Script recommendation method and device, terminal equipment and medium

Abstract

The application discloses a dialogue recommendation method and device, terminal equipment and a computer readable storage medium. The method comprises the following steps: obtaining current dialogue content, classifying the current dialogue content through a preset reply strategy classification model, obtaining a customer service reply prediction strategy output by the reply strategy classification model; based on the customer service reply prediction strategy, obtaining an initial candidate dialogue from a preset dialogue library, and screening the initial candidate dialogue to obtain a candidate dialogue; based on the dialogue content, scoring the candidate dialogue through a preset dialogue scoring model to obtain a target recommended dialogue matched with the dialogue content, and recommending the dialogue based on the target recommended dialogue. The application can realize flexible and accurate dialogue recommendation, meet user needs, and improve user experience.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, apparatus, terminal device, and computer-readable storage medium for recommending dialogue scripts. Background Technology

[0002] As people place increasingly higher demands on customer service, they often expect customer service representatives to understand user needs promptly and accurately and resolve user problems efficiently.

[0003] The current mainstream method for recommending dialogue is similar to a knowledge base question-and-answer approach, which focuses on returning matching dialogue based on the user's current question. This approach has at least the following problems:

[0004] 1) User responses are usually quite simple, with a lot of omissions and colloquial expressions, which makes it difficult for customer service to infer the user's true intentions by combining user attributes. Traditional word recommendation methods only match based on dialogue text, which makes it difficult to match the user's true needs.

[0005] 2) In service scenarios led and guided by customer service, customer service representatives need to guide users to achieve their goals through response strategies. There are multiple reasonable solutions to the same user problem, but traditional script recommendation methods often treat a fixed response as the standard answer, which cannot meet the diverse needs of users.

[0006] In summary, existing methods for recommending sales scripts cannot provide flexible and accurate recommendations to users, thus failing to meet their diverse needs. Summary of the Invention

[0007] The main objective of this invention is to provide a method, apparatus, terminal device, and computer-readable storage medium for recommending conversation scripts, aiming to achieve flexible and accurate script recommendations, meet user needs, and thereby improve user experience.

[0008] To achieve the above objectives, the present invention provides a method for recommending sales scripts, the method comprising the following steps:

[0009] The current dialogue content is obtained, and the current dialogue content is classified into response strategies using a preset response strategy classification model to obtain the customer service response prediction strategy output by the response strategy classification model.

[0010] Based on the customer service response prediction strategy, initial candidate scripts are obtained from a preset script library, and the initial candidate scripts are filtered to obtain candidate scripts.

[0011] Based on the dialogue content, a preset dialogue scoring model is used to score the candidate dialogues to obtain target recommended dialogues that match the dialogue content, and then dialogue recommendations are made based on the target recommended dialogues.

[0012] Optionally, before the steps of obtaining the current dialogue content, classifying the dialogue content into response strategies using a preset response strategy classification model, and obtaining the customer service response prediction strategy output by the response strategy classification model, the method further includes:

[0013] The response strategy classification model and the dialogue scoring model are constructed respectively.

[0014] Optionally, the step of constructing the response strategy classification model includes:

[0015] Obtain historical dialogue logs, and obtain the dialogue context, user attributes, and dialogue roles corresponding to the dialogue context from the historical dialogue logs;

[0016] The dialogue context, user attributes, and dialogue roles are input into the initial strategy classification model to obtain the strategy label distribution of the dialogue context, and the response strategy classification model is generated based on the strategy label distribution.

[0017] Optionally, the step of inputting the dialogue context, the user attributes, and the dialogue role into an initial strategy classification model to obtain the strategy label distribution of the dialogue context, and generating the response strategy classification model based on the strategy label distribution, includes:

[0018] The dialogue context is input into the pre-trained language representation layer of the initial policy classification model to obtain sentence vectors in the dialogue context.

[0019] The dialogue character and its corresponding relative position are respectively input into the character encoding layer and position encoding layer of the initial strategy classification model to obtain the character encoding corresponding to the dialogue character and the position encoding corresponding to the relative position of the character. The sentence vector is then accumulated with the character encoding and the position encoding.

[0020] The user attributes are input into the user attribute encoding layer of the initial strategy classification model to obtain the corresponding user features;

[0021] The sentence vector obtained by accumulating the role code and the position code is concatenated with the user feature, and the sentence vector after concatenation with the user feature is input into the interaction layer of the initial strategy classification model to obtain the dialogue sequence feature corresponding to the sentence vector;

[0022] The dialogue sequence features are input into the dialogue representation layer of the initial policy classification model to obtain the overall features of the dialogue content;

[0023] Based on the distribution of customer service response strategies for each user question in the historical dialogue logs, the overall features of the dialogue content are classified to determine the distribution of strategy tags corresponding to the dialogue context.

[0024] The response strategy classification model is obtained by training the model based on the strategy label distribution.

[0025] Optionally, the step of constructing the speech scoring model includes:

[0026] Based on the dialogue matching relationship between the user and customer service in the historical dialogue log, the training samples of the dialogue scoring model are obtained for post-training. The training samples include positive samples and negative samples, and the negative samples include easily distinguishable negative samples and difficult-to-distinguish negative samples.

[0027] Post-training is performed on the positive samples, the easily distinguishable negative samples, and the difficult-to-distinguish negative samples to construct an initial speech scoring model through the pre-trained language representation model;

[0028] The initial speech scoring model is fine-tuned based on the positive and negative samples to obtain the speech scoring model.

[0029] Optionally, after the step of obtaining training samples for the dialogue scoring model based on the dialogue matching relationship between the user and customer service in the historical dialogue log, the method includes:

[0030] The dialogue between the user and customer service in the historical dialogue log is spliced ​​together to obtain the dialogue content, and then the dialogue content is spliced ​​together with the customer service reply;

[0031] The concatenated dialogue content after the customer service reply is input into the word embedding layer of the dialogue scoring model to obtain the word vectors of the words in the dialogue content;

[0032] The role encoding and position encoding of the dialogue content are obtained through the role encoding layer and position encoding layer of the dialogue scoring model, respectively, and the word vector is accumulated with the role encoding and the position encoding;

[0033] Through the interaction layer of the speech scoring model, the word vectors accumulated from the role encoding and the position encoding are interacted to obtain word vector sequence features;

[0034] The word vector sequence features are input into the sentence representation layer of the dialogue scoring model to obtain the overall features of the dialogue context and customer response, thereby constructing the dialogue scoring model.

