Intelligent customer service guide method, system, device and medium based on difficulty learning

Abstract

The application relates to an intelligent customer service guiding method, system, device and medium based on difficulty learning. The method comprises the following steps: obtaining a plurality of interactive texts of a target user, determining an intention understanding text and an emotion understanding text of each interactive text; according to a preset text screening algorithm, screening a target difficulty text of a difficulty guiding demand from the plurality of interactive texts; according to a preset guiding model, determining a target guiding text of the target difficulty text, and according to difficulty analysis of the target difficulty text, determining a target language model corresponding to the target difficulty text from a plurality of candidate language models; and inputting the target guiding text and the target difficulty text into the target language model to obtain a target interactive text corresponding to the target difficulty text.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, specifically to an intelligent customer service guidance method, system, device, and medium based on difficulty learning. Background Technology

[0002] Currently, intelligent customer service is widely used in the customer service industry to serve users' inquiries and pre-sales and after-sales services. However, the complexity of human dialogue makes there are many difficult texts in chat interaction, which will reduce the response speed and accuracy of intelligent customer service. Especially in the recognition of user intent and emotion, intelligent customer service cannot meet all user needs. Human customer service is still needed to replace intelligent customer service in communication with users when intelligent customer service cannot provide effective services, in order to improve user satisfaction with services.

[0003] Therefore, when intelligent customer service encounters content that is difficult for them to understand, accurately identifying the user's intentions and emotions to improve customer service efficiency and user satisfaction has become an urgent problem to be solved. Summary of the Invention

[0004] The primary objective of this application is to address at least one of the aforementioned problems by providing a method, system, device, and medium for intelligent customer service guidance based on difficulty learning.

[0005] To achieve the various objectives of this application, the following technical solution is adopted:

[0006] A difficulty-based intelligent customer service guidance method provided for one of the purposes of this application includes the following steps:

[0007] The process involves acquiring multiple interactive texts from the target user and determining the intent-understanding text and emotion-understanding text for each interactive text. Based on a pre-defined text filtering algorithm, a target difficulty text for the required difficulty level is selected from the multiple interactive texts. According to a pre-defined guidance model, a target guidance text for the target difficulty text is determined. Based on the difficulty analysis of the target difficulty text, a target language model corresponding to the target difficulty text is determined from multiple candidate language models. The target guidance text and the target difficulty text are then input into the target language model to obtain the target interactive text corresponding to the target difficulty text.

[0008] In an optional embodiment, multiple interactive texts from the target user are acquired, and the intent-understanding text and sentiment-understanding text for each interactive text are determined, as follows:

[0009] For each obtained interaction text, a multi-turn interaction range of the interaction text is determined, question texts and answer texts in the multi-turn interaction range are obtained, semantic features, intent features and emotion features are extracted for each question text and each answer text, context dependency features are extracted for the multi-turn interaction range of each interaction text, intent understanding texts in the interaction text are determined according to the context dependency features, the semantic features and the intent features, and emotion understanding texts in the interaction text are determined according to the context dependency features, the semantic features and the emotion features, the intent understanding texts are used to represent the accuracy of the intelligent customer service in understanding the intent of the user question, and the emotion understanding texts are used to represent the accuracy of the intelligent customer service in understanding the emotion of the user.

[0010] In an optional embodiment, target difficulty texts of difficulty guide requirements are screened from the plurality of interaction texts according to a preset text screening algorithm, and the target difficulty texts are specifically as follows:

[0011] Based on the difference algorithm and the matching algorithm, the matching degree of the text difficulty and the matching degree of the understanding result corresponding to each interaction text are calculated according to the intent understanding text and the emotion understanding text.

[0012] According to the matching degree of the text difficulty and the matching degree of the understanding result, the target difficulty texts of difficulty guide requirements are screened from the plurality of interaction texts.

[0013] In an optional embodiment, the target difficulty texts of difficulty guide requirements are screened from the plurality of interaction texts according to the difference algorithm and the matching algorithm, and the matching degree of the text difficulty and the matching degree of the understanding result corresponding to each interaction text are calculated according to the intent understanding text and the emotion understanding text, and the target difficulty texts are specifically as follows:

[0014] For each interactive text, the mean of the text difference between the intent-understanding text of the interactive text and the intent-recognition text of multiple preset standard difficulty texts is calculated to obtain the difference parameter corresponding to the interactive text. According to preset text matching rules, the number of matching keywords, length similarity, and syntactic structure similarity between the intent-understanding text of the interactive text and the standard difficulty texts are calculated. The keyword number, length similarity, and syntactic structure similarity are then standardized and weighted to obtain the matching coefficient of the difference parameter. Finally, the product of the difference parameter, the matching coefficient, and the text information weight of the interactive text is calculated to obtain the matching degree of the text difficulty corresponding to the interactive text. The weight of the text information content is proportional to the difficulty of the interactive text. For each interactive text, the average similarity between the emotion-understanding text of the interactive text and the emotion recognition text of multiple preset historical interactive samples is calculated to obtain the first similarity parameter of the emotion-understanding text in the interactive text. For each interactive text, the average similarity between two emotion-understanding texts in adjacent time periods within the multi-round interaction range is calculated to obtain the second similarity parameter of the emotion-understanding text in the interactive text. The product of the first similarity parameter, the second similarity parameter, the context association weight, and the emotion weight is calculated to obtain the matching degree of the understanding result corresponding to the interactive text.

