A marketing strategy updating method, device and computer readable storage medium

Abstract

The present application discloses a method, device and computer-readable storage medium for updating a marketing strategy, the method comprising: obtaining marketing failure dialogue data and marketing success dialogue data; using an intent recognition model to perform intent recognition processing on the marketing failure dialogue data and the marketing success dialogue data respectively, and obtaining a first dialogue semantic label map and a second dialogue semantic label map; based on the first dialogue semantic label map and the second dialogue semantic label map, associating the marketing failure dialogue data and the marketing success dialogue data; analyzing and processing the marketing success dialogue data associated with the marketing failure dialogue data to generate a marketing success strategy; adding the marketing success strategy to the marketing strategy of the marketing failure dialogue data to update the marketing strategy of the marketing failure dialogue data. In the above manner, the present application can improve the success rate of marketing.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a marketing strategy updating method, device and computer-readable storage medium. Background Art

[0002] It is becoming increasingly important to use artificial intelligence technology for machine automated outbound marketing. In the current outbound marketing scenario, the conversion rate of machine automated marketing is much lower than that of manual marketing. The main reasons are: the marketing strategy during the human-machine dialogue, the verbal expression of the process node replies during the human-machine dialogue, the quality of the user's voice transcription during the human-machine dialogue, and the understanding of the semantics of the transcribed text. Among them, the quality of the marketing strategy during the human-machine dialogue mainly depends on the degree of refined design by human experts, and it cannot be directly reused in different businesses and scenarios. It not only requires a huge amount of manpower, but also the maintainability of the entire system is very low. Summary of the invention

[0003] The present application provides a marketing strategy updating method, device and computer-readable storage medium, which can improve the success rate of marketing.

[0004] To solve the above technical problems, the technical solution adopted in the present application is: to provide a method for updating a marketing strategy, the method comprising: obtaining marketing failure conversation data and marketing success conversation data; using an intent recognition model to perform intent recognition processing on the marketing failure conversation data and the marketing success conversation data respectively, to obtain a first conversation semantic label map and a second conversation semantic label map; based on the first conversation semantic label map and the second conversation semantic label map, associating the marketing failure conversation data and the marketing success conversation data; analyzing and processing the marketing success conversation data associated with the marketing failure conversation data to generate a marketing success strategy; adding the marketing success strategy to the marketing strategy of the marketing failure conversation data to update the marketing strategy of the marketing failure conversation data.

[0005] In order to solve the above-mentioned technical problems, another technical solution adopted in the present application is: to provide a marketing strategy updating device, which includes a memory and a processor connected to each other, wherein the memory is used to store a computer program, and when the computer program is executed by the processor, it is used to implement the marketing strategy updating method in the above-mentioned technical solution.

[0006] To solve the above technical problems, another technical solution adopted in the present application is: to provide a computer-readable storage medium, which is used to store a computer program. When the computer program is executed by a processor, it is used to implement the marketing strategy updating method in the above technical solution.

[0007] Through the above scheme, the beneficial effect of the present application is: first, the marketing failure dialogue data and the marketing success dialogue data are obtained; then, the intent recognition model is used to perform intent recognition processing on the marketing failure dialogue data to obtain a first dialogue semantic label map, and the intent recognition model is used to perform intent recognition processing on the marketing success dialogue data to obtain a second dialogue semantic label map; then, the first dialogue semantic label map and the second dialogue semantic label map are used to determine whether the marketing failure dialogue data and the marketing success dialogue data meet the association conditions, and if the association conditions are met, the marketing failure dialogue data and the marketing success dialogue data are associated; then, the marketing success dialogue data associated with the marketing failure dialogue data is analyzed and processed to obtain a marketing success strategy; finally, the marketing success strategy is added to the marketing strategy of the marketing failure dialogue data to update the marketing strategy of the marketing failure dialogue data; by associating the marketing success dialogue data with the marketing failure dialogue data, and extracting the marketing strategy adopted in the marketing success dialogue data, the marketing strategy in the marketing success dialogue data can be used to update the marketing strategy in the marketing failure dialogue data, so that the machine can use the updated marketing strategy for marketing in the next marketing, thereby increasing the chance of marketing success. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. Among them:

[0009] Figure 1 It is a flowchart of an embodiment of a method for updating a marketing strategy provided by the present application;

[0010] Figure 2 It is a schematic diagram of the human-computer dialogue provided by this application;

