A distributed-based intelligent customer service database updating method

Abstract

The application relates to the technical field of data processing, in particular to an intelligent customer service database updating method based on distribution, which comprises the following steps: matching an initial node of a fault tree based on configured customer service dialogue data; dividing process nodes according to the customer service dialogue data corresponding to the initial node, generating an execution state label by using the rule type and the text content of each process node; generating a communication portrait corresponding to the process node based on the execution state label of the current process node, the rule type violation record and the branch parameter of the current process node; comparing the intention recognition deviation between each process node and the adjacent process node by using the communication portraits of the process nodes, correcting the abnormal branch node according to the branch node corresponding to the intention recognition deviation; and performing node association analysis on the corrected branch node to adjust the branch position of the branch node in the fault tree; and the efficiency and accuracy of customer service database updating are realized.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of data processing, and in particular to an intelligent customer service database updating method based on distribution. BACKGROUND

[0002] In actual life, customer service business is an important content in the marketing system of enterprise products or services. In order to improve the customer service business level, customer service data of each business team of the customer service business needs to be collected offline, so as to analyze the business handling and business operation of each business team. However, it is found in practice that the accuracy of the existing customer service data is insufficient in a complex dialogue scene, and there is inconsistency in multi-node updating, which leads to long updating time and low updating efficiency of the customer service database.

[0003] For example, Chinese Patent Publication No. CN115544135A discloses a customer service data intelligent analysis method and device. The method comprises: collecting initial customer service data corresponding to customer service business from a customer service data source system associated with a data analysis platform according to pre-determined customer service business; inputting the initial customer service data into a data analysis model corresponding to the data analysis platform for analysis to obtain customer service index data corresponding to the customer service business, and outputting the customer service index data to the front end of the data analysis platform to enable the front end to display the customer service index data.

[0004] For example, Chinese Patent Publication No. CN120216707A discloses a customer service data quality inspection method, device, equipment and medium based on dynamic reasoning. The method comprises: analyzing customer service dialogue data to generate structured dialogue data containing intent labels and key problem nodes; according to the intent labels in the structured dialogue data, generating a framework by searching and enhancing, recalling matching domain knowledge from a standard knowledge base in real time to construct a dynamic context knowledge graph; inputting the dynamic context knowledge graph, using a thinking chain prompt template to guide a large model to perform step-by-step logical reasoning, and outputting a preliminary quality inspection conclusion chain; verifying the assertion nodes in the preliminary quality inspection conclusion chain to generate a traceable final quality inspection conclusion; based on the final quality inspection conclusion, calculating the quantitative score of the customer service dialogue data and generating a visual thinking chain report.

[0005] The existing technology respectively explains that the data indicators in the customer service data are analyzed to determine the combination processing of the underlying data, and the key positions in the customer service data are judged by step-by-step reasoning to identify the conflict alarm nodes of the customer service data. However, the existing technology ignores the related problems under multi-node synchronous updating, which leads to the inability to adjust the current customer service data according to the deviation of the user intent, and further causes the lag of the customer service database updating and the inability to respond to the business changes in time. SUMMARY

[0006] In order to solve the above technical problems, the technical scheme adopted by the present application is: a distributed-based intelligent customer service database updating method, comprising: S1, based on the configured customer service dialogue data, matching the initial node of the fault tree according to the question type of the customer service dialogue.

[0007] S2, according to the customer service dialogue data corresponding to the initial node, dividing the process node according to the business logic sequence, generating the execution state label corresponding to each process node by using the rule type and text content under each process node.

[0008] S3, based on the execution state label of the current process node, judging whether the current process node has a branch node, extracting the branch parameter corresponding to the branch node, and generating the communication portrait corresponding to the process node after merging the rule type violation record of the current process node and the branch parameter.

[0009] S4, using the communication portrait of each process node, comparing the intention recognition deviation between each process node and the adjacent process node, and correcting the abnormal branch node according to the branch node corresponding to the intention recognition deviation.

[0010] S5, performing node association analysis on the corrected branch node, and adjusting the branch position of the branch node in the fault tree according to the message hit rate of each branch node.

[0011] The beneficial effects of the present application are: first, the present application generates an execution state label by rule classification, and processes each branch node by combining horizontal / vertical rule fusion, to prevent path misjudgment in each node process flow, which limits the data updating efficiency of each node binding, and improves the customer service response efficiency and accuracy.

[0012] Second, the present application corrects the abnormal branch node by intention similarity deviation and time interval deviation, and adjusts the topology structure of the fault tree, to reduce the problem of reduced node updating efficiency caused by service processing time length, improve the accuracy of the communication portrait constructed for each user, and the consistency problem of the relative position of each node after correction. BRIEF DESCRIPTION OF DRAWINGS

[0013] The present application will be further described below in conjunction with the drawings and examples.

[0014] Figure 1 It is a process schematic diagram of a distributed-based intelligent customer service database updating method.

[0015] Figure 2 It is a process schematic diagram of step S1 of a distributed-based intelligent customer service database updating method.