[0015] In an optional embodiment, the target difficulty text for difficulty guidance is selected from multiple interactive texts based on the matching degree of the text difficulty and the matching degree of the comprehension result, including the following steps:

[0016] Calculate the first summation mean of the matching scores of all understanding results, and the second summation mean of the matching scores of all text difficulties. Calculate the covariance matrix of the first and second summation mean values ​​based on the accuracy and response speed of user feedback. Determine the preset first and second matching score thresholds based on the correlation characterization of text difficulty and understanding results using the covariance matrix. Based on the condition that the matching score of the understanding results is greater than the first matching score threshold, filter the top K interactive texts sorted by matching score to obtain candidate interactive texts. Sort the candidate interactive texts from largest to smallest according to the matching score of text difficulty, and filter out the candidate interactive texts whose matching score of text difficulty is greater than the second matching score threshold to obtain the target difficulty text.

[0017] In an optional embodiment, the target guidance text for the target difficulty text is determined according to a preset guidance model, including the following steps:

[0018] The target difficulty text is input into the guidance model. Based on the recall accuracy and recall rate thresholds, a first segment text is obtained. A first guidance text is generated based on the first segment text to prompt the intelligent customer service to understand the target difficulty text. The guidance model is trained through multiple guidance text training sets, knowledge base training sets, and interactive text training sets labeled with difficult texts. If the intent understanding text and emotion understanding text of the interactive text prompted by the first guidance text do not meet the preset first difficulty threshold, the top k second segment texts obtained from the input target difficulty text are obtained. The second segment texts are set as the context of the first guidance text and concatenated to obtain the second guidance text as the target guidance text.

[0019] If the intent-understanding text and emotion-understanding text of the interactive text under the second guiding text prompt do not meet the preset second difficulty threshold, then the original question of the target difficulty text is rewritten into multiple consecutive sub-questions based on the first guiding text and the second guiding text through a rewriting strategy. For each consecutive sub-question, optional assumptions are set based on the first guiding text, the second guiding text, and the knowledge base. The third fragment texts recalled by the consecutive sub-questions are sorted according to the optional assumptions to obtain the top k third fragment texts. The context of the second guiding text is then spliced ​​together to obtain the third guiding text as the target guiding text.

[0020] In an optional embodiment, based on the difficulty analysis of the target difficulty text, the target language model corresponding to the target difficulty text is determined from multiple candidate language models, including the following steps:

[0021] The process involves determining the number of texts corresponding to the target difficulty text, the text difficulty set, the text intent set, the text sentiment set, and the guidance text set. Based on the number of texts in the target difficulty text, the mean difference in text quantity between the target difficulty text and the training dataset of the candidate language model is calculated to obtain the text quantity similarity between the target difficulty text and the candidate language model training dataset. Based on the difficulty values ​​in the text difficulty set, the mean ratio of difficulty values ​​between the target difficulty text and the candidate language model training dataset is calculated to obtain the difficulty similarity between the target difficulty text and the candidate language model training dataset. Based on the text intent set and the text sentiment set, the similarity of intent understanding results and the similarity of sentiment understanding results between the target difficulty text and the candidate language model training data are calculated and summed. The weighted average is used to obtain the similarity of understanding results between the target difficulty text and the candidate language model training dataset; the mean of the guiding text similarity between the target difficulty text and the candidate language model training dataset is calculated based on the guiding text set to obtain the guiding similarity between the target difficulty text and the candidate language model training dataset; the product of the text quantity similarity, difficulty similarity, understanding result similarity, and guiding similarity is calculated to obtain the data similarity between the target difficulty text and the candidate language model; the target language model corresponding to the target difficulty text in the candidate language model is determined based on the data similarity, and the language model is trained by multiple difficulty text training datasets, guiding text training datasets, and annotated interactive texts generated corresponding to the training datasets.

[0022] On the other hand, a difficulty-based intelligent customer service guidance system provided to suit one of the purposes of this application includes:

[0023] The text recognition module is used to acquire multiple interactive texts from the target user and determine the intent-understanding text and sentiment-understanding text for each interactive text.

[0024] The difficulty filtering module is used to filter out the target difficulty text for difficulty guidance from multiple interactive texts based on a preset text filtering algorithm.

[0025] The difficulty learning module is used to determine the target guidance text of the target difficulty text based on the preset guidance model, and to determine the target language model corresponding to the target difficulty text from multiple candidate language models based on the difficulty analysis of the target difficulty text.

[0026] The text generation module is used to input the target guidance text and the target difficulty text into the target language model to obtain the target interactive text corresponding to the target difficulty text.

[0027] On another front, a difficulty-based intelligent customer service guidance device provided to suit one of the purposes of this application includes a central processing unit and a memory, wherein the central processing unit is used to invoke and run a computer program stored in the memory to execute the steps of the difficulty-based intelligent customer service guidance method described in this application.

[0028] In another aspect, a computer-readable storage medium is provided to suit one of the purposes of this application, the computer-readable storage medium storing computer-executable instructions for causing a computer to perform the intelligent customer service guidance method based on difficulty learning as disclosed in any of the first aspects of the present invention.

[0029] The technical solution of this application has many advantages, including but not limited to the following aspects:

[0030] This application firstly demonstrates that intent-based and emotion-based text understanding can directly represent the true state of intelligent customer service in understanding user intent and emotions. This facilitates subsequent guidance for intelligent customer service to learn the difficulty of text and provide data support for understanding intent and emotions, thereby improving the accuracy of intelligent customer service responses.

[0031] Secondly, by calculating the matching degree of text difficulty and understanding results, multiple interactive texts are filtered to select the target difficulty text that needs to be guided by difficulty. For other interactive processes, the computational cost of the system can be reduced. Filtering for the target difficulty text can improve the accuracy of intelligent customer service in identifying problems and also improve the accuracy of generating target interactive text based on the target difficulty text. This can reduce the consumption of internal computing resources of intelligent customer service and improve computing efficiency and effectiveness.