[0011] Figure 3 It is a schematic diagram of the person-to-person dialogue provided by this application;

[0012] Figure 4 It is a flowchart of another embodiment of the method for updating the marketing strategy provided by the present application;

[0013] Figure 5 is a schematic diagram of the intent recognition model provided by this application;

[0014] FIG6( a ) is another schematic diagram of a human-computer dialogue provided by the present application;

[0015] FIG6( b ) is a first semantic label graph corresponding to FIG6( a );

[0016] FIG6( c ) is a first conversation semantic label diagram corresponding to FIG6( a );

[0017] Figure 7 is a schematic diagram of a similarity calculation model provided by this application;

[0018] Figure 8 It is a structural schematic diagram of an embodiment of a device for updating a marketing strategy provided by the present application;

[0019] Fig. 9 It is a structural schematic diagram of an embodiment of a computer-readable storage medium provided by the present application. DETAILED DESCRIPTION

[0020] The present application is further described in detail below in conjunction with the accompanying drawings and examples. It is particularly noted that the following examples are only used to illustrate the present application, but are not intended to limit the scope of the present application. Similarly, the following examples are only some embodiments of the present application rather than all embodiments, and all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present application.

[0021] Reference to "embodiments" in this application means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0022] It should be noted that the terms "first", "second" and "third" in this application are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features. Thus, the features defined as "first", "second" and "third" may explicitly or implicitly include at least one of the features. In the description of this application, the meaning of "multiple" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but optionally also includes steps or units that are not listed, or optionally also includes other steps or units inherent to these processes, methods, products or devices.

[0023] This application proposes a high-quality marketing strategy mining solution for an outbound marketing scenario. This solution can not only determine the reasons for marketing failure in the current human-machine dialogue process, but also mine high-quality marketing strategies from similar marketing cases from human-to-human dialogues, thereby improving the conversion rate of machine customer service marketing. The technical solution adopted in this application is described in detail below.

[0024] See also Figure 1 , Figure 1 1 is a flow chart of an embodiment of a method for updating a marketing strategy provided by the present application, the method comprising:

[0025] Step 11: Obtain marketing failure conversation data and marketing success conversation data.

[0026] The data involved in this embodiment are divided into two categories: human-to-human dialogue data of successful marketing and human-to-machine dialogue data of failed marketing. The human-to-human dialogue data is data generated by communication between people, and the human-to-machine dialogue data is data generated by communication between people and machines. The dialogue between the customer service robot and the user can be recorded to obtain human-to-machine recording data, and then the human-to-machine recording data can be transcribed through automatic speech recognition technology (Automatic Speech Recognition, ASR) to obtain human-to-machine dialogue text data, and the human-to-machine dialogue text data of marketing failure (hang up) can be filtered out from the human-to-machine dialogue text data according to whether the marketing is successful, to obtain marketing failure dialogue data; similarly, the complete dialogue between the artificial customer service and the user can be recorded to obtain human-to-human recording data, and then the human-to-human recording data can be transcribed by ASR to obtain human-to-human dialogue text data, and according to whether the marketing is successful, the human-to-human dialogue text data of marketing success can be filtered out from the human-to-human dialogue text data to obtain marketing success dialogue data. This part of the data contains the high-quality marketing strategies of the artificial customer service.

[0027] Step 12: Use the intent recognition model to perform intent recognition processing on the marketing failure dialogue data and the marketing success dialogue data respectively to obtain the first dialogue semantic label map and the second dialogue semantic label map.

[0028] After obtaining the marketing failure dialogue data, the machine marketing strategy failure problem judgment can be performed, that is, the marketing failure data in the human-machine dialogue data is analyzed, and the user's intention to refuse to handle the transaction mentioned in the marketing failure dialogue data is identified and classified, and the type of machine marketing strategy failure in the marketing failure dialogue data is extracted. Specifically, the intent recognition model can be used to perform intent recognition processing on the marketing failure dialogue data to obtain dialogue intention information. The intent recognition model is a pre-trained model, which can be an intent recognition model in related technologies, such as: traditional machine learning algorithms (such as: deep belief networks or support vector machines) or based on deep learning algorithms (such as: convolutional neural networks, long short-term memory networks or regional convolutional neural networks), etc. The specific structure and implementation process of the model will not be described in detail here.