[0035] Optionally, before the steps of constructing the response strategy classification model and the dialogue scoring model respectively, the method further includes:

[0036] Obtain historical dialogue data;

[0037] Vector encoding is performed on the customer service response data in the historical dialogue data to obtain a set of response sentence vectors, and the set of response sentence vectors is clustered to obtain multiple candidate speech clusters;

[0038] Obtain the keywords of the candidate dialogue clusters, and based on the keywords, filter the candidate dialogue clusters to obtain clusters that represent customer service response strategies, and add the candidate dialogues in the clusters to the preset dialogue library;

[0039] The user questions in the historical dialogue corpus are encoded to obtain a set of user question sentence vectors, and the set of user question sentence vectors is clustered.

[0040] Based on the clusters representing customer service response strategies in the preset dialogue script library and the set of clustered user question sentence vectors, the distribution of customer service response strategies under each user question is obtained through the preset matching relationship between user questions and customer service responses. Based on the distribution of customer service response strategies, the distribution of strategy tags corresponding to the dialogue context is determined.

[0041] Optionally, the response scripts containing customer service response prediction strategies in the preset script library can be used as initial candidate scripts;

[0042] Determine the degree of match between the initial candidate dialogue and the context of the current dialogue;

[0043] Based on the matching degree, the initial candidate statements are filtered to obtain candidate statements.

[0044] Optionally, the step of scoring the candidate dialogues based on the dialogue content using a preset dialogue scoring model to obtain the target recommended dialogue includes:

[0045] The candidate dialogue is concatenated with the dialogue context in the current dialogue content, and the candidate dialogue with the concatenated dialogue context is input into the dialogue scoring model for scoring, so as to obtain the score corresponding to each group of dialogue context-candidate dialogue.

[0046] Based on the scores, target recommended scripts are determined from the candidate scripts.

[0047] To achieve the above objectives, the present invention also provides a script recommendation device, the script recommendation device comprising:

[0048] The strategy acquisition module is used to acquire the current dialogue content and classify the current dialogue content into response strategies through a preset response strategy classification model to obtain the customer service response prediction strategy output by the response strategy classification model.

[0049] The script filtering module is used to obtain initial candidate scripts from a preset script library based on the customer service response prediction strategy, and to filter the initial candidate scripts to obtain candidate scripts.

[0050] The dialogue scoring module is used to score the candidate dialogues based on the dialogue content and through a preset dialogue scoring model to obtain target recommended dialogues that match the dialogue content, so as to recommend dialogues based on the target recommended dialogues.

[0051] To achieve the above objectives, the present invention also provides a terminal device, the terminal device including a memory, a processor, and a script recommendation program stored in the memory and executable on the processor, wherein the script recommendation program, when executed by the processor, implements the steps of the script recommendation method as described above.

[0052] Furthermore, to achieve the above objectives, the present invention also proposes a computer-readable storage medium storing a script recommendation program, which, when executed by a processor, implements the steps of the script recommendation method described above.

[0053] To achieve the above objectives, the present invention also provides a computer program product, the computer program product comprising a computer program, which, when executed by a processor, implements the steps of the speech recommendation method described above.

[0054] This invention provides a method, apparatus, terminal device, computer-readable storage medium, and computer program product for recommending customer service scripts. The method involves acquiring the current dialogue content and classifying it into response strategies using a preset response strategy classification model to obtain a customer service response prediction strategy output by the model. Based on this prediction strategy, initial candidate scripts are retrieved from a preset script library and filtered to obtain more candidate scripts. Finally, based on the dialogue content, a preset script scoring model is used to score the candidate scripts, resulting in a target recommended script that matches the dialogue content. Script recommendations are then performed based on this target recommended script.

[0055] Compared to existing script recommendation methods, this invention uses a response strategy classification model to classify the current dialogue content into predictive customer service responses. This predictive response strategy then retrieves initial candidate scripts from a script database and filters them to select more candidate scripts. Based on this, a script scoring model scores the candidate scripts, ultimately yielding the target recommended script. Therefore, this invention, through the response strategy classification model and script scoring model, obtains a target recommended script that matches the current dialogue content. Using this target recommended script to reply to the user achieves accurate and flexible script recommendations, better aligning with real-world script recommendation scenarios, further meeting diverse user needs, and enhancing the user experience. Attached Figure Description

[0056] Figure 1 This is a schematic diagram of the hardware operating environment involved in the embodiments of the present invention;

[0057] Figure 2 This is a flowchart illustrating an embodiment of the speech recommendation method of the present invention;

[0058] Figure 3 This is a schematic diagram of the functional modules of an embodiment of the speech recommendation device of the present invention.

[0059] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0060] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0061] like Figure 1 As shown, Figure 1 This is a schematic diagram of the device structure of the hardware operating environment involved in the embodiments of the present invention.

[0062] The terminal device in this embodiment of the invention can be a smartphone, tablet computer, computer, server, or other network device. The terminal device in this embodiment can be used to achieve flexible and accurate recommendation of dialogue.

[0063] like Figure 1As shown, the terminal device may include: a processor 1001, such as a CPU; a network interface 1004; a user interface 1003; a memory 1005; and a communication bus 1002. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen and an input unit such as a keyboard. Optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be high-speed RAM or stable non-volatile memory, such as a disk drive. Optionally, the memory 1005 may also be a storage device independent of the aforementioned processor 1001.

[0064] Those skilled in the art will understand that Figure 1 The device structure shown does not constitute a limitation on the device recommended for the dialogue, and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0065] like Figure 1 As shown, the memory 1005, as a computer storage medium, may include an operation module, a network communication module, a user interface module, and a script recommendation program. The operation module is a program that manages and controls the hardware and software resources of the device, supporting the operation of the script recommendation program and other software or programs. Figure 1 In the device shown, the user interface 1003 is mainly used for data communication with the client; the network interface 1004 is mainly used for establishing a communication connection with the server; and the processor 1001 can be used to call the script recommendation program stored in the memory 1005 and perform the following operations:

[0066] The current dialogue content is obtained, and the current dialogue content is classified into response strategies using a preset response strategy classification model to obtain the customer service response prediction strategy output by the response strategy classification model.