[0032] Next, based on the difficulty level of the text, target guidance texts are provided for different levels of difficulty to serve as a basis for difficulty learning. This facilitates the intelligent customer service system to quickly identify user intentions and emotions, supporting the subsequent output of interactive text. This not only improves the interaction speed of the intelligent customer service system but also enhances the accuracy of its responses to user questions.

[0033] Finally, based on text similarity, difficulty similarity, understanding result similarity, and guidance similarity, the best-performing and most suitable language model was selected. This model can generate the target interactive text required for the target difficult text, thereby answering the user's questions during the interaction process. After combining the interactive text analysis results, the problems of intelligent customer service in understanding the difficulty, intent, and emotion of interactive text were identified. This can improve the accuracy of subsequent generation of target interactive text based on the target difficult text by the language model, thereby improving user satisfaction. Attached Figure Description

[0034] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0035] Figure 1 This is a flowchart of one embodiment of the difficulty-based intelligent customer service guidance method of this application;

[0036] Figure 2 This is a schematic diagram of the intelligent customer service guidance system based on difficulty learning used in this application. Detailed Implementation

[0037] The technical solution of this application is applicable to the field of natural language processing technology, and is particularly applicable to the interactive guidance scenario of intelligent customer service. In this context, intelligent customer service, as a very popular and promising application, is gradually being used in sales consultation scenarios. Intelligent customer service relies on natural language understanding technology and aims to provide users with various services in the form of text interaction or voice interaction, such as consultation, guidance, and daily chat. In the future, it will directly change the existing service delivery methods.

[0038] Currently, intelligent customer service is widely used in the customer service industry to serve users' inquiries and pre-sales and after-sales services. However, the complexity of human dialogue makes there are many difficult texts in chat interaction, which will reduce the response speed and accuracy of intelligent customer service. Especially in the recognition of user intent and emotion, intelligent customer service cannot meet all user needs. Human customer service is still needed to replace intelligent customer service in communication with users when intelligent customer service cannot provide effective services, in order to improve user satisfaction with services.

[0039] Therefore, when customer service representatives encounter content that is difficult to understand quickly during interactions with users, accurately identifying the user's intentions and emotions in order to improve customer service efficiency and user satisfaction has become an urgent problem to be solved.

[0040] Furthermore, to address the aforementioned technical issues, this application proposes an intelligent customer service guidance method based on difficulty learning. The overall approach is as follows: Based on the interactive text, determine the intent-understanding text and emotion-understanding text for the intelligent customer service; based on a preset text filtering algorithm, select the target difficulty text for which the intelligent customer service needs difficulty guidance from multiple interactive texts; based on a preset guidance model, determine the target guidance text for the target difficulty text; then, based on the difficulty analysis of the target difficulty text, determine the target language model corresponding to the target difficulty text from multiple candidate language models; input the target guidance text and the target difficulty text into the target language model to obtain the target interactive text corresponding to the target difficulty text.

[0041] The above technical solution will be described in detail below with reference to the accompanying drawings and specific embodiments. The specific embodiments below can be combined with each other, and the same or similar concepts or processes will not be described again in some embodiments.

[0042] See Figure 1 This application discloses an intelligent customer service guidance method based on difficulty learning, which includes the following steps in a typical stress test embodiment:

[0043] Step 1100: Obtain multiple interactive texts from the target user and determine the intent-understanding text and sentiment-understanding text for each interactive text;

[0044] Optionally, interactive text can be collected through speech recognition and text recognition of the communication content between the intelligent customer service and the user, and the communication content can be converted into interactive text that the computer can read and recognize. It is understandable that the communication content between the intelligent customer service and the user will have a large time span and multiple rounds of conversation. Therefore, when acquiring the user's interactive text, it is necessary to consider the context and time relationship to summarize and classify the interactive text. It is easy to know that the interactive text usually consists of the user's question text and the intelligent customer service's answer text. The accuracy of the answer text is an important indicator for evaluating the intelligent customer service in solving user questions.

[0045] In an optional embodiment, multiple interactive texts from the target user are acquired, and the intent-understanding text and sentiment-understanding text for each interactive text are determined, including the following steps:

[0046] For each acquired interactive text, determine the multi-round interaction range of the interactive text, and acquire the question text and answer text within the multi-round interaction range;

[0047] Semantic features, intent features, and sentiment features are extracted for each question text and each answer text. Context-dependent features are extracted for the multi-turn interaction scope of each interactive text.

[0048] It is evident that distinguishing between question and answer texts and considering the scope of interaction in acquiring interactive texts can increase the range of interactive data that can be selected, ensuring that the analysis results obtained from subsequent data processing meet the actual needs of the real interaction process. The identification tasks of question texts mainly involve the difficulty of text understanding, the identification of user intent, and the identification of user emotions. The output tasks of answer texts mainly involve the accuracy of the answer and the correlation between the difficulty of text understanding, the identification of user intent, and user emotions on the accuracy of the answer.

[0049] Next, the intent understanding text in the interaction text is determined based on context-dependent features, semantic features, and intent features. The emotion understanding text in the interaction text is determined based on context-dependent features, semantic features, and emotion features. The intent understanding text is used to characterize the accuracy of the intelligent customer service in understanding the user's question intent, and the emotion understanding text is used to characterize the accuracy of the intelligent customer service in understanding the user's emotion.

[0050] It is evident that understanding text through intent and emotion can directly represent the true extent to which intelligent customer service understands user intent and emotion. This facilitates subsequent guidance for intelligent customer service to learn the difficulty of the text and provide data support for understanding intent and emotion, thereby improving the accuracy of intelligent customer service responses.