[0029] Furthermore, the marketing type can be extracted from the marketing failure conversation data through the intention recognition model, where the marketing type is the type of the product being promoted, and the user's intention in the marketing success conversation data can be extracted in the same way as the marketing failure conversation data.

[0030] Step 13: Based on the first conversation semantic label graph and the second conversation semantic label graph, the marketing failure conversation data and the marketing success conversation data are associated.

[0031] The obtained first conversation semantic label map and the second conversation semantic label map are used to associate the marketing failure conversation data with the marketing success conversation data. Specifically, the similarity between the first conversation semantic label map and the second conversation semantic label map can be calculated, and the similarity can be compared with the set similarity threshold. If the similarity is greater than the similarity threshold, it indicates that the marketing failure conversation data is similar to the marketing success conversation data, and a correlation relationship between the two can be established.

[0032] Step 14: Analyze and process the marketing success conversation data associated with the marketing failure conversation data to generate a marketing success strategy.

[0033] After associating the marketing failure conversation data with the marketing success conversation data, the mining of high-quality conversation strategies in the marketing success conversation data can be performed; specifically, through the association relationship established in step 13, the marketing success strategies (i.e., marketing success strategies) in the marketing success conversation data associated with the marketing failure conversation data are located.

[0034] Step 15: Add the successful marketing strategy to the marketing strategy of the failed marketing conversation data to update the marketing strategy of the failed marketing conversation data.

[0035] After obtaining the successful marketing strategy, the successful marketing strategy is used to update the marketing strategy in the associated failed marketing conversation data (recorded as the failed marketing strategy), for example: directly replacing the failed marketing strategy; or combining the failed marketing strategy with the successful marketing strategy to generate a new marketing strategy, thereby turning the failed marketing strategy into a successful marketing strategy, which can improve the success rate of machine marketing and increase the conversion rate of machine customer service marketing.

[0036] In a specific embodiment, Figure 2-Figure 3 The marketing example of the data package business of xx communication company shown in the figure is: Figure 2 This is an example of human-machine dialogue marketing. In this example, when the robot was selling a data package to the user, the user rejected the robot's sales pitch on the grounds that "I should go to the business hall to handle it myself". At this time, the robot tried to market again with the strategy of "the user had excessive data usage in the past, so it is not cost-effective", but was hung up by the user halfway, resulting in a marketing failure. Figure 3 For example, in the human-machine marketing example, similar to the problems encountered in the human-machine marketing process, when the amateur promoted the data package to the user, the user refused the amateur's promotion on the grounds that "I can go to the business hall to handle it myself". However, at this time, the amateur took the strategy of "the customer is a VIP user and can enjoy the privilege of online handling" and actively asked the user whether he could help him handle the business online, and successfully promoted the data package to the user. It can be seen that high-quality marketing strategies can significantly improve the conversion rate in the machine marketing process.

[0037] This embodiment provides a solution for mining high-quality marketing strategies in outbound marketing scenarios. This solution can not only quickly locate the problems with the current marketing strategies in human-computer dialogues, but also specifically mine improvement solutions that can improve the current marketing failure problems from the person-to-person dialogue data, thereby improving the effectiveness of the human-computer dialogue marketing strategy in the outbound marketing scenario and improving the conversion success rate of human-computer marketing.

[0038] See also Figure 4 , Figure 4 1 is a flow chart of another embodiment of a method for updating a marketing strategy provided by the present application, the method comprising:

[0039] Step 401: Acquire marketing failure conversation data.

[0040] Step 401 is the same as step 11 in the above embodiment and will not be described again.

[0041] The user's intention type for rejecting marketing in a specific marketing task scenario can be manually defined first (L = [l1, l2, ..., ln]). For example, if the marketing service is a 4G to 5G service, the type label of the user's rejection can be "busy", "wifi available", "sufficient traffic", "acceptable speed reduction", "forgot to cancel" or "unclear intention", etc. Then, the intent recognition model is used to identify the intent of the content output by the current user. Specifically, it is found in practical applications that there are many difficult problems in the recognition of user intentions, such as: multiple types of intentions, negative intentions, accounting intentions or consulting intentions, etc., which leads to the mainstream semantic representation method based on bidirectional encoder representations (Bidirectional Encoder Representations from Transformers, BERT) encoding cannot capture the above difficult intentions well; for this, this embodiment proposes an intention recognition scheme based on semantic structure analysis, which mainly constructs the current user's expression content (utterance) structured based on semantic labels, and then uses a graph decoding scheme to construct a semantic label graph; finally, the user's intention is recognized by matching metrics, and the specific implementation scheme will be described below.