[0067] Based on the customer service response prediction strategy, initial candidate scripts are obtained from a preset script library, and the initial candidate scripts are filtered to obtain candidate scripts.

[0068] Based on the dialogue content, a preset dialogue scoring model is used to score the candidate dialogues to obtain target recommended dialogues that match the dialogue content, and then dialogue recommendations are made based on the target recommended dialogues.

[0069] Further, before the step of acquiring the current dialogue content, classifying the dialogue content into response strategies using a preset response strategy classification model, and obtaining the customer service response prediction strategy output by the response strategy classification model, the processor 1001 can call the script recommendation program stored in the memory 1005 and perform the following operations:

[0070] The response strategy classification model and the dialogue scoring model are constructed respectively.

[0071] Furthermore, the processor 1001 can be used to call the script recommendation program stored in the memory 1005 and perform the following operations:

[0072] Obtain historical dialogue logs, and obtain the dialogue context, user attributes, and dialogue roles corresponding to the dialogue context from the historical dialogue logs;

[0073] The dialogue context, user attributes, and dialogue roles are input into the initial strategy classification model to obtain the strategy label distribution of the dialogue context, and the response strategy classification model is generated based on the strategy label distribution.

[0074] Furthermore, the processor 1001 can be used to call the script recommendation program stored in the memory 1005 and perform the following operations:

[0075] The dialogue context is input into the pre-trained language representation layer of the initial policy classification model to obtain sentence vectors in the dialogue context.

[0076] The dialogue character and its corresponding relative position are respectively input into the character encoding layer and position encoding layer of the initial strategy classification model to obtain the character encoding corresponding to the dialogue character and the position encoding corresponding to the relative position of the character. The sentence vector is then accumulated with the character encoding and the position encoding.

[0077] The user attributes are input into the user attribute encoding layer of the initial strategy classification model to obtain the corresponding user features;

[0078] The sentence vector obtained by accumulating the role code and the position code is concatenated with the user feature, and the sentence vector after concatenation with the user feature is input into the interaction layer of the initial strategy classification model to obtain the dialogue sequence feature corresponding to the sentence vector;

[0079] The dialogue sequence features are input into the dialogue representation layer of the initial policy classification model to obtain the overall features of the dialogue content;

[0080] Based on the distribution of customer service response strategies for each user question in the historical dialogue logs, the overall features of the dialogue content are classified to determine the distribution of strategy tags corresponding to the dialogue context.

[0081] The response strategy classification model is obtained by training the model based on the strategy label distribution.

[0082] Furthermore, the processor 1001 can be used to call the script recommendation program stored in the memory 1005 and perform the following operations:

[0083] Based on the dialogue matching relationship between the user and customer service in the historical dialogue log, the training samples of the dialogue scoring model are obtained for post-training. The training samples include positive samples and negative samples, and the negative samples include easily distinguishable negative samples and difficult-to-distinguish negative samples.

[0084] Post-training is performed on the positive samples, the easily distinguishable negative samples, and the difficult-to-distinguish negative samples to construct an initial speech scoring model through the pre-trained language representation model;

[0085] The initial speech scoring model is fine-tuned based on the positive and negative samples to obtain the speech scoring model.

[0086] Furthermore, after the step of obtaining training samples for the dialogue scoring model based on the dialogue matching relationship between the user and customer service in the historical dialogue log, the processor 1001 can be used to call the dialogue recommendation program stored in the memory 1005 and perform the following operations:

[0087] The dialogue between the user and customer service in the historical dialogue log is spliced ​​together to obtain the dialogue content, and then the dialogue content is spliced ​​together with the customer service reply;

[0088] The concatenated dialogue content after the customer service reply is input into the word embedding layer of the dialogue scoring model to obtain the word vectors of the words in the dialogue content;

[0089] The role encoding and position encoding of the dialogue content are obtained through the role encoding layer and position encoding layer of the dialogue scoring model, respectively, and the word vector is accumulated with the role encoding and the position encoding;

[0090] Through the interaction layer of the speech scoring model, the word vectors accumulated from the role encoding and the position encoding are interacted to obtain word vector sequence features;

[0091] The word vector sequence features are input into the sentence representation layer of the dialogue scoring model to obtain the overall features of the dialogue context and customer response, thereby constructing the dialogue scoring model.

[0092] Furthermore, before the steps of constructing the response strategy classification model and the dialogue scoring model respectively, the processor 1001 can call the dialogue recommendation program stored in the memory 1005 and perform the following operations:

[0093] Obtain historical dialogue data;

[0094] Vector encoding is performed on the customer service response data in the historical dialogue data to obtain a set of response sentence vectors, and the set of response sentence vectors is clustered to obtain multiple candidate speech clusters;

[0095] Obtain the keywords of the candidate dialogue clusters, and based on the keywords, filter the candidate dialogue clusters to obtain clusters that represent customer service response strategies, and add the candidate dialogues in the clusters to the preset dialogue library;

[0096] The user questions in the historical dialogue corpus are encoded to obtain a set of user question sentence vectors, and the set of user question sentence vectors is clustered.

[0097] Based on the clusters representing customer service response strategies in the preset dialogue script library and the set of clustered user question sentence vectors, the distribution of customer service response strategies under each user question is obtained through the preset matching relationship between user questions and customer service responses. Based on the distribution of customer service response strategies, the distribution of strategy tags corresponding to the dialogue context is determined.

[0098] Furthermore, the processor 1001 can be used to call the script recommendation program stored in the memory 1005 and perform the following operations:

[0099] The response scripts containing customer service response prediction strategies in the preset script library are used as initial candidate scripts.

[0100] Determine the degree of match between the initial candidate dialogue and the context of the current dialogue;

[0101] Based on the matching degree, the initial candidate statements are filtered to obtain candidate statements.

[0102] Furthermore, the processor 1001 can be used to call the script recommendation program stored in the memory 1005 and perform the following operations:

[0103] The candidate dialogue is concatenated with the dialogue context in the current dialogue content, and the candidate dialogue with the concatenated dialogue context is input into the dialogue scoring model for scoring, so as to obtain the score corresponding to each group of dialogue context-candidate dialogue.