[0016] Figure 3is a flowchart of step S2 of the distributed-based intelligent customer service database updating method.

[0017] Figure 4 is a flowchart of step S3 of the distributed-based intelligent customer service database updating method.

[0018] Figure 5 is a flowchart of step S4 of the distributed-based intelligent customer service database updating method.

[0019] Figure 6 is a flowchart of step S5 of the distributed-based intelligent customer service database updating method. DETAILED DESCRIPTION

[0020] Embodiments of the present application are described in detail below. The embodiments described below are exemplary only, and are not to be construed as limiting the present application. Unless otherwise defined, scientific and technical terms used in the embodiments have the same meaning as commonly understood by one of ordinary skill in the art to which the application belongs. Unless otherwise required by context, singular terms shall include pluralities and vice versa. The singular forms "a", "an", and "the" include plural referents unless the context clearly indicates otherwise.

[0021] Reference Figure 1 A distributed-based intelligent customer service database updating method includes: S1, based on the configured customer service dialogue data, matching the initial node of the fault tree according to the question type of the customer service dialogue.

[0022] S2, according to the customer service dialogue data corresponding to the initial node, dividing the flow nodes in the order of business logic, using the rule type and text content under each flow node to generate the execution state label corresponding to each flow node.

[0023] S3, based on the execution state label of the current flow node, judging whether the current flow node has a branch node, extracting the branch parameters corresponding to the branch node, and generating the communication portrait corresponding to the flow node after merging the rule type violation record of the current flow node and the branch parameters.

[0024] S4, using the communication portrait of each flow node, comparing the intention recognition deviation between each flow node and the adjacent flow node, and performing abnormal branch node correction according to the branch node corresponding to the intention recognition deviation.

[0025] S5, performing node association analysis on the corrected branch node, and adjusting the branch position of the branch node in the fault tree according to the message hit rate of each branch node.

[0026] When describing the service attitude and other contents, the initial description content or the question consulted using the guide capsule is used to map the corresponding data to the initial node of the fault tree, and the content corresponding to the initial node is determined according to the classification of the current question type.

[0027] When the type of the question in the current customer service conversation is obtained through the guided capsule, the guided capsule can be directly positioned to the initial node. However, when the content described after the guided capsule is used is inconsistent with the guided capsule or the question is directly answered, the corresponding initial node cannot be directly positioned, as shown in step S1. Figure 2 The implementation manner of step S1 includes the following steps.

[0028] S12, based on the effective classification information, mapping the problem type under the effective classification information to the mapping relationship between the initial node of the fault tree.

[0029] When the time recall retrieval is performed, the actual intention of the user in the current customer service conversation is viewed according to the processing mode of the text recall and the vector recall.

[0030] The text recall performs keyword matching based on the customer service conversation data, quickly finds the classification label related to the current description content, such as the content corresponding to the after-sales service, and the vector recall converts the customer service conversation data into a vector in a manner of semantic similarity calculation, views the current classification label, and considers the classification label corresponding to the text recall and the vector recall at this time as the effective classification information. As for the type of the question under the effective information, the text corresponding to the effective classification information is viewed, and the corresponding question type is obtained through the text interpretation.

[0031] That is, the implementation manner of step S11 further includes the following steps.

[0032] After the initial node of the fault tree is updated, the latest round of customer service conversation data is subjected to vector regression, and the classification label under the text recall and the vector regression is set as the effective classification information.

[0033] Preferably, the text recall directly associates the known classification label by scanning the dialogue content through a keyword dictionary or a rule engine; the vector recall adopts a BERT, Word2Vec or the like model, mines the potential classification label by calculating the semantic similarity between vectors, embeds the dialogue text into a high-dimensional space, and performs similarity matching with a pre-labeled classification label vector library to determine the most relevant label, and considers this classification label as the classification label of the vector recall.

[0034] Preferably, based on the dialogue text corresponding to the effective classification information, the core meaning of the text is extracted through natural language understanding (NLU) technology, and is mapped to the pre-defined question type, such as the user intention related to the return and exchange process, product damage and the like.

[0035] Preferably, the identification information of the user after each consultation is identified during each processing of the customer service dialogue data. This information is marked in the note information processed by the customer service for the user. By directly searching for the keywords related to the problem solving, it is determined whether the user's problem is solved after the customer service dialogue. In the scenario where the problem is not solved, the time interval between the current access time point and the last access time point is used as the basis for judgment. When the corresponding time interval is greater than the preset time interval, it is considered that the current problem has been solved. The preset time interval can be 1 day. Generally, the user's problem can be solved within 1 day after communication with the customer service. If it cannot be solved, the latest round of data of the user can be obtained to update the current customer service dialogue data and complete the storage and construction of the customer service related data.

[0036] In an embodiment of the present application, the rule type described above is the rule type of the basic behavior of each flow node, such as the combination of fast speech, talk over, mute, response timeout, etc. The rule type of the content form of sensitive words and illegal operations is taken as the execution status of each flow node as the description content. At the same time, each flow node is divided into service attitude level, trip, and rule type about answer quality such as incomplete answer according to the text speed of the current user and the customer service during communication. At this time, the flow nodes divided by the answer flow constitute a fault tree, and the effectiveness of the customer service in answering questions under different problems and intentions is counted by analyzing the current fault tree for the user's answer and consultation questions.