[0042] Step 402: Use the label prediction layer to identify and process the content data to obtain a predicted content label.

[0043] The marketing failure dialogue data includes at least one content data, the content data may be text data corresponding to a sentence said by the user during the human-computer dialogue, and the predicted content tag includes a first content tag and a second content tag; specifically, Figure 5 As shown, the intent recognition model includes a label prediction layer (not shown in the figure) and a graph parsing layer. The label prediction layer includes an encoding layer, a labeling layer, and a token extraction layer. The token extraction layer includes a first self-attention layer.

[0044] Furthermore, the content data is input into the encoding layer to obtain content feature data; the content feature data is input into the annotation layer to obtain the position information of the first content tag; then based on the position information, the first content tag is extracted from the content data; the content feature data is input into the first self-attention layer to obtain the second content tag. Specifically, the encoding layer can be a BERT encoder, and the annotation layer can be a Conditional Random Fields (CRF) tagger; the first content tag is an interval-class token tag (i.e., a keyword / word that actually exists in the content data), and the second content tag is a tag-class token tag (i.e., a keyword / word that does not actually exist in the content data, which is a virtual token tag). The extracted token tags (including interval-class token tags and tag-class token tags) can be called concept nodes in the constructed semantic tag graph.

[0045] For example, Figure 5 As shown, the content data is "I want to change the query password of my savings card, can I do it?" Through BERT encoding and CRF-based decoding, the starting and ending positions of the interval-like token labels in the input content data are predicted, and the label-like token labels are predicted through attention interaction. Vtoken is the feature vector of the input label-like token label, V'token is the feature vector of the predicted label-like token label, "B" is the end position of the interval-like token label, and "E" is the end position of the interval-like token label. Therefore, the first content label includes "savings card" and "query password", and the second content label includes "change" and "processing", that is, the predicted content label (W) includes "savings card", "query password", "change" and "processing".

[0046] Step 403: Use the graph parsing layer to process the predicted content labels to construct a first semantic label graph.

[0047] The first semantic label graph is a semantic label graph corresponding to the content data. The content data may be query data. According to the predicted content label identified in step 402, the main intent concept node is obtained. The main intent concept node is a token label in the predicted content label. If there are multiple main intent concept nodes, multiple first semantic label graphs (trees) need to be constructed. Then, the first semantic label graph is initialized as the main intent concept node, that is, the main intent concept node is used as the initial graph state. Then, the utterance-level semantic label graph (that is, the first semantic label graph) is constructed by adopting the following steps:

[0048] (1) Select a node t from the extracted concepts nodes.

[0049] (2) In the constructed partial graph, select a parent node for t.

[0050] When searching for a parent node, the information of some subtrees that have been constructed can be integrated.

[0051] (3) Update the utterance-level semantic label graph and Graph State;

[0052] (4) Repeat steps (1)-(3) until the END node is selected.

[0053] For example, Figure 5 For example, f(·) is the node select module function, g(·) is the head select module function, G t is the current utterance-level semantic label map, x i (i=0,1,3,…) represents the selected concept node, g i It is a graph state. It can be understood that the graph parsing layer is similar to the related art in which the abstract meaning representation (AMR) parsing is realized through iterative reasoning of graphs and sequences, and will not be described in detail here.

[0054] Step 404: Obtain a second semantic label map, and match the first semantic label map with the second semantic label map to determine a third semantic label map.

[0055] The second semantic label map is a semantic label map corresponding to each marketing conversation data in a preset database, where the marketing conversation data is person-to-person conversation data for successful marketing. The third semantic label map is a second semantic label map whose matching degree with the first semantic label map is greater than a preset matching degree, where the preset matching degree is a value set based on experience or application requirements.

[0056] Furthermore, the utterance-level semantic label graph (i.e., the second semantic label graph) corresponding to the natural sentence patterns under each intent in the preset database is constructed offline in the above manner, and the utterance-level semantic label graph corresponding to the user content is encoded using a graph encoder (Graph Transformer), and the encoded sentence representation is saved offline. When obtaining the currently input content data, the semantic label graph (i.e., the first semantic label graph) and the encoded semantic representation of the current content data are also constructed first; by matching the first semantic label graph with the second semantic label graph (for example, calculating whether the similarity between the first semantic label graph and the second semantic label graph is greater than a set value), a third semantic label graph is selected from all the second semantic label graphs, so that the third semantic label graph can be used to obtain the intent corresponding to the current content data.