[0104] Based on the scores, target recommended scripts are determined from the candidate scripts.

[0105] Considering the existing methods of recommending sales scripts, at least the following problems exist:

[0106] 1) Due to the large number of omissions and colloquial expressions in the customer service-user dialogue process, it is difficult to capture the user's true intentions through the dialogue text;

[0107] 2) There is usually not one solution that meets the user's needs in the conversation, and customer service replies are often not enough to list them all;

[0108] 3) Data annotation in dialogue recommendation tasks is very cumbersome.

[0109] To address the aforementioned problems, this invention proposes a dialogue recommendation method. Compared to traditional knowledge-based simple matching dialogue recommendation methods, this invention is divided into an offline mining stage, a model training stage, and a dialogue prediction stage.

[0110] In the offline mining phase, by using historical logs of user-customer service conversations, clustering algorithms are used to merge similar customer service replies into the same cluster, mining standard customer service reply scripts and corresponding reply strategies, and establishing a standard script library; similar user questions in the context of the conversation are merged into the same cluster, and the distribution of reply strategy labels under similar user questions is obtained based on the correspondence between user questions and customer service replies.

[0111] During the model training phase, historical dialogue logs between users and customer service representatives are used to train both a strategy prediction model and a speech scoring model based on imitation learning. The strategy prediction model, based on Label Distribution Learning (LDL), constructs labels using strategy distributions from similar samples and learns the strategy distribution by modeling the dialogue context, dialogue roles, and user labels using a hierarchical transformer model. The speech scoring model employs a two-stage training approach of post-training and fine-tuning, enhancing the model's discriminative scoring ability by constructing difficult-to-distinguish negative samples.

[0112] The prediction phase includes three modules: strategy prediction, strategy-script recall, and script ranking. First, based on the existing dialogue between the user and customer service representative in the current conversation, the data is input into the strategy prediction model for strategy classification. The output strategy is selected based on the strategy's threshold or ranking. In the recall phase, candidate scripts are recalled through the matching relationship between strategies and scripts. Finally, a script scoring model scores the current dialogue content and candidate scripts, completing the ranking and selection of candidate scripts.

[0113] Reference Figure 2 , Figure 2 This is a flowchart illustrating the first embodiment of the speech recommendation method of the present invention.

[0114] This invention provides an embodiment of a script recommendation method. It should be noted that although the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than that shown here.

[0115] Specifically, the script recommendation method in this embodiment includes the following steps:

[0116] Step S10: Obtain the current dialogue content, and classify the current dialogue content into response strategies using a preset response strategy classification model to obtain the customer service response prediction strategy output by the response strategy classification model.

[0117] It should be noted that in this embodiment, a response strategy classification model and a dialogue scoring model are pre-built. Through these models, accurate, efficient and flexible responses to the questions raised by each user can be achieved.

[0118] The terminal device first obtains the current dialogue content, which includes the current dialogue context, dialogue roles, and user attributes. In this embodiment, the dialogue context includes several dialogues between the user and customer service, the dialogue roles include the user and customer service, and the user attributes include the user's age, education level, gender, whether they have outstanding debts, and whether they are a VIP customer.

[0119] Then, the terminal device will classify the current dialogue content into response strategies using a pre-established response strategy classification model, and obtain the customer service response prediction strategy output by the response strategy classification model.

[0120] Specifically, for example, the terminal device inputs the current dialogue context, corresponding role, and user attributes into the response strategy classification model to classify response strategies, thereby obtaining the predicted strategy distribution corresponding to the current dialogue content.<p1, p2, p3, ……> Alternatively, select the top-ranked strategies (e.g., the top 3) in the probability distribution of the prediction strategy distribution as the customer service response prediction strategy, or select the customer service response prediction strategy by using a threshold division method.

[0121] Step S20: Based on the customer service response prediction strategy, obtain initial candidate scripts from the preset script library, and filter the initial candidate scripts to obtain candidate scripts.

[0122] After obtaining the customer service response prediction strategy output by the response strategy classification model, the terminal device will obtain initial candidate scripts from the preset script library based on the customer service response prediction strategy, and further filter out candidate scripts from the initial candidate scripts.

[0123] It is worth noting that, in this embodiment, the criteria for selecting candidate dialogues from the initial candidate dialogues can be specifically as follows: by solving the matching degree between the initial candidate dialogues and the preceding dialogue, the initial candidate dialogues corresponding to the lower matching degree are removed, and finally the candidate dialogues are obtained.

[0124] Step S30: Based on the dialogue content, the candidate dialogues are scored using a preset dialogue scoring model to obtain the target recommended dialogue.

[0125] As described above, based on the pre-built dialogue scoring model, after obtaining candidate dialogues, the terminal device will score the candidate dialogues based on the dialogue context and dialogue roles, and then obtain the target recommended dialogues.

[0126] Specifically, for example, after scoring the candidate dialogues, the terminal device can select the top three dialogues (e.g., the top 3) as the target recommended dialogues.

[0127] In this embodiment, the terminal device first acquires the current dialogue content, which includes the current dialogue context, dialogue roles, and user attributes. Then, the terminal device uses a pre-established response strategy classification model to classify the current dialogue content into response strategies, obtaining a customer service response prediction strategy output by the model. After acquiring the customer service response prediction strategy, the terminal device uses this strategy to obtain initial candidate scripts from a preset script library, and further filters out candidate scripts from these initial candidate scripts. After acquiring the candidate scripts, the terminal device uses a preset script scoring model to score the candidate scripts based on the dialogue content including the dialogue context, dialogue roles, and user attributes, thereby obtaining the target recommended script.

[0128] Compared to existing script recommendation methods, this invention uses a response strategy classification model to classify the current dialogue content into predictive customer service responses. This predictive response strategy then retrieves initial candidate scripts from a script database and filters them to select more candidate scripts. Based on this, a script scoring model scores the candidate scripts, ultimately yielding the target recommended script. Therefore, this invention, through the response strategy classification model and script scoring model, obtains a target recommended script that matches the current dialogue content. Using this target recommended script to reply to the user achieves accurate and flexible script recommendations, better aligning with real-world script recommendation scenarios, further meeting diverse user needs, and enhancing the user experience.