[0037] As for the above text content, it is recorded in the form of voice text, text, etc. under the corresponding rule type. The execution state label such as talk over + response timeout is generated in real time to explain the state of the user's expression of the problem and the customer service's expression of the problem. Then the number of labels and the dialogue duration under each flow node are recorded.

[0038] It should be noted that when the rule type of trip occurs, it represents the problem of fracture and logical jump in the current service flow, such as not executing step by step according to the normal business flow node, and skipping some key links. For example, the customer asks about the return flow, and the customer service does not guide the customer to confirm the order information first, but directly jumps to the step of filling out the application form, resulting in the customer needing to repeatedly supplement the information. At the same time, trip can be the problem of interrupting the dialogue due to the customer service dialogue data not meeting the customer's expectations. These parts will affect the accuracy of the current customer service dialogue processing.

[0039] Preferably, the flow nodes can be divided based on the business logic sequence of consultation→guidance→answer→confirmation, or the logic sequence of related processing of the corresponding problem type is directly obtained, and the data of the current customer service and user dialogue is divided into multiple node chain structure flow nodes.

[0040] For example,Figure 3 As shown, the implementation of step S2 includes: S21, sequentially determining the basic behavior rule, the content form rule and the answer quality rule of the current flow node under the rule classification, and setting the rule trigger state of each rule classification.

[0041] S22, verifying the occurrence number and the repetition number of each rule classification after rule triggering according to the rule trigger state of each rule classification, and sequentially setting the label frequency of each rule classification.

[0042] S23, performing rule matching on each rule classification according to the label frequency of each rule classification, combining the data after rule matching, and setting the execution state label.

[0043] Preferably, the rule trigger state of the basic behavior rule can be expressed as: fast speech, which is the customer voice > 4 words per second, or the customer interrupts the user when the user has not finished speaking, or the user interrupts the customer when the customer has not finished speaking; silence, which is that the user continues to be silent for > 8 seconds, or response timeout, which is that the customer response time > 10 seconds, to illustrate the trigger state of the basic behavior rule in the real-time communication scenario between the user and the customer. At this time, the basic behavior rule not only contains the current summarized content, but also contains other content, and these data can be processed by using the trigger form set in advance in the database.

[0044] As for the rule trigger state of the content form rule, it is sensitive words, which is when sensitive words or taboo words appear, or violation operation, which is to guide the customer to perform a non-standard process, to illustrate the problem in the content form.

[0045] As for the rule trigger state of the answer quality rule, it can be to identify the service attitude according to the negative emotion and positive emotion expressed by the current vocabulary, that is, to count the emotional intensity of the positive vocabulary and the negative vocabulary explained by the customer, and to select the service attitude grade according to the mean value of the emotional intensity. At this time, the emotional dictionary can be used for vocabulary query to mark the emotional intensity contained in the current customer text; the trip is to check the interruption and jumping, breaking part of the dialogue, the answer is not comprehensive, which is to explain the number of customer problems, and when the customer's answer to the user's question has a coverage less than 80%, it is considered that the answer is not comprehensive.

[0046] For the above-mentioned checking of the occurrence number and the repetition number of each rule classification, the purpose is mainly to determine the rules triggered by the current flow node and the number of times of triggering, to determine the place where problems are likely to occur under the current flow node, and the deficiencies of the customer in the dialogue. After synchronizing these data to the database, it is convenient for subsequent evaluation of the customer's work and adjustment of the current dialogue.

[0047] As for the rule matching processing, the weak real-time existing under the current processing, the label conflict management, and the parameter identification missing are reduced to prevent the existence of partial abnormal bias after data combination.

[0048] Preferably, when performing rule matching on each rule category, the processing mode includes: taking the label under the basic behavior rule as the initial label, and sequentially calculating the matching degree of the initial label under different label frequencies with the labels under the content form rule and the answer quality rule. During the matching degree calculation, the text can be converted into a vector to calculate the similarity based on the current matching degree value. The calculation formula can be calculated based on the cosine similarity form. Then, the part with the maximum value in the matching degree of the initial label with the content form rule and the answer quality rule is compared, which is taken as the cover label of the initial label. The initial label and the corresponding cover label are taken as the execution state label. This part emphasizes the conflict-free part of the label combination. The label combination with strong real-time and relevance is combined as the output execution state label to strengthen the different and description bias existing in the current customer service dialogue. If there is no relevant content with the rule category at this time, it means that the current customer service dialogue adopts the same way as the preset script, and the stored customer service data does not need to be updated.

[0049] Preferably, the set label frequency groups the execution state label to output in the form of grouping to determine the processing process of multiple groups about the current rule category.

[0050] In an embodiment of the present application, the collected branch parameters are recorded in the content corresponding to the rule type and the text type, and the monitoring description state between the customer service and the user during the call is completed to build the current communication portrait.

