based on the second feature vector set and the intent relationship graph.

[0072] After obtaining the second feature vector set, the score of each second feature vector in the second feature vector set can be calculated, and the main image and auxiliary intention can be determined from the second candidate intention set based on the score and the intention relationship graph.

[0073] The step of determining the main idea and auxiliary intention from the second candidate intention set based on the second feature vector set and the intention relationship graph includes: The second score of each second feature vector in the second feature vector set is calculated using the intent relationship graph; The second candidate intent corresponding to the second feature vector with the highest second score is determined as the primary intent. Based on the intent relationship graph, auxiliary intents are determined from other second candidate intents.

[0074] Specifically, for any feature vector in the second feature vector, the score (denoted as the "second score") can be calculated using the following formula:

[0075] in, Candidate intention i Compared with other candidate intentions k Cumulative value of the support relationship between them; Candidate intention i Compared with other candidate intentions k Cumulative value of conflict relationships between them; Candidate intention i The final score; ~ These are all weighting coefficients, among which, to These correspond to the current round's semantic support, contextual continuity, historical prior knowledge, business priority, and behavioral support, respectively. Corresponding cumulative support value, The cumulative value of corresponding conflict relationships.

[0076] After calculating the second score of all second feature vectors, the second candidate intent corresponding to the second feature vector with the highest second score can be determined as the primary intent. Then, auxiliary intents are determined from the other second feature vectors based on the intent relationship graph. Here, the intent graph represents the customer's most core and priority need, and the system should prioritize responding to or guiding the business process around this intent; related intents represent other valid needs expressed by the customer at the same time, or related needs that have a supporting, dependent, or complementary relationship with the intent graph.

[0077] In this way, by combining the semantic scores of each feature in the feature vector and the edge weights in the intent relationship graph with the weight coefficients, the scores of each candidate intent are further calculated, and the candidate intent corresponding to the highest score is determined as the primary intent, which further improves the accuracy of intent recognition and achieves accurate resolution of intent conflicts.

[0078] In this embodiment of the application, the step of calculating the second score of each second feature vector in the second feature vector set using the intent relationship graph includes: For any target second feature vector in the second feature vector set, obtain the current business type; If the current business type is not a preset type, then the initial weight coefficient is obtained, and the second score of the target second feature vector is calculated using the initial weight coefficient and the intent relationship graph; If the current business type is the preset type, then obtain the initial weight coefficient; The initial weight coefficients are adjusted according to the preset features in the target second feature vector to obtain the target weight coefficients; The second score of the second feature vector of the target is calculated using the target weight coefficient and the intent relationship graph.

[0079] Specifically, weighting coefficients can be of two types: initial weighting coefficients and dynamic weighting coefficients.

[0080] The initial weight coefficients can be obtained through training with historical labeled data, optimization of the validation set, or configuration of business rules, and are used to directly calculate the second score of the second feature vector when the current business type is not a preset type.

[0081] The dynamic weighting coefficients are the weighting coefficients obtained after dynamically adjusting the initial weighting coefficients. As mentioned earlier, the feature vector contains "individual features," including "business priority" and "behavioral support." These "individual features" are factors related to actual production applications, and their settings are relatively flexible. Because in In the calculation process, in addition to the defined feature vector of the candidate intent itself F This also includes the "supporting edge weights" and "conflicting edge weights" of the candidate intent's relationship with other candidate intents. Therefore, it is possible to combine the candidate intent's "individual characteristics" (including "business priority" and "behavioral support") to conduct... The weights support dynamic adjustment of edge weights and conflict edge weight coefficients.

[0082] For example, during e-commerce promotions, for candidate intents with high business priority and high behavioral support (such as "purchase intent"), and for candidate intents with lower business priority (such as "inquiry intent"), the "purchase intent" is... In the calculation, the coefficient of the supporting edge weight can be increased (e.g., from 0.2 to 0.4), while the coefficient of the conflicting edge weight can be decreased (e.g., from 0.3 to 0.2). This improves the final score of the candidate intent, making it prioritized in the business scenario. In other words, the score of "purchase intent" will ultimately be higher than that of "consultation intent," avoiding errors in determining the primary intent when intents conflict. At this point, the system can treat purchase as the primary intent graph and provide answers regarding purchase knowledge, avoiding focusing on the consultation level and wasting customer time and reducing the experience.

[0083] Furthermore, dynamic weighting coefficients can be enabled when preset business types (such as e-commerce activities). "Business priority" can be the primary criterion for judgment, that is, the "business priority" of the candidate intent is judged first to ensure that the "business priority" is high, and "behavioral support" is a secondary judgment factor (also relatively high). "Behavioral support" can be used as a secondary factor depending on the actual business situation.