[0129] Furthermore, based on the first embodiment of the speech recommendation of the present invention, a second embodiment of the speech recommendation of the present invention is proposed.

[0130] In this embodiment, before step S10, "obtaining the current dialogue content and classifying the current dialogue content into response strategies using a preset response strategy classification model to obtain the customer service response prediction strategy output by the response strategy classification model," the following may be included:

[0131] Step S40: Construct the response strategy classification model and the speech scoring model respectively.

[0132] When a terminal device uses the response strategy classification model and the dialogue scoring model to recommend dialogue based on the current conversation content, it needs to pre-build the response strategy classification model and the dialogue scoring model. In this way, by utilizing the constructed response strategy classification model and the dialogue scoring model respectively, flexible and accurate dialogue recommendations can be achieved for each user, assisting customer service in serving users, meeting diverse user needs, resolving various user problems, improving user experience, and further enhancing user stickiness.

[0133] Furthermore, in step S40 above, "constructing the response strategy classification model" may include:

[0134] Step S401: Obtain historical dialogue logs, and obtain the dialogue context, user attributes, and dialogue roles corresponding to the dialogue context from the historical dialogue logs.

[0135] Step S402: Input the dialogue context, the user attributes, and the dialogue role into the initial strategy classification model to obtain the strategy label distribution of the dialogue context, and generate the response strategy classification model based on the strategy label distribution.

[0136] When building a response strategy classification model, the terminal device first needs to obtain historical dialogue logs, and then obtain the dialogue context, user attributes, and dialogue roles from those logs. The historical dialogue logs can be understood as all the content shared between the customer service representative and the user during a single call.

[0137] In one embodiment, after obtaining the historical dialogue logs, they need to be cleaned and preprocessed beforehand. Specifically, for example, the terminal device filters dialogues generated by the top 10% of customer service representatives based on their overall rating, and performs text preprocessing on the corresponding dialogue text and customer service replies, such as removing stop words, standardizing capitalization, and normalizing numerical amounts / times. In this way, the dialogue text, user attributes, and dialogue roles can be further obtained from the cleaned and preprocessed historical dialogue logs.

[0138] Then, the terminal device inputs the obtained dialogue context, user attributes, and dialogue roles into the pre-trained language representation model. In this embodiment, the initial policy classification model can be a hierarchical transformer. Its underlying layer uses pre-trained BERT initialization and untrained sentence representation layer and user encoding layer, etc., to obtain the policy label distribution corresponding to the dialogue context output by the model. Then, based on the policy label distribution, a response policy classification model is generated.

[0139] Further, in step S402 above, "inputting the dialogue context, the user attributes, and the dialogue role into the initial strategy classification model to obtain the strategy label distribution of the dialogue context, and generating the response strategy classification model based on the strategy label distribution" can include:

[0140] Step S4021: Input the dialogue context into the pre-trained language representation layer in the initial policy classification model to obtain the sentence vector in the dialogue context;

[0141] Step S4022: Input the dialogue role and the corresponding relative position of the role into the role encoding layer and the position encoding layer of the initial strategy classification model respectively to obtain the role encoding corresponding to the dialogue role and the position encoding corresponding to the relative position of the role, and then accumulate the sentence vector with the role encoding and the position encoding.

[0142] Step S4023: Input the user attributes into the user attribute encoding layer of the initial strategy classification model to obtain the corresponding user features;

[0143] Step S4024: Concatenate the sentence vector obtained by accumulating the role code and the position code with the user feature, and input the sentence vector after concatenation with the user feature into the interaction layer of the initial strategy classification model to obtain the dialogue sequence feature corresponding to the sentence vector;

[0144] Step S4025: Input the dialogue sequence features into the dialogue representation layer of the initial policy classification model to obtain the overall features of the dialogue content;

[0145] Step S4026: Based on the distribution of customer service response strategies for each user question in the historical dialogue log, classify the overall features of the dialogue content to determine the distribution of strategy tags corresponding to the dialogue context.

[0146] Step S4027: Train the response strategy classification model based on the strategy label distribution to obtain the response strategy classification model.

[0147] It is worth noting that in this embodiment, the terminal device inputs the dialogue context <,,…,>, user attributes <,,…,>, and dialogue roles <,,…,> into the aforementioned initial strategy classification model. The final strategy label distribution of the dialogue context is obtained by taking the maximum of the one-hot encoding strategy label <0, 1, 0, …> corresponding to the customer service reply 'a' and the strategy label distribution <0.3, 0.7, 0.1, …> corresponding to the dialogue context.<p1, p2, p3, ……> The target is used as the target for training to generate a policy classification model M1.

[0148] Specifically, for example:

[0149] 1) Input the dialogue context <,,…,> into the pre-trained language representation layer (BERT) to obtain the sentence vector representation <,,…,> of each sentence in the dialogue content;

[0150] 2) Input the dialogue characters <,,…,> and their corresponding relative positions into the character encoding layer and position encoding layer of the initial policy classification model. Obtain the corresponding character encoding and position encoding through the character encoding layer and position encoding layer, respectively. Then, add the above sentence vector representation with the corresponding character encoding and position encoding.

[0151] 3) Input the user attributes <,,…,> into the user attribute encoding layer of the initial policy classification model to obtain the corresponding user features;

[0152] 4) Concatenate the above user features with the sentence vectors of accumulated role and position encodings, and input the concatenated user feature sentence vectors into the interaction layer of the initial policy classification model. The dialogue sequence features corresponding to each sentence vector are obtained through self-attention.

[0153] 5) Input the above dialogue sequence features into the dialogue representation layer of the initial policy classification model to obtain the overall features of the dialogue content;

[0154] 6) Based on the distribution of customer service response strategies for each user question in the historical dialogue logs, classify the representation of the above dialogue content to obtain the probability that the dialogue text belongs to each customer service response strategy, i.e., the strategy label distribution.<p1, p2,p3, ……> ;

[0155] 7) Regarding the above strategy label distribution<p1, p2, p3, ……> Train the model to generate a multi-label response strategy classification model M1.