[0051] At this time, the branch node determines whether the current flow node has extended script content according to the recognized intention of the corresponding text in the flow node, compares the intention composition degree of the extended content under these multiple branches, selects the branch content more suitable for the current scene, and improves the update and storage efficiency of the script data. The branch parameter represents the rule type and the text content contained in the branch node.

[0052] In step S3, the branch node under the current flow node is first identified, and data update and iteration processing are performed according to the content contained in the branch node. Meanwhile, the branch node also represents the extended description and other description modes of the text corresponding to the current flow node. These contents are used to verify whether there is other replaceable description content in the current customer service dialogue scene to enrich the composition of the customer service dialogue data.

[0053] As Figure 4As shown, the implementation manner of the branch node processing in step S3 includes: S31, extracting a branch condition table corresponding to the process node according to the execution state label of the current process node, generating a multi-level branch node according to the rule type triggered by the branch condition table, and integrating multiple processes according to the flow transfer order of the multi-level branch node; the branch condition table is used to describe the rule types covered in the current process node, and multiple branch nodes are generated according to the value of each rule type, for example, if there is a specific service consultation, a branch description of the corresponding service is triggered, for example, card application → card application inquiry branch, at this time, according to the order of process flow transfer triggered by the rule type, a multi-level branch node is generated, or the branch node existing under the normal inquiry of the current process can be used to generate multiple process flow transfer related branch nodes according to the normal process flow transfer, and the corresponding process node can be regarded as the branch node of the current process node according to the state reservation and data persistence processing of the current execution state label under multiple processing steps, to describe the level related to process flow transfer, data temporary storage reservation and rule type, and the branch node is divided into levels.

[0054] S32, the execution state label of the integrated branch node is reserved, and the process nodes covered after reservation are regarded as the branch nodes of the current process node; the process nodes pointed to by each rule type are configured into the fault tree to determine the dialogue flow control of the current fault tree, and data storage is performed according to the form of flow transfer.

[0055] As for the implementation manner of forming the communication portrait in step S3, the following contents are included: S33, hot spot marking is performed on the branch node satisfying the execution state label, rule type violation records of the current branch node are extracted at the time point of the hot spot marking, and the rule type violation records are merged into the current process node.

[0056] S34, when merging the branch parameters, the frequency of the adjacent level branch node at the upper level branch node and the lower level branch node is judged, and the branch parameters of each branch node are added to the current process node according to the connection relationship between the adjacent two level branch nodes.

[0057] S35, when the query times of the branch node are greater than the preset query times threshold, the rule type of the corresponding branch node is taken as the main rule type of the current process node, and the branch node of the adjacent level under the main rule type is taken as the secondary rule type, and the main rule type and the secondary rule type are used to form the communication portrait of the current process node.

[0058] S36, when the query times of the branch node are greater than the preset query times threshold, the order of the rule type triggered by each branch node is used to form the communication portrait of the process node according to the level of each branch node.

[0059] It should be noted that the rule type violation record is used to record the negative labels of the current customer service, such as sensitive words, taboo words, word stealing, response timeout, etc., which represent that the customer service does not fully answer the user's question. The negative labels are counted at the corresponding process node to indicate the problems that may occur during the customer service conversation. As for the branch parameter merging, the number of times of using different branches is recorded to determine the effect of different branch nodes in answering the current process node for different problem types.

[0060] Preferably, the execution state label satisfies the execution state label of the current process node and the preset rule type, such as detecting the sensitive word label to trigger the content compliance rule. At this time, the number of triggering times of the sensitive word label is set as a hot spot mark in its surge period, that is, when the label exists in a certain time period within a day or a week and is greater than the average number of triggering times in the corresponding time period, it is marked as a hot spot mark. At this time, it is merged into the corresponding process node to prompt to review the current customer service conversation data to update the related customer service answer content and the related record of the use of the speech data.

[0061] Preferably, the query times are the number of times of being triggered or accessed, and the preset query times threshold is used to set the critical value of the branch node under different rule types. When the value is exceeded, it means that this part of the content is the main content described in the current process node. The preset query times threshold can be set based on the unit time period, such as selecting the maximum allowed access value of each rule type as the preset query times threshold, that is, taking the upper limit value of the confidence interval of the 95% confidence level of the normal customer service conversation as the preset query times threshold currently set, to indicate the form of the record of the problem and the related content in the customer service conversation.

[0062] Preferably, when generating the communication portrait of the process node, the fitting relationship in the current branch node also needs to be determined to update the current customer service conversation data according to the fitting relationship.

[0063] For example, based on the content quantitatively calculated under each rule type, the branch nodes corresponding to each rule type are preliminarily associated, and the newly added content of the branch node under each rule type is used to establish a new connection. The branch nodes that are recombined and connected are regarded as the current output portrait to form the data connection structure of the current customer service conversation update.

[0064] Therefore, when the communication portrait is output, the implementation manner further includes: taking the associated branch node as a connection point, and sequentially performing horizontal rule fusion on each connection point according to the rule type.