[0051] Step 2100: Based on the preset text filtering algorithm, select the target difficulty text for difficulty guidance requirements from multiple interactive texts;

[0052] It is evident that by setting a text filtering algorithm, difficult texts in the interaction process between intelligent customer service and users can be automatically obtained. After filtering out texts with high comprehension difficulty, the intelligent customer service is guided to learn the difficulty level before outputting corresponding answers. This not only improves the accuracy rate but also enhances the intelligent customer service's ability to handle difficult texts and increases the efficiency of answering questions.

[0053] In an optional embodiment, a target difficulty text for difficulty guidance is selected from multiple interactive texts according to a preset text filtering algorithm, including the following steps:

[0054] Based on the difference algorithm and the matching algorithm, the matching degree of the text difficulty and the matching degree of the understanding result are calculated according to the intent understanding text and the sentiment understanding text.

[0055] Based on the matching degree of the text difficulty and the matching degree of the understanding result, the target difficulty text for difficulty guidance is selected from multiple interactive texts.

[0056] Through the above embodiments, the matching degree of text difficulty and understanding results can be calculated. Then, multiple interactive texts can be filtered to select the target difficulty text that needs difficulty guidance. For other interactive processes that do not require difficulty guidance, the computational cost of the system can be reduced. Filtering for target difficulty text can improve the accuracy of intelligent customer service in identifying problems and also improve the accuracy of generating target interactive text based on the target difficulty text. This can reduce the consumption of internal computing resources of intelligent customer service and improve computational efficiency and effectiveness.

[0057] In practical implementation, the method based on the difference degree algorithm and the matching degree algorithm, according to the intent understanding text and the sentiment understanding text, calculates the matching degree of the text difficulty and the matching degree of the understanding result for each interactive text, including the following steps:

[0058] For each interactive text, calculate the average text difference between the intent-understanding text of the interactive text and the intent-recognition text of multiple preset standard difficulty texts, and obtain the difference parameter corresponding to the interactive text.

[0059] According to the preset text matching rules, the number of keywords, length similarity, and grammatical structure similarity that match between the intention-understanding text of the interactive text and the standard difficulty text are calculated. The number of keywords, length similarity, and grammatical structure similarity are then standardized and weighted to obtain the matching coefficient of the difference parameter.

[0060] The product of the difference parameter, the matching coefficient, and the text information weight of the interactive text is calculated to obtain the matching degree of the text difficulty corresponding to the interactive text. The text information weight is proportional to the difficulty of the interactive text.

[0061] For each interactive text, calculate the average similarity between the emotion-understanding text of the interactive text and the emotion-recognition text of multiple preset historical interactive samples to obtain the first similarity parameter of the emotion-understanding text in the interactive text.

[0062] For each interactive text, the mean similarity between two emotion understanding texts at adjacent times within the multi-turn interaction range is calculated to obtain the two similarity parameters of the emotion understanding text in the interactive text.

[0063] The matching degree of the understanding result corresponding to the interactive text is obtained by calculating the product of the first similarity parameter, the second similarity parameter, the contextual association weight, and the sentiment weight.

[0064] As can be seen, the above implementation method can realize the calculation of whether the interactive text is the target difficulty text that needs to be guided by difficulty learning in terms of intent understanding and emotion understanding based on the preset difficulty reference text and historical interactive text. This can reduce the computational loss of the system. The screening of target difficulty text can improve the accuracy of intelligent customer service in identifying problems and also improve the accuracy of subsequent generation of target interactive text based on target difficulty text. This can reduce the consumption of internal computing resources of intelligent customer service and improve computational efficiency and effectiveness.

[0065] Next, based on the matching degree of the text difficulty and the matching degree of the comprehension results, the target difficulty text for the difficulty guidance requirement is selected from multiple interactive texts, including the following steps:

[0066] Calculate the first summation mean of the matching degree of all understanding results, and calculate the second summation mean of the matching degree of all text difficulties. Calculate the covariance matrix of the first and second summation mean based on the accuracy and response speed of user feedback. Determine the preset first and second matching degree thresholds based on the correlation characterization of text difficulty and understanding results using the covariance matrix.

[0067] Based on the condition that the matching degree of the understanding result is greater than the first matching degree threshold, the top K interactive texts sorted by matching degree are selected to obtain candidate interactive texts;

[0068] Candidate interactive texts are sorted from highest to lowest difficulty based on their matching degree. Candidate interactive texts with a matching degree greater than the second matching degree threshold are selected to obtain the target difficulty text.

[0069] Through the above implementation method, multiple interactive texts can be screened a second time based on the matching degree of the understanding result and the matching degree of the text difficulty. This allows for the selection of interactive texts that most likely indicate difficulties in handling user problems by the intelligent customer service, based on the degree of intent recognition, emotion recognition, and understanding of the difficult text. In particular, the influence between text difficulty and understanding result is positively correlated. Using a covariance matrix can better characterize the impact of text difficulty and understanding result on the accuracy of the interactive text output by the intelligent customer service, thereby accurately selecting the target difficulty text that needs guidance. This improves the accuracy of subsequent output of target interactive text based on the target difficulty text in the interactive text, reduces the consumption of internal computing resources of the intelligent customer service, and improves computing efficiency and effectiveness.

[0070] Step 3100: Based on the preset guidance model, determine the target guidance text for the target difficulty text; based on the difficulty analysis of the target difficulty text, determine the target language model corresponding to the target difficulty text from multiple candidate language models.

[0071] It should be noted that the intelligent customer service system is mainly built on FAQs and RAGs to summarize and output answers to common questions using data such as historical user consultation records, product manuals or official documents. However, for more difficult texts, it needs to learn and understand the questions further. Therefore, it is necessary to guide the intelligent customer service system to learn the specific intentions and emotions of users expressed in interactive texts of different difficulty levels.