[0156] Furthermore, in step S40 above, "constructing the speech scoring model" may include:

[0157] Step S403: Based on the dialogue matching relationship between the user and customer service in the historical dialogue log, obtain the training samples of the dialogue scoring model, and perform post-training on the training samples. The training samples include positive samples and negative samples, and the negative samples include easily distinguishable negative samples and difficult-to-distinguish negative samples.

[0158] Step S404: Post-training is performed on the positive samples, the easily distinguishable negative samples, and the difficult-to-distinguish negative samples to construct an initial speech scoring model;

[0159] Step S405: Based on the positive samples and the negative samples, fine-tune the initial speech scoring model to obtain the speech scoring model.

[0160] When building a dialogue scoring model, the terminal device first needs to construct training samples for the model based on the matching relationship between user questions and customer service responses in historical dialogue logs. Specifically, for each dialogue context <,,…,>, positive samples, easily identifiable negative samples, and difficult-to-identify negative samples are constructed respectively. The ratio of these three types of negative samples can be 2:1:1. The positive samples are the original customer service responses. Customer service responses can be randomly sampled from the entire corpus as easily identifiable negative samples, and customer service responses can be randomly sampled from the same dialogue as difficult-to-identify negative samples. The label values ​​range from {0,1,2} to represent positive samples, difficult-to-identify negative samples, and easily identifiable negative samples, respectively.

[0161] Based on this, the terminal device will perform post-training on the aforementioned training samples to construct an initial dialogue scoring model M2'. Specifically, several words in the original text (i.e., the original historical dialogue log) will be randomly replaced with the special character [MASK], original words, or random words. Then, the original text will be predicted through context, and initialized using the parameters of the pre-trained BERT model to generate the initial dialogue scoring model M2'. In addition, the above post-training also includes the following objectives: a three-class classification task to determine whether the matching relationship between the dialogue context and the customer service response is a positive sample, an easily classifiable negative sample, or a difficult-to-classify negative sample.

[0162] Then, the terminal device will fine-tune the initial dialogue scoring model M2' to obtain the dialogue scoring model M2. Specifically, firstly, data labeled as positive samples and easily distinguishable negative samples from the historical dialogue log are selected, and the initial dialogue scoring model M2' is fine-tuned. The training objective is binary classification, and the output is a value between 0 and 1, thus generating the dialogue scoring model M2.

[0163] In this embodiment, the model construction process of the initial dialogue scoring model M2' and the dialogue scoring model M2 is consistent with that of the strategy classification model M1. Specifically, for example:

[0164] 1) Concatenate the user output and customer service reply in the historical dialogue log to obtain the dialogue content, where different roles are separated by delimiters. <eos>The dialogue was then separated and combined with the customer service response using a separator. <sep>Mark;

[0165] 2) Input the concatenated dialogue content into the word embedding layer of the speech scoring model to obtain the word vector of each word;

[0166] 3) Obtain the role code and position code of the dialogue text through the role encoding layer and position encoding layer of the speech scoring model, respectively, and then add the above word vector representations to the corresponding role code and position code;

[0167] 4) Input the above word vectors into the interaction layer of the speech scoring model, and obtain the word vector sequence features of each word vector through self-attention;

[0168] 5) Input the above word vector sequence features into the sentence representation layer of the dialogue scoring model to obtain the overall features of the dialogue context and customer service response.

[0169] Furthermore, before step S40 above, "constructing the response strategy classification model and the dialogue scoring model respectively", the following may also be included:

[0170] Step S50: Obtain historical dialogue data;

[0171] Step S60: Vector encoding is performed on the customer service response data in the historical dialogue data to obtain a set of response sentence vectors, and the set of response sentence vectors is clustered to obtain multiple candidate speech clusters;

[0172] Step S70: Obtain the keywords of the candidate dialogue clusters, and based on the keywords, filter the candidate dialogue clusters to obtain clusters that represent customer service response strategies, and add the candidate dialogues in the clusters to the preset dialogue library.

[0173] Step S80: Encode the user questions in the historical dialogue corpus to obtain a set of user question sentence vectors, and cluster the set of user question sentence vectors;

[0174] Step S90: Based on the clusters representing customer service response strategies in the preset dialogue script library and the set of clustered user question sentence vectors, the distribution of customer service response strategies under each user question is obtained through the preset matching relationship between user questions and customer service responses, so as to determine the distribution of strategy tags corresponding to the dialogue context based on the distribution of customer service response strategies.

[0175] Before constructing the response strategy classification model and the script scoring model, the terminal device also needs to perform offline data mining in advance. That is, by using the historical logs of user and customer service dialogues, clustering algorithms are used to merge similar customer service replies into the same cluster, standard scripts and corresponding strategies are mined, and a standard script library is established; similar user questions in the dialogue are merged into the same cluster, and the distribution of strategy labels under similar questions is obtained according to the correspondence between user questions and customer service replies.

[0176] Specifically, for example:

[0177] 1) Obtain historical dialogue data, and then use the SimCSE model pre-trained based on domain corpus to perform vector encoding on the customer service response data in the historical dialogue data to obtain a set of response sentence vectors. Then, use algorithms such as Kmeans and DBSCAN to cluster the set of response sentence vectors to obtain K candidate speech clusters.

[0178] 2) Use keyword extraction algorithms (such as TextRank) to obtain keywords (such as annual fee, interest, etc.) for each candidate dialogue cluster. Based on these keywords, select clusters that represent customer service response strategies from the candidate dialogues. Then, based on expert experience, select candidate dialogues from the clusters and add them to the preset dialogue library.

[0179] 3) Using the SimCSE model, user questions in the historical dialogue corpus are vectorized to obtain a set of user question sentence vectors, and the set of user question sentence vectors is clustered using algorithms such as Kmeans and DBSCAN;

[0180] 4) Based on the clusters representing customer service response strategies in the aforementioned pre-defined dialogue script library and the clustered set of user question sentence vectors, the strategy distribution of the corresponding customer service response for each user question is obtained through the matching relationship between user questions and customer service responses. Based on this strategy distribution, the strategy tag distribution corresponding to the preceding dialogue text is determined.