[0084] It should be noted that dynamic weighting coefficients may overemphasize the effect of factors. Therefore, it is emphasized that they should be used when the current business type is a preset type. However, they can also be used in other situations where the actual business deems it necessary to use dynamic weighting coefficients. This application does not limit this.

[0085] Based on this, for any second feature vector in the second feature vector set (denoted as the "target second feature vector"), the current business type can be obtained. If the current business type is not a preset type, then the initial weight coefficient can be obtained, and the second score of the target second feature vector can be calculated using this initial weight coefficient and the intent relationship graph (supporting cumulative relationship value and cumulative conflict relationship value). .

[0086] If the current business type is a preset type (such as an e-commerce activity), then the initial weight coefficients can be obtained first. Then, the initial weight coefficients can be adjusted according to the preset features in the target second feature vector (such as business priority and behavioral support) to obtain the target weight coefficients. Finally, the target weight coefficients and the intent relationship graph (cumulative values ​​of support relationships and cumulative values ​​of conflict relationships) can be used to calculate the second score of the target second feature vector. .

[0087] In this way, by setting different weight coefficients for different business types, when calculating the score of each candidate intent, different weight coefficients can be selected according to different business types to calculate the score of each candidate intent, which further improves the accuracy of intent recognition and achieves accurate resolution of intent conflicts.

[0088] In this embodiment of the application, determining the auxiliary intent from other second candidate intents based on the intent relationship graph includes: At least one third candidate intent is determined from other second candidate intents; the second score of the third candidate intent is less than the highest second score, and greater than the second score threshold. The intent relationship diagram is used to determine whether there is a conflict between each third candidate intent and the intent diagram; If a conflict exists, the third candidate intent that does not conflict will be identified as the auxiliary intent.

[0089] Specifically, after the main idea is determined, at least one candidate intention (denoted as "third candidate intention") can be determined from other second candidate intentions in the second candidate intention set. The second score of the third candidate intention is less than the highest second score and greater than the second score threshold.

[0090] Then, based on the support edge weights and conflict edge weights in the intent relationship graph, it is determined whether each third candidate intent conflicts with the main intent graph. Third candidate intents with conflicts are filtered out, and third candidate intents without conflicts are determined as auxiliary intents.

[0091] Candidate intents that are below the second score threshold and do not have a clear supporting relationship with the intent graph are filtered out to avoid irrelevant intents interfering with the final identification, thus resolving intent conflicts.

[0092] In this way, since the support edge weights and conflict edge weights in the intent relationship graph clearly indicate whether there is a conflict between each candidate intent, when determining auxiliary intents, the auxiliary intents that are more dependent on the main intent graph can be determined by the support edge weights and conflict edge weights, thereby improving the accuracy of auxiliary intent identification and realizing the accurate resolution of intent conflicts.

[0093] For ease of understanding, please refer to Figure 2 The diagram shows an example of an intent relationship graph. It includes candidate intents such as "Event Inquiry," "Refund Progress Inquiry," "Exchange Request," "Logistics Inquiry," "Return Request," and "Purchase Intention." "Exchange Request" is the primary intent. "Event Inquiry" and "Refund Progress Inquiry" both have weak support edge weights with the intent graph, so they are not connected to the intent graph. "Logistics Inquiry" and "Purchase Intention" are both auxiliary intents. The support edge weight (dashed line) between "Logistics Inquiry" and the intent graph is 0.22, the support edge weight between "Purchase Intention" and the intent graph is 0.65, and the conflict edge weight (solid line) between "Return Request" and the intent graph is 0.72.

[0094] Furthermore, the following examples will illustrate this further.

[0095] Example 1: A customer enters, "My air conditioner just shipped, but I want to exchange it for a different model, so I need to process an exchange." The system parses three candidate intentions: "Exchange Request," "Logistics Inquiry," and "Purchase Intention." If the final scores are: [Exchange Request: 0.7], [Purchase Intention: 0.5], and [Logistics Inquiry: 0.2], and the minimum score threshold is 0.3, then "Logistics Inquiry" is removed. Since "Exchange Request" has the highest score, it is identified as the primary intention; "Purchase Intention" has reached the score threshold and does not have a strong conflict with "Exchange Request," so it is identified as a secondary intention. The system should prioritize processing the exchange issue and supplement the response with information about purchasing a new model or reordering.