[0072] In an optional embodiment, the target guidance text for the target difficulty text is determined according to a preset guidance model, including the following steps:

[0073] The target difficulty text is input into the guidance model, and the first segment text is obtained by filtering based on the recall accuracy and recall rate threshold. The first guidance text used to prompt the intelligent customer service to understand the target difficulty text is generated based on the first segment text. The guidance model is obtained by training multiple guidance text training sets, knowledge base training sets and interactive text training sets labeled with difficult text.

[0074] If the intent understanding text and emotion understanding text of the interactive text under the prompt of the first guiding text do not meet the preset first difficulty threshold, then obtain the top k second fragment texts obtained from the input target difficulty text, set the second fragment texts as the context of the first guiding text and concatenate them to obtain the second guiding text as the target guiding text;

[0075] If the intent-understanding text and emotion-understanding text of the interactive text under the second guiding text prompt do not meet the preset second difficulty threshold, then the original question of the target difficulty text is rewritten into multiple consecutive sub-questions based on the first guiding text and the second guiding text through a rewriting strategy. For each consecutive sub-question, optional assumptions are set based on the first guiding text, the second guiding text, and the knowledge base. The third fragment texts recalled by the consecutive sub-questions are sorted according to the optional assumptions to obtain the top k third fragment texts. The context of the second guiding text is then spliced ​​together to obtain the third guiding text as the target guiding text.

[0076] As can be seen, the above embodiments can provide intelligent customer service with target guidance text as a basis for difficulty learning for target difficulty text based on the difficulty level of the interpreted text. This facilitates intelligent customer service to quickly identify user intentions and emotions, providing support for subsequent output of interactive text. This not only improves the interaction speed of intelligent customer service, but also improves the accuracy of intelligent customer service in responding to user questions.

[0077] In a further embodiment, based on the difficulty analysis of the target difficulty text, the target language model corresponding to the target difficulty text is determined from multiple candidate language models, including the following steps:

[0078] Determine the number of texts corresponding to the target difficulty level, the set of text difficulty, the set of text intent, the set of text emotion, and the set of guiding texts;

[0079] The mean difference in the number of texts between the target difficulty text and the training dataset of the candidate language model is calculated based on the number of texts in the target difficulty text to obtain the text quantity similarity between the target difficulty text and the training dataset of the candidate language model.

[0080] The average ratio of the difficulty values ​​of the target difficulty text to the training dataset of the candidate language model is calculated based on the difficulty values ​​of the text difficulty set, thus obtaining the difficulty similarity between the target difficulty text and the training dataset of the candidate language model.

[0081] The weighted average of the similarity between the intent understanding results and the emotion understanding results between the target difficulty text and the training data of the candidate language model is calculated based on the text intent set and the text emotion set. This yields the similarity between the understanding results of the target difficulty text and the training dataset of the candidate language model.

[0082] The mean similarity between the target difficulty text and the training dataset of the candidate language model is calculated based on the set of guiding texts, thus obtaining the guiding similarity between the target difficulty text and the training dataset of the candidate language model.

[0083] The product of the text quantity similarity, difficulty similarity, comprehension result similarity, and guidance similarity is calculated to obtain the data similarity between the target difficulty text and the candidate language model;

[0084] The target language model corresponding to the target difficulty text in the candidate language model is determined based on the data similarity. The language model is trained by multiple difficulty text training datasets, guidance text training datasets, and annotated interactive texts generated corresponding to the training datasets.

[0085] As can be seen, the above embodiments enable the selection of the most effective and best-matching language model based on text quantity similarity, difficulty similarity, understanding result similarity, and guidance similarity. The language model can be an LLM model that can generate the target interactive text required for the target difficult text, thereby answering the user's questions during the interaction. After combining the interactive text analysis results, the problems of intelligent customer service in understanding the difficulty, intent, and emotion of the interactive text are determined, which can improve the accuracy of subsequent generation of target interactive text based on the target difficult text by the language model, thereby improving user satisfaction.

[0086] Step 4100: Input the target guidance text and the target difficulty text into the target language model to obtain the target interactive text corresponding to the target difficulty text.

[0087] In practical implementation, the target guidance text assists the intelligent customer service in identifying the user's true intentions and emotions in the target difficulty text, while improving the accuracy of the intelligent customer service in understanding difficult text. The target language model is obtained by filtering the target difficulty text and the target guidance text, thereby outputting the most accurate target interaction text, improving the accuracy of generating target interaction text based on the target difficulty text. Furthermore, the difficulty learning for different target difficulty texts and the difficulty learning for the corresponding output target interaction text can be iterated in the intelligent customer service system, enabling the intelligent customer service to gradually improve the accuracy of handling difficult user questions. This will not only gradually improve processing efficiency but also gradually reduce computing resource consumption, providing users with a better user experience and increasing user satisfaction.

[0088] The unique technical advantage of this application lies in the fact that, firstly, by using intent-understanding text and emotion-understanding text, it can directly represent the true situation of intelligent customer service in understanding user intent and emotions, so as to guide intelligent customer service to learn the difficulty of text and provide data support for understanding intent and emotions, thereby improving the accuracy of intelligent customer service responses.

[0089] Secondly, by calculating the matching degree of text difficulty and understanding results, multiple interactive texts are filtered to select the target difficulty text that needs to be guided by difficulty. For other interactive processes, the computational cost of the system can be reduced. Filtering for the target difficulty text can improve the accuracy of intelligent customer service in identifying problems and also improve the accuracy of generating target interactive text based on the target difficulty text. This can reduce the consumption of internal computing resources of intelligent customer service and improve computing efficiency and effectiveness.