[0181] Further, in step S20 above, "based on the customer service response prediction strategy, obtaining initial candidate scripts from a preset script library, and filtering the initial candidate scripts to obtain candidate scripts" may include:

[0182] Step S201: Use the reply scripts containing customer service reply prediction strategies in the preset script library as initial candidate scripts;

[0183] Step S202: Determine the matching degree between the initial candidate dialogue and the context of the current dialogue content;

[0184] Step S203: Based on the matching degree, the initial candidate dialogue is filtered to obtain candidate dialogue.

[0185] The terminal device first selects response scripts containing customer service response prediction strategies from the aforementioned preset script library as initial candidate scripts. Then, it determines the matching degree corresponding to each initial candidate script and further selects candidate scripts from the aforementioned initial candidate scripts based on the matching degree.

[0186] In one embodiment, the determination of the above similarity can be specifically as follows: input the above reply statement (i.e. the initial candidate statement) into the pre-trained language representation layer in the strategy classification model M1 to obtain the sentence vector of the above reply statement, and then calculate the similarity (cosine similarity) between the sentence vector and the overall features of the dialogue in the M1 model.

[0187] Then, the statements with a similarity lower than the preset similarity threshold are removed from the initial candidate statements to obtain candidate statements. In this embodiment, the value of the preset similarity threshold is not specifically limited, and can be flexibly set according to the actual dialogue recommendation scenario.

[0188] Furthermore, in step S30 above, "based on the dialogue content, scoring the candidate dialogues using a preset dialogue scoring model to obtain the target recommended dialogue" may include:

[0189] Step S301: The candidate dialogue is concatenated with the dialogue context in the dialogue content, and the candidate dialogue with the concatenated dialogue context is output to the dialogue scoring model for scoring, so as to obtain the scores corresponding to multiple sets of dialogue context-candidate dialogue.

[0190] Step S302: Based on the score, determine the target recommended script from the candidate scripts.

[0191] After acquiring candidate dialogues, the terminal device concatenates each candidate dialogue with the preceding dialogue in the current conversation. Then, it inputs each candidate dialogue with the concatenated preceding dialogue into the dialogue scoring model M2 for scoring, obtaining the score corresponding to each group of preceding dialogue and candidate dialogue. This allows for the selection of several top-ranked candidate dialogues as target recommended dialogues. For example, the top 3 dialogues can be selected as target recommended dialogues to assist customer service in responding to users.

[0192] In this invention, a response strategy classification model and a dialogue scoring model were established separately. The dialogue context, user roles, and dialogue roles were trained together, improving the accuracy and effectiveness of the dialogue recommendation model. Furthermore, the labels were augmented by predicting the distribution of strategy labels corresponding to the dialogue context of the current dialogue content, enhancing data utilization efficiency and model training performance on small sample data, while also improving the accuracy of strategy selection. Simultaneously, a dialogue scoring model was constructed using post-training and fine-tuning, significantly improving the model's scoring variability and accuracy, thus better facilitating the recommendation of appropriate dialogue.

[0193] Furthermore, embodiments of the present invention also propose a script recommendation device, referring to... Figure 3 The script recommendation device includes:

[0194] The strategy acquisition module is used to acquire the current dialogue content and classify the current dialogue content into response strategies through a preset response strategy classification model to obtain the customer service response prediction strategy output by the response strategy classification model.

[0195] The script filtering module is used to obtain initial candidate scripts from a preset script library based on the customer service response prediction strategy, and to filter the initial candidate scripts to obtain candidate scripts.

[0196] The dialogue scoring module is used to score the candidate dialogues based on the dialogue content and through a preset dialogue scoring model to obtain target recommended dialogues that match the dialogue content, so as to recommend dialogues based on the target recommended dialogues.

[0197] The extended content of the specific implementation of the speech recommendation system of the present invention is basically the same as the various embodiments of the speech recommendation method described above, and will not be repeated here.

[0198] Furthermore, embodiments of the present invention also propose a computer-readable storage medium storing a script recommendation program, wherein when the script recommendation program is executed by a processor, it implements the steps of the script recommendation method described below.

[0199] The embodiments of the speech recommendation device and computer-readable storage medium of the present invention can be referred to the embodiments of the speech recommendation method of the present invention, and will not be repeated here.

[0200] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0201] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0202] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (such as a smartphone or tablet computer) to execute the methods described in the various embodiments of the present invention.

[0203] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.< / sep> < / eos>

Claims

1. A method for recommending sales scripts, characterized in that, The recommended scripts include: Construct a response strategy classification model and a dialogue scoring model respectively; The current dialogue content is obtained, and the current dialogue content is classified into response strategies using a preset response strategy classification model to obtain the customer service response prediction strategy output by the response strategy classification model. Based on the customer service response prediction strategy, initial candidate scripts are obtained from a preset script library, and the initial candidate scripts are filtered to obtain candidate scripts. Based on the dialogue content, a preset dialogue scoring model is used to score the candidate dialogues to obtain target recommended dialogues that match the dialogue content, and dialogue recommendations are made based on the target recommended dialogues. The step of constructing the response strategy classification model includes: Obtain historical dialogue logs, and obtain the dialogue context, user attributes, and dialogue roles corresponding to the dialogue context from the historical dialogue logs; The initial strategy classification model is input into the dialogue context, the user attributes, and the dialogue role to obtain the strategy label distribution of the dialogue context, and the response strategy classification model is generated based on the strategy label distribution; this step includes: The dialogue context is input into the pre-trained language representation layer of the initial policy classification model to obtain sentence vectors in the dialogue context. The dialogue character and its corresponding relative position are respectively input into the character encoding layer and position encoding layer of the initial strategy classification model to obtain the character encoding corresponding to the dialogue character and the position encoding corresponding to the relative position of the character. The sentence vector is then accumulated with the character encoding and the position encoding. The user attributes are input into the user attribute encoding layer of the initial strategy classification model to obtain the corresponding user features; The sentence vector, which is the sum of the role code and the position code, is concatenated with the user feature, and the sentence vector after concatenation with the user feature is input into the interaction layer of the initial strategy classification model to obtain the dialogue sequence feature corresponding to the sentence vector; The dialogue sequence features are input into the dialogue representation layer of the initial policy classification model to obtain the overall features of the dialogue content; Based on the distribution of customer service response strategies for each user question in the historical dialogue logs, the overall features of the dialogue content are classified to determine the distribution of strategy tags corresponding to the dialogue context. The response strategy classification model is obtained by training on the strategy label distribution.