[0096] Example 2: A customer inputs, "There seems to be an e-commerce promotion going on these days, I need to exchange the item and repurchase it." The system analyzes and identifies three candidate intents: "Exchange Request," "Purchase Intent," and "Promotion Inquiry." If the final scores are: [Exchange Request: 0.7], [Purchase Intent: 0.5], and [Promotion Inquiry: 0.4], then "Exchange Request" is identified as the primary intent. Since both "Purchase Intent" and "Promotion Inquiry" have reached the score threshold, and there is no strong conflict between them and "Exchange Request," and "Promotion Inquiry" can also supplement information on discounts in the repurchase scenario, "Purchase Intent" and "Promotion Inquiry" can be jointly identified as secondary intents. The system should prioritize answering the exchange processing method, while supplementing the repurchase path and current promotional information.

[0097] In this embodiment, when the system obtains the current dialogue text, it generates a first candidate intent set based on the current dialogue text. Then, it constructs a first feature vector set and an intent relationship graph based on the first candidate intent set. The first candidate intent set is then filtered based on the first feature vector set to obtain a second candidate intent set. The second feature vector set is then determined from the first feature vector set. Based on the second feature vector set and the intent relationship graph, the main intent and auxiliary intent are determined from the second candidate intent set. In this way, a unified, multi-dimensional feature vector set is constructed for each candidate intent. Beyond the feature vectors, an intent relationship graph is further constructed, and the conflict edge weights and support edge weights between candidate intents are determined, describing the pairwise relationships and global competition structure between candidate intents. The feature vectors are then weighted and scored using the intent relationship graph and weight coefficients corresponding to different business types. Candidate intents are filtered based on the scores, and finally, the main intent and auxiliary intent are determined from the filtered candidate intents based on the scores. This accurately identifies the customer's main intent, achieves intent resolution under multiple intent conflicts, and improves the user experience of intelligent customer service.

[0098] It should be noted that the execution entity of the intent conflict handling method for customer service dialogue provided in this application embodiment can be a virtual device, or a control module in the virtual device for executing the method for loading intent conflict handling for customer service dialogue. This application embodiment uses the execution of the method for loading intent conflict handling for customer service dialogue by a virtual device as an example to illustrate the intent conflict handling method for customer service dialogue provided in this application embodiment.

[0099] Reference Figure 3 The diagram illustrates a structural block diagram of an embodiment of an intent conflict handling device for customer service dialogue, which may specifically include the following modules: The generation module 301 is used to generate a first candidate intent set based on the current dialogue text when the current dialogue text is obtained; Construction module 302 is used to construct a first feature vector set and an intent relationship graph based on the first candidate intent set; the first feature vector set includes each candidate intent in the first candidate intent set. Figure 1 A corresponding feature vector is provided, and the intent relationship graph is used to characterize the relationship between any two candidate intents; The first determining module 303 is used to filter the first candidate intent set based on the first feature vector set to obtain a second candidate intent set, and to determine the second feature vector set from the first feature vector set. The second determining module 304 is used to determine the main idea and auxiliary intention from the second candidate intention set based on the second feature vector set and the intention relationship graph.

[0100] In this embodiment of the application, the generation module is specifically used for: Obtain the historical dialogue text from the most recent N rounds, the last historical idea graph, the confidence level of the last historical idea graph, and the current business stage status; N is a positive integer; Intent identification is performed using the current dialogue text, the historical dialogue text, the last historical idea graph, the confidence level, and the current business stage state to obtain a first candidate intent set including at least one candidate intent.

[0101] In this embodiment of the application, the construction module is specifically used for: The associated factors of each first candidate intent in the first candidate intent set are quantified to obtain multiple features corresponding to each first candidate intent; The feature vectors of each first candidate intent are constructed using multiple features of each first candidate intent, resulting in a first feature vector set that includes the feature vectors corresponding to each first candidate intent. Construct an intent relationship graph using each first candidate intent.

[0102] In this embodiment of the application, the first determining module is specifically used for: The first score of the corresponding first candidate intent is calculated using each first feature vector in the first feature vector set; The first candidate intent set is filtered out when the first score is less than the first score threshold, resulting in the second candidate intent set; The first feature vector set is filtered based on the second candidate intent set to obtain the second feature vector set.

[0103] In this embodiment of the application, the second determining module: The calculation submodule is used to calculate the second score of each second feature vector in the second feature vector set using the intent relationship graph; The idea graph determination submodule is used to determine the second candidate idea corresponding to the second feature vector with the highest second score as the main idea; The auxiliary intent determination submodule is used to determine the auxiliary intent from other second candidate intents based on the intent relationship graph.