[0090] Next, based on the difficulty level of the text, target guidance texts are provided for different levels of difficulty to serve as a basis for difficulty learning. This facilitates the intelligent customer service system to quickly identify user intentions and emotions, supporting the subsequent output of interactive text. This not only improves the interaction speed of the intelligent customer service system but also enhances the accuracy of its responses to user questions.

[0091] Finally, based on text similarity, difficulty similarity, understanding result similarity, and guidance similarity, the best-performing and most suitable language model was selected. This model can generate the target interactive text required for the target difficult text, thereby answering the user's questions during the interaction process. After combining the interactive text analysis results, the problems of intelligent customer service in understanding the difficulty, intent, and emotion of interactive text were identified. This can improve the accuracy of subsequent generation of target interactive text based on the target difficult text by the language model, thereby improving user satisfaction.

[0092] Please see Figure 2According to one aspect of this application, an intelligent customer service guidance system based on difficulty learning is provided. The system includes: a text recognition module, used to acquire multiple interactive texts of a target user and determine the intent understanding text and emotion understanding text of each interactive text; a difficulty filtering module, used to filter out the target difficulty text for difficulty guidance needs from multiple interactive texts according to a preset text filtering algorithm; a difficulty learning module, used to determine the target guidance text of the target difficulty text according to a preset guidance model, and determine the target language model corresponding to the target difficulty text from multiple candidate language models based on the difficulty analysis of the target difficulty text; and a text generation module, used to input the target guidance text and the target difficulty text into the target language model to obtain the target interactive text corresponding to the target difficulty text.

[0093] Based on any embodiment of the system in this application, the system of this application further includes: an extraction module, configured to determine the multi-turn interaction range of each acquired interactive text, and acquire question text and answer text within the multi-turn interaction range; extract semantic features, intent features, and emotion features for each question text and each answer text, and extract context-dependent features for the multi-turn interaction range of each interactive text.

[0094] Based on any embodiment of the system in this application, the system of this application further includes: a feature understanding module, configured to determine intent understanding text in the interactive text based on context-dependent features, semantic features, and intent features, and to determine sentiment understanding text in the interactive text based on context-dependent features, semantic features, and sentiment features, wherein the intent understanding text is used to characterize the accuracy of the intelligent customer service in understanding the user's question intent, and the sentiment understanding text is used to characterize the accuracy of the intelligent customer service in understanding the user's sentiment.

[0095] Based on any embodiment of the system in this application, the system of this application further includes: a text filtering module, configured to calculate the matching degree of text difficulty and the matching degree of understanding result corresponding to each interactive text based on the difference degree algorithm and the matching degree algorithm, according to the intent understanding text and the emotion understanding text; and to filter out the target difficulty text for difficulty guidance needs from multiple interactive texts based on the matching degree of text difficulty and the matching degree of understanding result.

[0096] Based on any embodiment of the system in this application, the system further includes: a matching module, configured to, for each interactive text, calculate the average text difference between the intent-understanding text of the interactive text and the intent-recognition text of a preset set of standard difficulty texts, to obtain a difference parameter corresponding to the interactive text; calculate, according to preset text matching rules, the number of matching keywords, length similarity, and syntactic structure similarity between the intent-understanding text of the interactive text and the standard difficulty texts, perform standardized weighted processing on the number of matching keywords, length similarity, and syntactic structure similarity, to obtain a matching coefficient for the difference parameter; and calculate the product of the difference parameter, the matching coefficient, and the text information weight of the interactive text. The matching degree of the text difficulty corresponding to the interactive text is obtained, and the text information weight is proportional to the difficulty of the interactive text. For each interactive text, the average similarity between the emotion understanding text of the interactive text and the emotion recognition text of multiple preset historical interaction samples is calculated to obtain the first similarity parameter of the emotion understanding text in the interactive text. For each interactive text, the average similarity between two emotion understanding texts of the emotion understanding text of the interactive text in adjacent time periods within the range of multiple rounds of interaction is calculated to obtain the second similarity parameter of the emotion understanding text in the interactive text. The product of the first similarity parameter, the second similarity parameter, the context association weight, and the emotion weight is calculated to obtain the matching degree of the understanding result corresponding to the interactive text.

[0097] Based on any embodiment of the system in this application, the system in this application further includes: a guidance model selection module, configured to input the target difficulty text into the guidance model, filter the first segment text based on the recall accuracy and recall rate thresholds, generate a first guidance text as the target difficulty text based on the first segment text to prompt the intelligent customer service to understand the target difficulty text, and the guidance model is obtained by training through multiple guidance text training sets, knowledge base training sets and interactive text training sets labeled with difficult texts;

[0098] If the intent and emotion understanding texts of the interactive text prompted by the first guiding text do not meet the preset first difficulty threshold, then the top k second fragment texts obtained from the input target difficulty text are acquired, and the second fragment texts are set as the context of the first guiding text and concatenated to obtain the second guiding text as the target guiding text. If the intent and emotion understanding texts of the interactive text prompted by the second guiding text do not meet the preset second difficulty threshold, then the original question of the target difficulty text is rewritten into multiple consecutive sub-questions based on the first and second guiding texts using a rewriting strategy. For each consecutive sub-question, optional assumptions are set based on the first guiding text, the second guiding text, and the knowledge base. The third fragment texts recalled by the consecutive sub-questions are sorted according to the optional assumptions to obtain the top k third fragment texts, and the context of the second guiding text is concatenated to obtain the third guiding text as the target guiding text.