[0065] The branch node under the transverse rule fusion is combined with the new content of the branch node and other branch nodes. If there is a rule type conflict in the combined branch node, the newly added content is quantified, and the vertical rule fusion is performed. The intersection set of the transverse rule fusion and the vertical rule fusion is used to reset the connection relationship of each branch node. The branch node with adjusted connection relationship is re-input to the corresponding process node.

[0066] Preferably, the part regarded as a connection point represents a plurality of branch nodes for contents such as seizing a word, response timeout, etc. during process flow transfer. These parts that have normally performed process flow transfer are then processed in a transverse fusion manner, and similar business flow processing contents are combined to improve the branches that each problem type can contain after time regression of the customer service database, thereby improving the accuracy of question and answer. During transverse rule fusion, the plurality of branch nodes currently regarded as connection points are compared in a semantic similarity manner, and branch nodes in similar situations are combined. When the semantic similarity is greater than 0.6, the branch nodes in the transverse direction are fused. The selected value represents that the branch nodes from a plurality of problem types are relatively consistent when flowing to the next step. When the value is 0.6, it represents that the branch nodes described by the branch nodes are similar except for individual extreme cases. The branch nodes in the transverse direction can be connected. The branch nodes connected from a plurality of problem types can be obtained to illustrate that the branch nodes in the plurality of problem types are related. Ultimately, the data set can be expanded, the data conflict and contradiction can be reduced, the real-time data updating requirement of the customer service dialogue can be enhanced, a relatively complete user communication portrait can be constructed, and personalized processing of the customer service dialogue can be formed based on the formed portrait.

[0067] As for the rule type conflict, after transverse rule fusion, the branch node at the corresponding position adds new description content when the customer service dialogue data is updated to the database. For example, after the customer service dialogue, the user introduces picture or link content and text dialogue form content, and non-current problem type description appears in the voice description. At this time, it is necessary to compare whether the newly added content conflicts with the current rule type. At this time, the data about the rule type set in the database is compared in a semantic similarity manner to further illustrate that in the scene where the semantic similarity is less than 0, and the semantic similarity between the current rule type and the rule type after transverse rule fusion is less than 0 for multiple times, the content is irrelevant, and vertical fusion is performed. The newly added content is sorted according to the semantic similarity value from small to large, and is marked at the branch node for vertical rule fusion to illustrate the update of the newly added content in the branch node.

[0068] Vertical rule fusion aims to ensure the integrity of business processes and prevent data loss in scenarios where the conversation flow jumps or the topic is interrupted during customer service dialogues. This would prevent newly added content from being related to the current configured branch nodes, thus preventing the generation of user profiles from completing the service loop for users. The data combined from the two methods will emphasize the breadth and relative depth of user behavior under different question types, so as to complete the correlation and update between multiple questions and types in customer service dialogues, and form a sufficiently complete customer service database.

[0069] The subsequent selection of horizontal rule fusion and vertical rule fusion involves jointly describing the connection relationship between the new content and the branch nodes, as well as the connection relationship of the branch nodes under different question types, to obtain the relative data of the updated branch nodes, thereby assisting in the efficiency of subsequent manual review.

[0070] Ultimately, process nodes such as order inquiries, refund applications, and fault reporting are formed. Each process node includes corresponding basic behaviors, content compliance, and service quality-related rule types. After that, branch nodes are formed for each branch type, as well as lower-level branch nodes related to the transfer and process initiation under the corresponding function. When there are related branch nodes under the problem type required by the process node, the corresponding content is combined with annotations to further expand the customer service data.

[0071] In one embodiment of the present invention, when comparing intent recognition deviations, multiple branch nodes are compared based on the communication profiles corresponding to the process nodes, and the content such as intent compositeness and intent similarity is compared to determine the weight and recognition deviation under each branch process.

[0072] Intent similarity represents the semantic similarity between the text of the branch node under the current communication profile and the user's intent. The user intent is identified based on the question type configured on the current process node, or it can be calculated based on the user intent represented by the execution status label mapped to the data. Intent complexity represents the ratio of the repetition of the corresponding user intent under multiple nodes, indicating the consistency of the user's intent before and after. At this time, the intent recognition deviation between adjacent process nodes is compared to identify the degree of difference in the overall intent needs of the current user under multiple branches, and whether there are branch nodes with completely consistent intents, so as to further adjust the customer service database for setting the dialogue response for users according to semantic similarity.

[0073] like Figure 5 As shown, the implementation of step S4 includes: S41, based on the intent similarity and intent compositeness of each branch node under the current process node, comparing the deviation between the branch node under the current process node and the corresponding upstream and downstream nodes.

[0074] S42, determine the deviation between the current branch node and the upstream and downstream nodes based on the compared intention similarity deviation and intention composite deviation; the intention similarity deviation is obtained by calculating the difference between the intention similarity of the branch node under the current process node and the upstream and downstream nodes, and the intention composite is obtained by determining whether the branch node under the current process node represents consistent content based on the intention similarity of the branch node under the current process node, and the number of times of occurrence of the intention of the branch node is counted as the number of times of occurrence of all intentions of the current branch node and the upstream and downstream nodes, thereby obtaining the intention composite; and the intention composite deviation is obtained by calculating the intention composite of the branch node with the upstream and downstream nodes respectively; the branch node under the standard condition when the intention similarity deviation and the intention composite deviation are greater than the standard condition is regarded as the processing deviation.