[0057] Step 405: Determine the first conversation intention information based on the third semantic label graph.

[0058] After obtaining the third semantic tag graph, the intent corresponding to the content data is obtained by matching; specifically, the preset database also includes intent information corresponding to the marketing conversation data, and the intent information corresponding to the third semantic tag graph is used as sub-intent information, and all sub-intent information is merged to obtain the first conversation intent information.

[0059] Step 406: Merge the first semantic label graphs corresponding to all content data to obtain a first conversation semantic label graph.

[0060] The first conversation semantic label graph is a chapter-level semantic label graph. Through the cross-link strategy, the root node of the conversation-level semantic label graph is defined, and the root node is a dummy node. Then, the user-related content in all content data is connected, and the conversation turn number is used as the type of the connected edge to form a conversation-level semantic label graph. For example, taking the conversation record between the user and the robot shown in Figure 6(a) as an example, the conversation record is parsed by AMR to generate four first semantic label graphs shown in Figure 6(b); these four first semantic label graphs are merged to obtain the first conversation semantic label graph shown in Figure 6(c).

[0061] This embodiment obtains the user's utterance-level intention (i.e., the first dialogue intention information) through intent matching, and constructs a dialogue-level semantic label map through a cross-link strategy, and uses it as the type of human-computer dialogue failure; for the intention of long and difficult sentences output by the user, such as: multi-category intention, negative intention, accounting intention or consultation intention, etc., a token-based parsing method is used for intent recognition, which has higher recognition accuracy.

[0062] Step 407: Based on the first semantic label graph and the second semantic label graph, the similarity between the marketing failure dialogue data and the marketing success dialogue data is calculated to obtain data similarity.

[0063] After obtaining the failure problem of the marketing failure conversation data, the person-to-person conversation data with similar conversation content is mined from the person-to-person conversation data of the marketing success. This embodiment adopts the method of conversation label graph matching to mine similar conversation data from the person-to-person conversation data, specifically including:

[0064] A) According to the above-mentioned method of constructing the semantic label graph of human-computer dialogue (i.e., the first dialogue semantic label graph), a semantic label graph structure corresponding to a human-computer dialogue dataset is constructed offline, and the human-computer dialogue dataset includes multiple marketing failure dialogue data.

[0065] B) Offline construction of the semantic label graph structure corresponding to the human-to-human conversation dataset.

[0066] The person-to-person dialogue data set includes multiple marketing success dialogue data. Since the number of dialogue rounds of the marketing success person-to-person dialogue data is generally greater than the number of dialogue rounds of the human-computer dialogue data, when sampling the person-to-person dialogue data, the data in the front area of ​​the person-to-person dialogue data can be selected as the marketing success dialogue data.

[0067] C) Perform sentence feature encoding and semantic label map feature encoding on the marketing failure dialogue data and the marketing success dialogue data, then fuse the two sets of features through attention, and finally determine the similarity score (i.e., data similarity) between the marketing failure dialogue data and the marketing success dialogue data through mixed similarity calculation.

[0068] In a specific embodiment, it is possible to use Figure 7 The similarity calculation model shown is used to calculate the data similarity, and the model includes: a first encoder, a second encoder, a second self-attention layer, a product layer, and a pooling layer, specifically including:

[0069] The marketing failure dialogue data is encoded using a first preset encoding method to obtain first feature data; the first dialogue semantic label map is encoded using a second preset encoding method to obtain second feature data; the marketing success dialogue data is encoded using the first preset encoding method to obtain third feature data; the second dialogue semantic label map is encoded using the second preset encoding method to obtain fourth feature data; based on the first feature data, the second feature data, the third feature data and the fourth feature data, the data similarity is calculated. Specifically, the first preset encoding method is implemented by a first encoder, and the first preset encoding method can be BERT; the second preset encoding method is implemented by a second encoder, and the second preset encoding method can be a graph convolution network (GCN) or a tree-structured long short-term memory network (Tree-LSTM).