[0099] Based on any embodiment of the system in this application, the system further includes: a language model selection module, configured to determine the number of texts corresponding to the target difficulty text, a text difficulty set, a text intent set, a text sentiment set, and a guiding text set; calculate the average difference in text quantity between the target difficulty text and the training dataset of the candidate language model based on the number of texts in the target difficulty text, to obtain the text quantity similarity between the target difficulty text and the training dataset of the candidate language model; calculate the average ratio of the difficulty values ​​between the target difficulty text and the training dataset of the candidate language model based on the difficulty values ​​in the text difficulty set, to obtain the difficulty similarity between the target difficulty text and the training dataset of the candidate language model; and calculate the intent similarity between the target difficulty text and the training data of the candidate language model based on the text intent set and the text sentiment set. The weighted average of the similarity between the solution results and the similarity between the emotion understanding results is used to obtain the similarity between the target difficulty text and the candidate language model training dataset. The mean of the guide text similarity between the target difficulty text and the candidate language model training dataset is calculated based on the guide text set to obtain the guide similarity between the target difficulty text and the candidate language model training dataset. The product of the text quantity similarity, difficulty similarity, understanding result similarity, and guide similarity is calculated to obtain the data similarity between the target difficulty text and the candidate language model. Based on the data similarity, the target language model corresponding to the target difficulty text in the candidate language model is determined. The language model is trained using multiple difficulty text training datasets, guide text training datasets, and annotated interactive texts generated corresponding to the training datasets.

[0100] Another embodiment of this application provides a difficulty-based intelligent customer service guidance device, which includes a processor, a computer-readable storage medium, a memory, and a network interface connected via a system bus. The computer-readable, non-volatile storage medium of the difficulty-based intelligent customer service guidance device stores an operating system, a database, and computer-readable instructions. The database may store information sequences, and when the computer-readable instructions are executed by the processor, the processor can implement a difficulty-based intelligent customer service guidance method.

[0101] The processor of this difficulty-learning-based intelligent customer service guidance device provides computational and control capabilities, supporting the operation of the entire device. The device's memory can store computer-readable instructions, which, when executed by the processor, cause the processor to perform the difficulty-learning-based intelligent customer service guidance method of this application. The network interface of the device is used for communication with a terminal.

[0102] In this embodiment, the processor is used to execute... Figure 2 The system defines the specific functions of each module, and the memory stores the program code and various data required to execute these modules or submodules. The network interface is used to enable data transmission between user terminals or servers.

[0103] The non-volatile readable storage medium in this embodiment stores the program code and data required to execute all modules in the difficulty-based intelligent customer service guidance system of this application. The server can call the program code and data of the server to execute the functions of all modules.

[0104] This application also provides a non-volatile readable storage medium storing computer-readable instructions, which, when executed by one or more processors, cause the one or more processors to perform the steps of the difficulty-based intelligent customer service guidance method of any embodiment of this application.

[0105] This application also provides a computer program product, including a computer program / instructions that, when executed by one or more processors, implement the steps of the method described in any embodiment of this application.

Claims

1. A method for intelligent customer service guidance based on difficulty learning, characterized in that, Includes the following steps: Acquire multiple interactive texts from the target user, and determine the intent-understanding text and sentiment-understanding text for each interactive text; Based on a preset text filtering algorithm, the target difficulty text for difficulty guidance is selected from multiple interactive texts; Based on the preset guidance model, the target guidance text for the target difficulty text is determined. Based on the difficulty analysis of the target difficulty text, the target language model corresponding to the target difficulty text is determined from multiple candidate language models. The target guidance text and the target difficulty text are input into the target language model to obtain the target interactive text corresponding to the target difficulty text; The step of determining the target guidance text based on the preset guidance model includes the following steps: The target difficulty text is input into the guidance model, and the first segment text is obtained by filtering based on the recall accuracy and recall rate threshold. The first guidance text used to prompt the intelligent customer service to understand the target difficulty text is generated based on the first segment text. The guidance model is obtained by training multiple guidance text training sets, knowledge base training sets and interactive text training sets labeled with difficult text. If the intent understanding text and emotion understanding text of the interactive text under the prompt of the first guiding text do not meet the preset first difficulty threshold, then obtain the top k second fragment texts obtained from the input target difficulty text, set the second fragment texts as the context of the first guiding text and concatenate them to obtain the second guiding text as the target guiding text; If the intent-understanding text and emotion-understanding text of the interactive text under the second guiding text prompt do not meet the preset second difficulty threshold, then the original question of the target difficulty text is rewritten into multiple consecutive sub-questions based on the first guiding text and the second guiding text through a rewriting strategy. For each consecutive sub-question, optional assumptions are set based on the first guiding text, the second guiding text, and the knowledge base. The third fragment texts recalled by the consecutive sub-questions are sorted according to the optional assumptions to obtain the top k third fragment texts. The context of the second guiding text is then spliced ​​together to obtain the third guiding text as the target guiding text.

2. The intelligent customer service guidance method based on difficulty learning according to claim 1, characterized in that, Obtain multiple interaction texts from the target user, and determine the intent-based text and sentiment-based text for each interaction text, as follows: For each acquired interactive text, determine the multi-round interaction range of the interactive text, and acquire the question text and answer text within the multi-round interaction range; Semantic features, intent features, and sentiment features are extracted for each question text and each answer text. Context-dependent features are extracted for the multi-turn interaction scope of each interactive text. The intent-understanding text in the interactive text is determined based on context-dependent features, semantic features, and intent features. The emotion-understanding text in the interactive text is determined based on context-dependent features, semantic features, and emotion features. The intent-understanding text is used to characterize the accuracy of the intelligent customer service in understanding the user's question intent, and the emotion-understanding text is used to characterize the accuracy of the intelligent customer service in understanding the user's emotion.