2. The script recommendation method as described in claim 1, characterized in that, The steps for constructing the speech scoring model include: Based on the dialogue matching relationship between the user and customer service in the historical dialogue log, the training samples of the dialogue scoring model are obtained for post-training. The training samples include positive samples and negative samples, and the negative samples include easily distinguishable negative samples and difficult-to-distinguish negative samples. Post-training is performed on the positive samples, the easily distinguishable negative samples, and the difficult-to-distinguish negative samples to construct an initial speech scoring model through a pre-trained language representation model; The initial speech scoring model is fine-tuned based on the positive and negative samples to obtain the speech scoring model.

3. The script recommendation method as described in claim 2, characterized in that, After the step of obtaining training samples for the dialogue scoring model based on the dialogue matching relationship between the user and customer service in the historical dialogue logs, the method further includes: The dialogue between the user and customer service in the historical dialogue log is spliced ​​together to obtain the dialogue content, and then the dialogue content is spliced ​​together with the customer service reply; The concatenated dialogue content after the customer service reply is input into the word embedding layer of the dialogue scoring model to obtain the word vectors of the words in the dialogue content; The role encoding and position encoding of the dialogue content are obtained through the role encoding layer and position encoding layer of the dialogue scoring model, respectively, and the word vector is accumulated with the role encoding and the position encoding; Through the interaction layer of the speech scoring model, the word vectors accumulated from the role encoding and the position encoding are interacted to obtain word vector sequence features; The word vector sequence features are input into the sentence representation layer of the dialogue scoring model to obtain the overall features of the dialogue context and customer response, thereby constructing the dialogue scoring model.

4. The script recommendation method as described in claim 1, characterized in that, Before the steps of constructing the response strategy classification model and the dialogue scoring model respectively, the method further includes: Obtain historical dialogue data; Vector encoding is performed on the customer service response data in the historical dialogue data to obtain a set of response sentence vectors, and the set of response sentence vectors is clustered to obtain multiple candidate speech clusters; Obtain the keywords of the candidate dialogue clusters, and based on the keywords, filter the candidate dialogue clusters to obtain clusters that represent customer service response strategies, and add the candidate dialogues in the clusters to the preset dialogue library; The user questions in the historical dialogue corpus are encoded to obtain a set of user question sentence vectors, and the set of user question sentence vectors is clustered. Based on the clusters representing customer service response strategies in the preset dialogue script library and the set of clustered user question sentence vectors, the distribution of customer service response strategies under each user question is obtained through the preset matching relationship between user questions and customer service responses. Based on the distribution of customer service response strategies, the distribution of strategy tags corresponding to the dialogue context is determined.

5. The script recommendation method as described in claim 1, characterized in that, The step of obtaining initial candidate scripts from a preset script library based on the customer service response prediction strategy, and filtering the initial candidate scripts to obtain candidate scripts, includes: The response scripts containing customer service response prediction strategies in the preset script library are used as initial candidate scripts. Determine the degree of match between the initial candidate dialogue and the context of the current dialogue; Based on the matching degree, the initial candidate statements are filtered to obtain candidate statements.

6. The script recommendation method as described in claim 5, characterized in that, The step of scoring the candidate dialogues based on the dialogue content using a preset dialogue scoring model to obtain the target recommended dialogue includes: The candidate dialogue is concatenated with the dialogue context in the current dialogue content, and the candidate dialogue with the concatenated dialogue context is input into the dialogue scoring model for scoring, so as to obtain the score corresponding to each group of dialogue context-candidate dialogue. Based on the scores, target recommended scripts are determined from the candidate scripts.

7. A script recommendation device, characterized in that, The script recommendation device includes: The module is used to construct a response strategy classification model and a dialogue scoring model, respectively. The module is also used to acquire historical dialogue logs and obtain the dialogue context, user attributes, and dialogue roles corresponding to the dialogue context from the historical dialogue logs. The dialogue context, user attributes, and dialogue roles are input into an initial strategy classification model to obtain the strategy label distribution of the dialogue context, and the response strategy classification model is generated based on the strategy label distribution. The module is also used to input the dialogue context into the pre-trained language representation layer of the initial strategy classification model to obtain sentence vectors in the dialogue context; and to input the dialogue roles and their corresponding relative positions into the role encoding layer and position encoding layer of the initial strategy classification model, respectively, to obtain the role encoding corresponding to the dialogue role and the relative position pair of the roles. The system first encodes the corresponding position and accumulates the sentence vector with the role code and the position code. Then, it inputs the user attributes into the user attribute encoding layer of the initial strategy classification model to obtain the corresponding user features. Next, it concatenates the accumulated sentence vector with the user features and inputs the concatenated sentence vector into the interaction layer of the initial strategy classification model to obtain the dialogue sequence features corresponding to the sentence vector. Finally, it inputs the dialogue sequence features into the dialogue representation layer of the initial strategy classification model to obtain the overall dialogue content features. Based on the distribution of customer service response strategies for each user question in the historical dialogue logs, it classifies the overall dialogue content features to determine the strategy label distribution corresponding to the preceding dialogue text. Finally, it trains the response strategy classification model based on the strategy label distribution. The strategy acquisition module is used to acquire the current dialogue content and classify the current dialogue content into response strategies through a preset response strategy classification model to obtain the customer service response prediction strategy output by the response strategy classification model. The script filtering module is used to obtain initial candidate scripts from a preset script library based on the customer service response prediction strategy, and to filter the initial candidate scripts to obtain candidate scripts. The dialogue scoring module is used to score the candidate dialogues based on the dialogue content and through a preset dialogue scoring model to obtain target recommended dialogues that match the dialogue content, so as to recommend dialogues based on the target recommended dialogues.

8. A terminal device, characterized in that, The terminal device includes a memory, a processor, and a basic script recommendation program stored in the memory and executable on the processor. When the script recommendation program is executed by the processor, it implements the steps of the script recommendation method 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 a script recommendation program, which, when executed by a processor, implements the steps of the script recommendation method as described in any one of claims 1 to 6.

Images