[0075] It should be noted that the standard condition represents the average value of the intention similarity deviation and the intention composite deviation obtained from the historical data, which represents whether the content of the current customer service dialogue is inconsistent, topic jumping, etc.

[0076] S43, set the time sequence association path of the deviation item based on the relative position of the deviation item in the process node and the dependence relationship between the upstream and downstream nodes; use the time sequence association path of the deviation item for path analysis to determine the time sequence interaction bias between the branch nodes.

[0077] Preferably, the time sequence association path is based on the relationship between the current branch node after the execution of the upstream and downstream nodes, and is arranged in the order of time sequence execution, such as verifying whether there is redundancy in the upstream and downstream nodes when executing in time sequence, if there is redundancy, the corresponding branch node is configured to a separate path, if it is not a redundant node, it is combined into a time sequence association path according to the order of the upstream and downstream nodes along with the time sequence, if it is a redundant node, the redundant node is merged into a single node according to the time interval between the redundant node and the adjacent branch node, and the two or more branch nodes represented by the redundant node are combined to form a sub-path of the time sequence association path.

[0078] For example, after the user inquires about the delivery time, the customer service replies "48 hours", and then the user asks again "how long can I receive the goods?", the customer service again repeats "expected to arrive within 4 days"; at this time, the branch node corresponding to the delivery time is regarded as a redundant node, and the corresponding redundant node is merged into the time sequence association path, the redundant node represents the repeated answer content under the same intention of the user, and this part is regarded as the node of redundant reply processing. At the same time, the redundant node forms a time sequence association path with respect to other branch nodes, and the redundant node is regarded as a sub-path of the time sequence association path, which is regarded as an independent branch content. At this time, the time sequence association path formed is regarded as the time sequence interaction bias of the branch node.

[0079] S44, when the branch node of the multi-time sequence interaction bias is obtained, the order of each branch node is adjusted according to the bias degree of the time sequence interaction bias, and the branch node after the order is adjusted is output.

[0080] For the quantitative index of the time sequence interaction bias, the time interval deviation and the intention similarity deviation are normalized, and the weighted sum is taken as the bias degree. The branch node with high bias degree is regarded as the node with priority order, and the branch node order is adjusted according to the bias degree value in sequence.

[0081] Then, weights are set for the branch nodes on the time sequence correlation path. Each weight is based on the frequency of the time interval deviation and the intention similarity deviation of the corresponding branch node in the historical data, and the total frequency of using the corresponding branch node in the historical data, to indicate the weight of the branch node on the time sequence correlation path.

[0082] Preferably, the time interval deviation represents the difference between the average processing time between nodes and the expected time. The expected time represents the processing time of the corresponding branch node in the historical data. This time interval deviation will be used as a standard to measure the effectiveness of the current dialogue and whether the user is easy to communicate. After being set as a label, the user's communication portrait is enriched.

[0083] That is, the processing mode of step S44 includes: calculating the time interval deviation of the average processing time and the expected time of the plurality of branch nodes on the time sequence correlation path, combining the time interval deviation and the intention similarity deviation as the bias degree of the time sequence interaction bias, adjusting the order of each branch node on the time sequence correlation path according to the value of the bias degree from large to small, and outputting the branch node after the order is adjusted.

[0084] In an embodiment of the application, after the branch position of the fault tree is adjusted, the adjusted fault tree is stored in the database to complete the update storage of the data after the user dialogue, so as to facilitate the subsequent adjustment of the stored dialogue template and the warning record during the customer service dialogue according to the difference between the customer service and the user, and finally realize the update iteration of the customer service database and complete the closed loop iteration processing of the knowledge base.

[0085] When performing node correlation analysis, the branch nodes can also be expressed differently according to the periodic factor, which is regarded as a periodic factor node, and the state of the plurality of flow nodes under node transfer is adjusted to generate the fault tree.

[0086] Preferably, the message hit rate represents the number of conversations in which node A is followed by node B divided by the total number of occurrences of node A, and is used to illustrate the message hit rate of the branch node in the state transition process using the frequent item set method; then the periodic nodes existing in the frequent item set are extracted, and after the periodic state is bound to the branch node, the state bound to the branch node is determined according to the frequent item set method, and the node level configured in the current fault tree is adjusted to complete the storage of the relevant dialogue and conversation data under the fault tree.

[0087] As shown in Figure 6 , the implementation mode of step S5 includes: S51, based on the message hit rate of the current branch node, checking the frequent item set of each branch node, and extracting node pairs of multiple branch nodes in the frequent item set that meet the minimum support threshold; wherein the minimum support threshold is set to 5%, which is used to illustrate the node pair of the execution process of the current branch node from A→B, and these node pairs will represent the frequent item after the node exception is corrected.