3. The intelligent customer service guidance method based on difficulty learning according to claim 2, characterized in that, Based on a preset text filtering algorithm, the target difficulty text for difficulty guidance is selected from multiple interactive texts, as follows: Based on the difference algorithm and the matching algorithm, the matching degree of the text difficulty and the matching degree of the understanding result are calculated according to the intent understanding text and the sentiment understanding text. Based on the matching degree of the text difficulty and the matching degree of the understanding result, the target difficulty text for difficulty guidance is selected from multiple interactive texts.

4. The intelligent customer service guidance method based on difficulty learning according to claim 3, characterized in that, The method based on the difference algorithm and the matching algorithm calculates the matching degree of text difficulty and the matching degree of understanding result for each interactive text according to intent-based and emotion-based text understanding, including the following steps: For each interactive text, calculate the average text difference between the intent-understanding text of the interactive text and the intent-recognition text of multiple preset standard difficulty texts, and obtain the difference parameter corresponding to the interactive text. According to the preset text matching rules, the number of keywords, length similarity, and grammatical structure similarity that match between the intention-understanding text of the interactive text and the standard difficulty text are calculated. The number of keywords, length similarity, and grammatical structure similarity are then standardized and weighted to obtain the matching coefficient of the difference parameter. The product of the difference parameter, the matching coefficient, and the text information weight of the interactive text is calculated to obtain the matching degree of the text difficulty corresponding to the interactive text. The text information weight is proportional to the difficulty of the interactive text. For each interactive text, calculate the average similarity between the emotion-understanding text of the interactive text and the emotion-recognition text of multiple preset historical interactive samples to obtain the first similarity parameter of the emotion-understanding text in the interactive text. For each interactive text, the mean similarity between two emotion understanding texts at adjacent times within the multi-turn interaction range is calculated to obtain the two similarity parameters of the emotion understanding text in the interactive text. The matching degree of the understanding result corresponding to the interactive text is obtained by calculating the product of the first similarity parameter, the second similarity parameter, the contextual association weight, and the sentiment weight.

5. The intelligent customer service guidance method based on difficulty learning according to claim 4, characterized in that, Based on the matching degree of the text difficulty and the matching degree of the comprehension results, the target difficulty text for difficulty guidance is selected from multiple interactive texts, including the following steps: Calculate the first summation mean of the matching degree of all understanding results, and calculate the second summation mean of the matching degree of all text difficulties. Calculate the covariance matrix of the first and second summation mean based on the accuracy and response speed of user feedback. Determine the preset first and second matching degree thresholds based on the correlation characterization of text difficulty and understanding results using the covariance matrix. Based on the condition that the matching degree of the understanding result is greater than the first matching degree threshold, the top K interactive texts sorted by matching degree are selected to obtain candidate interactive texts; Candidate interactive texts are sorted from highest to lowest difficulty based on their matching degree. Candidate interactive texts with a matching degree greater than the second matching degree threshold are selected to obtain the target difficulty text.

6. The intelligent customer service guidance method based on difficulty learning according to claim 1, characterized in that, Based on the difficulty analysis of the target difficulty text, the target language model corresponding to the target difficulty text is determined from multiple candidate language models, including the following steps: Determine the number of texts corresponding to the target difficulty level, the set of text difficulty, the set of text intent, the set of text emotion, and the set of guiding texts; The mean difference in the number of texts between the target difficulty text and the training dataset of the candidate language model is calculated based on the number of texts in the target difficulty text to obtain the text quantity similarity between the target difficulty text and the training dataset of the candidate language model. The average ratio of the difficulty values ​​of the target difficulty text to the training dataset of the candidate language model is calculated based on the difficulty values ​​of the text difficulty set, thus obtaining the difficulty similarity between the target difficulty text and the training dataset of the candidate language model. The weighted average of the similarity between the intent understanding results and the emotion understanding results between the target difficulty text and the training data of the candidate language model is calculated based on the text intent set and the text emotion set. This yields the similarity between the understanding results of the target difficulty text and the training dataset of the candidate language model. The mean similarity between the target difficulty text and the training dataset of the candidate language model is calculated based on the set of guiding texts, thus obtaining the guiding similarity between the target difficulty text and the training dataset of the candidate language model. The product of the text quantity similarity, difficulty similarity, comprehension result similarity, and guidance similarity is calculated to obtain the data similarity between the target difficulty text and the candidate language model; The target language model corresponding to the target difficulty text in the candidate language model is determined based on the data similarity. The language model is trained by multiple difficulty text training datasets, guidance text training datasets, and annotated interactive texts generated corresponding to the training datasets.

7. An intelligent customer service guidance system based on difficulty learning, characterized in that, The system is used to execute the intelligent customer service guidance method based on difficulty learning as described in any one of claims 1-6, the system comprising: The text recognition module is used to acquire multiple interactive texts from the target user and determine the intent-understanding text and sentiment-understanding text for each interactive text. The difficulty filtering module is used to filter out the target difficulty text for difficulty guidance from multiple interactive texts based on a preset text filtering algorithm. The difficulty learning module is used to determine the target guidance text of the target difficulty text based on the preset guidance model, and to determine the target language model corresponding to the target difficulty text from multiple candidate language models based on the difficulty analysis of the target difficulty text. The text generation module is used to input the target guidance text and the target difficulty text into the target language model to obtain the target interactive text corresponding to the target difficulty text.

8. An intelligent customer service guidance device based on difficulty learning, comprising: At least one processor, and, A memory communicatively connected to the at least one processor; characterized in that the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the intelligent customer service guidance method based on difficulty learning as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores, in the form of computer-readable instructions, a computer program implemented according to any one of claims 1 to 6, which, when invoked by a computer, executes the steps included in the corresponding method.

Images