[0104] In this embodiment of the application, the idea mapping determination submodule is specifically used for: For any target second feature vector in the second feature vector set, obtain the current business type; If the current business type is not a preset type, then the initial weight coefficient is obtained, and the second score of the target second feature vector is calculated using the initial weight coefficient and the intent relationship graph; If the current business type is the preset type, then obtain the initial weight coefficient; The initial weight coefficients are adjusted according to the preset features in the target second feature vector to obtain the target weight coefficients; The second score of the second feature vector of the target is calculated using the target weight coefficient and the intent relationship graph.

[0105] In this embodiment of the application, the auxiliary intent determination submodule is specifically used for: At least one third candidate intent is determined from other second candidate intents; the second score of the third candidate intent is less than the highest second score, and greater than the second score threshold. The intent relationship diagram is used to determine whether there is a conflict between each third candidate intent and the intent diagram; If a conflict exists, the third candidate intent that does not conflict will be identified as the auxiliary intent.

[0106] In this embodiment, when the system obtains the current dialogue text, it generates a first candidate intent set based on the current dialogue text. Then, it constructs a first feature vector set and an intent relationship graph based on the first candidate intent set. The first candidate intent set is then filtered based on the first feature vector set to obtain a second candidate intent set. The second feature vector set is then determined from the first feature vector set. Based on the second feature vector set and the intent relationship graph, the main intent and auxiliary intent are determined from the second candidate intent set. In this way, a unified, multi-dimensional feature vector set is constructed for each candidate intent. Beyond the feature vectors, an intent relationship graph is further constructed, and the conflict edge weights and support edge weights between candidate intents are determined, describing the pairwise relationships and global competition structure between candidate intents. The feature vectors are then weighted and scored using the intent relationship graph and weight coefficients corresponding to different business types. Candidate intents are filtered based on the scores, and finally, the main intent and auxiliary intent are determined from the filtered candidate intents based on the scores. This accurately identifies the customer's main intent, achieves intent resolution under multiple intent conflicts, and improves the user experience of intelligent customer service.

[0107] The virtual device in this application embodiment can be a device, or a component, integrated circuit, or chip in a terminal. The device can be a mobile electronic device or a non-mobile electronic device. For example, a mobile electronic device can be a mobile phone, tablet computer, laptop computer, PDA, in-vehicle electronic device, wearable device, ultra-mobile personal computer (UMPC), netbook, or personal digital assistant (PDA), etc. A non-mobile electronic device can be a server, network attached storage (NAS), personal computer (PC), television (TV), ATM, or self-service machine, etc. This application embodiment does not impose specific limitations.

[0108] This application also provides an electronic device, including a processor, a memory, and a program or instructions stored in the memory and executable on the processor. When the program or instructions are executed by the processor, they implement the various processes of the above-described embodiment of the intent conflict handling method for customer service dialogue and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0109] It should be noted that the electronic devices in the embodiments of this application include the mobile electronic devices and non-mobile electronic devices described above.

[0110] This application also provides a readable storage medium storing a program or instructions. When the program or instructions are executed by a processor, they implement the various processes of the above-described embodiment of the intent conflict handling method for customer service dialogue and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0111] The processor is the processor in the electronic device described in the above embodiments. The readable storage medium includes computer-readable storage media, such as computer read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0112] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0113] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0114] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A method for handling intent conflicts in customer service dialogues, characterized in that, include: When the current dialogue text is obtained, a first candidate intent set is generated based on the current dialogue text; A first feature vector set and an intent relationship graph are constructed based on the first candidate intent set; the first feature vector set includes feature vectors that correspond one-to-one with each candidate intent in the first candidate intent set, and the intent relationship graph is used to characterize the relationship between any two candidate intents; The first candidate intent set is filtered based on the scores of the first feature vector set and the candidate intents to obtain the second candidate intent set, and the second feature vector set is determined from the first feature vector set based on the second candidate intent set. The main intent and auxiliary intent are determined from the second candidate intent set based on the second feature vector set and the intent relationship graph.

2. The method for handling intent conflicts in customer service dialogue according to claim 1, characterized in that, The generation of the first candidate intent set based on the current dialogue text includes: Obtain the historical dialogue text from the most recent N rounds, the last historical idea graph, the confidence level of the last historical idea graph, and the current business stage status; N is a positive integer; Intent identification is performed using the current dialogue text, the historical dialogue text, the last historical idea graph, the confidence level, and the current business stage state to obtain a first candidate intent set including at least one candidate intent.