[0088] S52, binding the node pair of the branch node to a periodic type, and determining the occurrence frequency deviation of the node pair in a specific periodic type; the occurrence frequency deviation is the number of occurrences of the node pair in the corresponding period minus the average number in the period, and the difference is divided by the average number in the period to illustrate the deviation relative to the average case; at this time, the periodic type includes four periodic forms of day, week, month and season, and the occurrence frequency deviation can be represented as the occurrence frequency in the current day minus the average number per day to illustrate the deviation, and the same for each week, each month and each season, and the value deviating from the overall average number in a single periodic type is used to illustrate the periodic factor in each branch node, and if the occurrence probability deviation value is too large, it will indicate that there is a significant periodic state in the corresponding period; when the occurrence probability deviation is greater than the preset deviation threshold, the corresponding node pair is marked as a periodic node.

[0089] At this time, the preset deviation threshold can be based on the occurrence probability deviation under normal circumstances, and the upper limit value of the confidence interval set at a 95% confidence level is set as the preset deviation threshold to adjust the configuration of the branch node in the fault tree.

[0090] S53, when the branch node is a periodic node, the position of the corresponding periodic node is split into an independent subtree of the current fault tree; if the branch node is a periodic node, it means that the customer service conversation data set for the corresponding branch node is commonly used in a specific period, and these related data need to be stored and processed separately, and the periodic content is updated in time according to the independent subtree, so as to prevent the update of the customer service database from affecting the description of the periodic content.

[0091] S54, when the branch node is not a periodic node, the branch adjustment is performed with the occurrence probability deviation of the branch node and the message hit rate, and the adjusted fault tree is output.

[0092] The branch adjustment manner can be hierarchical promotion, hierarchical sinking, branch merging, etc. The hierarchical promotion can be that if the current message hit rate is continuously greater than 40% and the periodic deviation is less than 20%, the current branch node can be merged into the previous branch node at this time. The 40% and 20% used at this time are only used for illustration, and the actual values need to be viewed in combination with the multiple branches under the current flow node. The average value of the message hit rate calculated by the user in the normal dialogue selection main branch node is used as the value of the message hit rate judgment at this time, and the periodic deviation is also set by using the average value of the periodic deviation under the main branch node to indicate whether the current branch node is the main branch. If both of the above two values are satisfied, the level of the current branch node is promoted to indicate that the content described by the current node is a relatively main level, and these data need to be updated to the customer service database in time to improve the accuracy of the customer service answer to the user.

[0093] The hierarchical sinking is to judge the continuous message hit rate, and to construct a confidence interval with the message hit rate at this time. In the case that the message hit rate is continuously less than the lower limit value of the confidence interval, the message hit rates of the last three periods are used as the basis when the conditions are met, and the hierarchical level of the current branch node is lowered to make it a leaf node to describe the relative content of the current fault tree.

[0094] The branch merging is to indicate that the current branch node has two child nodes, and the hit rates of the two child nodes are close, such as less than 5% or less than 1%. The two branch nodes under this branch node are merged to form a selectable child node, such as the form of A→[B / C] selector, to further modify the node at the branch position of the fault tree.

[0095] Although the embodiments of the present application have been shown and described above, it should be understood that the above embodiments are exemplary and should not be construed as limiting the present application. Those skilled in the art can make changes, modifications, replacements and variations to the above embodiments within the scope of the present application, which are still covered by the protection scope of the present application.

Claims

1. A distributed intelligent customer service database update method, characterized in that, include: S1, based on the configured customer service dialogue data, matches the initial node of the fault tree according to the question type in the customer service dialogue; S2, based on the customer service dialogue data corresponding to the initial node, divide the process nodes according to the business logic order, and generate the execution status label corresponding to each process node using the rule type and text content under each process node; S3, based on the execution status label of the current process node, determine whether there is a branch node in the current process node, extract the branch parameters corresponding to the branch node, and generate the communication profile corresponding to the process node after merging the rule type violation record and branch parameters of the current process node. S4. Utilize the communication profile of each process node to compare the intent recognition deviation between each process node and its adjacent process nodes, and correct abnormal branch nodes according to the branch nodes corresponding to the intent recognition deviation. S5 performs node association analysis on the corrected branch nodes and adjusts the branch position of each branch node in the fault tree based on the message hit rate of each branch node.

2. The method for updating a distributed intelligent customer service database according to claim 1, characterized in that, The implementation methods for step S1 include: S11, Based on the acquired customer service dialogue data, perform time-sensitive recall retrieval of the customer service dialogue data to determine the valid classification information of the current user during the consultation; S12 uses a fault tree mapped by effective classification information to construct a mapping relationship between the problem types under the effective classification information and the initial nodes of the fault tree.

3. The method for updating a distributed intelligent customer service database according to claim 2, characterized in that, The implementation of step S11 also includes: At the current access time, retrieve the latest round of customer service dialogue data, perform text retrieval on the customer service dialogue data, and update the initial node of the fault tree with the latest round of customer service dialogue data. After updating the initial node of the fault tree, vector regression is performed on the latest round of customer service dialogue data. Effective classification information is set using text recall and classification labels under vector regression.

4. The method for updating a distributed intelligent customer service database according to claim 1, characterized in that, Step S2 can be implemented in the following ways: S21, sequentially determine the basic behavior rules, content format rules, and answer quality rules of the current process node under the rule category, and set the rule triggering status of each rule category; S22, based on the rule triggering status of each rule category, verify the number of occurrences and repetitions of each rule category after the rule is triggered, and set the tag frequency of each rule category in turn; S23. Based on the label frequency of each rule category, perform rule matching for each rule category, combine the matched data, and set the execution status label.

5. The method for updating a distributed intelligent customer service database according to claim 4, characterized in that, When performing rule matching for each rule category, the processing methods include: Using the tags under the basic behavior rules as initial tags, the matching degree of the initial tags under different tag frequencies is calculated with the tags under the content format rules and the answer quality rules in turn. The largest part of the matching degree of the initial tags with the content format rules and the answer quality rules in turn is compared and used as the covering tag of the initial tag. The initial tag and the corresponding covering tag are used as the execution status tag.

6. The method for updating a distributed intelligent customer service database according to claim 1, characterized in that, Step S3 can be implemented in the following ways: S31. Based on the execution status label of the current process node, extract the branch condition table corresponding to the process node, generate multi-level branch nodes according to the rule type triggered by the branch condition table, and integrate multiple processes according to the flow order of the multi-level branch nodes. S32, retain the execution status label of the integrated branch node, and regard the process node covered by the retention as the branch node of the current process node; configure the process node pointed to by each rule type into the fault tree; S33, mark hotspots for branch nodes that meet the execution status label, extract rule type violation records of the current branch node at the time point of hotspot marking, and merge the rule type violation records into the current process node; S34, when merging branch parameters, determine the frequency of the branch nodes at the previous and next level branch nodes, and add the branch parameters of each branch node to the current process node based on the connection relationship between the two levels of branch nodes. S35, when the number of queries for a branch node exceeds the preset query threshold, the rule type of the corresponding branch node is taken as the main rule type of the current process node, and the branch nodes at the adjacent level under the main rule type are taken as the secondary rule types. The main rule type and the secondary rule type are used to form the communication profile of the current process node. S36, when the number of queries for a non-branch node exceeds the preset query threshold, the communication profile of the process node is formed by the order in which the rule types corresponding to each branch node are triggered and the hierarchy of each branch node.

7. The method for updating a distributed intelligent customer service database according to claim 6, characterized in that, When outputting a communication profile, the implementation methods also include: Using related branch nodes as connection points, the connection points are sequentially merged horizontally according to the rule type; The branch nodes under the horizontal rule fusion are combined with other branch nodes based on the newly added content of the branch nodes. If there is a rule type conflict in the combined branch nodes, the newly added content quantified at the time of conflict is used for vertical rule fusion. The connection relationship of each branch node is reset based on the intersection set of the horizontal rule fusion and the vertical rule fusion. The branch nodes with the adjusted connection relationship are then re-entered into the corresponding process nodes.

8. The method for updating a distributed intelligent customer service database according to claim 1, characterized in that, Step S4 can be implemented in the following ways: S41, Based on the intent similarity and intent complexity of each branch node under the current process node, compare the deviation between the branch node under the current process node and the corresponding upstream and downstream nodes respectively. S42, using the compared intent similarity deviation and intent composite deviation, branch nodes whose calculated intent similarity deviation and intent composite deviation are greater than those in the standard case are considered as deviation items to be processed; S43, Based on the relative position of the deviation item in the process node and the dependency relationship between upstream and downstream nodes, set the temporal association path of the deviation item; use the temporal association path of the deviation item to perform path analysis and determine the temporal interaction bias between each branch node; S44: When obtaining branch nodes with multiple temporal interaction biases, adjust the order of each branch node according to the degree of bias of the temporal interaction bias, and output the branch nodes after the adjustment.

9. The method for updating a distributed intelligent customer service database according to claim 8, characterized in that, The processing method of step S44 includes: Calculate the time interval deviation between the average processing time of multiple branch nodes on the temporal association path and the expected time. Combine the time interval deviation with the intent similarity deviation as the bias degree of the temporal interaction bias. Adjust the order of each branch node on the temporal association path from large to small according to the bias degree value, and output the branch nodes after the adjustment order.

10. The method for updating a distributed intelligent customer service database according to claim 1, characterized in that, Step S5 can be implemented in the following ways: S51, based on the message hit rate of the current branch node, examine the frequent itemsets of each branch node, and extract the node pairs of multiple branch nodes that meet the minimum support threshold from the frequent itemsets; S52, bind the node pairs of the branch nodes by period type, and determine the frequency deviation of the node pairs in a specific period type; S53, when the branch node is a periodic node, split the position of the corresponding periodic node into an independent subtree of the current fault tree; S54, when the branch node is not a periodic node, adjust the branch based on the probability deviation of the branch node's occurrence and the message hit rate, and output the adjusted fault tree.

Images