3. The method for handling intent conflicts in customer service dialogue according to claim 1, characterized in that, The construction of a first feature vector set and an intent relationship graph based on the first candidate intent set includes: The associated factors of each first candidate intent in the first candidate intent set are quantified to obtain multiple features corresponding to each first candidate intent; The feature vectors of each first candidate intent are constructed using multiple features of each first candidate intent, resulting in a first feature vector set that includes the feature vectors corresponding to each first candidate intent. Construct an intent relationship graph using each first candidate intent.

4. The method for handling intent conflicts in customer service dialogues according to claim 1, characterized in that, The step of filtering the first candidate intent set based on the scores of the first feature vector set and the candidate intents to obtain a second candidate intent set, and determining the second feature vector set from the first feature vector set based on the second candidate intent set, includes: The first score of the corresponding first candidate intent is calculated using each first feature vector in the first feature vector set; The first candidate intent set is filtered out when the first score is less than the first score threshold to obtain the second candidate intent set; The first feature vector set is filtered based on the second candidate intent set to obtain the second feature vector set.

5. The method for handling intent conflicts in customer service dialogues according to claim 1, characterized in that, The step of determining the main idea and auxiliary intention from the second candidate intention set based on the second feature vector set and the intention relationship graph includes: The second score of each second feature vector in the second feature vector set is calculated using the intent relationship graph; The second candidate intent corresponding to the second feature vector with the highest second score is determined as the primary intent. Based on the intent relationship graph, auxiliary intents are determined from other second candidate intents.

6. The method for handling intent conflicts in customer service dialogue according to claim 5, characterized in that, The step of calculating the second score of each second feature vector in the second feature vector set using the intent relationship graph includes: For any target second feature vector in the second feature vector set, obtain the current business type; If the current business type is not a preset type, then the initial weight coefficient is obtained, and the second score of the target second feature vector is calculated using the initial weight coefficient and the intent relationship graph; If the current business type is the preset type, then obtain the initial weight coefficient; The initial weight coefficients are adjusted according to the preset features in the target second feature vector to obtain the target weight coefficients; The second score of the second feature vector of the target is calculated using the target weight coefficient and the intent relationship graph.

7. The method for handling intent conflicts in customer service dialogue according to claim 5, characterized in that, The step of determining the auxiliary intent from other second candidate intents based on the intent relationship graph includes: At least one third candidate intent is determined from other second candidate intents; the second score of the third candidate intent is less than the highest second score, and greater than the second score threshold. The intent relationship diagram is used to determine whether there is a conflict between each third candidate intent and the intent diagram; If a conflict exists, the third candidate intent that does not conflict will be identified as the auxiliary intent.

8. An intent conflict handling device for customer service dialogue, characterized in that, include: The generation module is used to generate a first set of candidate intents based on the current dialogue text when the current dialogue text is obtained; A construction module is used to construct a first feature vector set and an intent relationship graph based on the first candidate intent set; the first feature vector set includes feature vectors that correspond one-to-one with each candidate intent in the first candidate intent set, and the intent relationship graph is used to characterize the relationship between any two candidate intents; The first determining module is used to filter the first candidate intent set based on the scores of the first feature vector set and the candidate intents to obtain a second candidate intent set, and to determine the second feature vector set from the first feature vector set based on the second candidate intent set. The second determining module is used to determine the main intent and auxiliary intent from the second candidate intent set based on the second feature vector set and the intent relationship graph.

9. The intent conflict handling device for customer service dialogue according to claim 8, characterized in that, The generation module is specifically used for: Obtain the historical dialogue text from the most recent N rounds, the last historical idea graph, the confidence level of the last historical idea graph, and the current business stage status; N is a positive integer; Intent identification is performed using the current dialogue text, the historical dialogue text, the last historical idea graph, the confidence level, and the current business stage state to obtain a first candidate intent set including at least one candidate intent.

10. The intent conflict handling apparatus for customer service dialogue according to claim 8, characterized in that, The building module is specifically used for: The associated factors of each first candidate intent in the first candidate intent set are quantified to obtain multiple features corresponding to each first candidate intent; The feature vectors of each first candidate intent are constructed using multiple features of each first candidate intent, resulting in a first feature vector set that includes the feature vectors corresponding to each first candidate intent. Construct an intent relationship graph using each first candidate intent.

11. An electronic device, characterized in that, It includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the intent conflict resolution method for customer service dialogue as described in claims 1-7.

12. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a processor, implement the steps of the intent conflict resolution method for customer service dialogue as described in claims 1-7.

Images