[0070] Further, the first feature data and the second feature data are processed by the second self-attention layer to obtain the fifth feature data; the fifth feature data and the first feature data are processed to obtain the sixth feature data; the sixth feature data and the first feature data are processed by the pooling layer to obtain the seventh feature data; the seventh feature data and the second feature data are fused to obtain the first fused feature data; the third feature data and the fourth feature data are processed by the second self-attention layer to obtain the eighth feature data; the eighth feature data and the third feature data are processed to obtain the ninth feature data; the ninth feature data and the third feature data are processed by the pooling layer to obtain the tenth feature data; the tenth feature data and the fourth feature data are fused to obtain the second fused feature data; based on the first fused feature data and the second fused feature data, the data similarity is calculated. Specifically, the fifth feature data and the first feature data are multiplied by the product layer to obtain the sixth feature data; the eighth feature data and the third feature data are multiplied by the product layer to obtain the ninth feature data.

[0071] In one embodiment, the seventh feature data and the second feature data can be weightedly summed to generate first fused feature data; the tenth feature data and the fourth feature data can be weightedly summed to generate second fused feature data; the similarity between the first fused feature data and the second fused feature data is calculated to obtain data similarity.

[0072] In another embodiment, the similarity between the second feature data and the fourth feature data can be calculated to obtain the first similarity; the similarity between the seventh feature data and the tenth feature data can be calculated to obtain the second similarity; the first similarity and the second similarity are weightedly summed to obtain data similarity.

[0073] Step 408: Determine whether the data similarity is greater than a preset similarity.

[0074] The preset similarity may be a similarity comparison threshold set in advance based on experience or application requirements.

[0075] Step 409: If the data similarity is greater than a preset similarity, an association relationship between the marketing failure dialogue data and the marketing success dialogue data is established.

[0076] This embodiment first constructs a token semantic label graph structure of marketing failure dialogue data and marketing success dialogue data; then obtains the similarity score of marketing failure dialogue data and marketing success dialogue data by fusing graph encoding and sentence encoding. When the similarity score is large, the similarity association between marketing failure dialogue data and marketing success dialogue data is achieved.

[0077] Step 410: Analyze and process the marketing success conversation data associated with the marketing failure conversation data to generate a marketing success strategy.

[0078] Step 410 is the same as step 14 of the above embodiment and will not be described again here.

[0079] Step 411: Add the marketing success strategy to the marketing strategy of the marketing failure dialogue data to update the marketing strategy of the marketing failure dialogue data.

[0080] The marketing success strategies are classified according to the type of the first conversation intention information to obtain classification results; the classification results are filtered based on the user selection instructions issued by the developer to obtain filtered marketing success strategies; the filtered success strategies are added to the marketing strategies corresponding to the marketing failure conversation data to obtain updated marketing strategies.

[0081] This embodiment analyzes different types of marketing failure dialogue data, summarizes the successful marketing strategies adopted by sales personnel when they encounter similar marketing problems in human-to-human dialogues, and classifies the successful marketing strategies, which are then manually confirmed; finally, the confirmed successful marketing strategies are added to the human-computer dialogue strategies as high-quality marketing strategies.

[0082] The technical solution provided in this embodiment adopts a token-based parsing method to identify the intent of long and difficult sentences, with higher recognition accuracy and a wider range of recognition types; and adopts an intent matching method to obtain the user's utterance-level intent, and constructs a dialogue-level semantic label graph through a cross-link strategy, which is used as the type of human-computer dialogue failure; by constructing a token semantic label graph structure of human-to-human dialogue data and human-to-computer dialogue data, a graph encoding and sentence encoding fusion method is adopted to obtain the similarity score of human-to-human dialogue data and human-to-computer dialogue data, and realize the similar association between human-to-computer dialogue data and human-to-human dialogue data; by associating similar human-to-human dialogue data with human-to-computer dialogue data, high-quality marketing strategies in human-to-human dialogue data in the scenario of human-to-computer dialogue failure are extracted, which helps to improve the success rate of machine sales.

[0083] See also Figure 8 , Figure 8 It is a structural diagram of an embodiment of a marketing strategy updating device provided in the present application. The marketing strategy updating device 80 includes a memory 81 and a processor 82 connected to each other. The memory 81 is used to store a computer program. When the computer program is executed by the processor 82, it is used to implement the marketing strategy updating method in the above embodiment.

[0084] See also Fig. 9 , Fig. 9 It is a structural diagram of an embodiment of a computer-readable storage medium provided in the present application. The computer-readable storage medium 90 is used to store a computer program 91. When the computer program 91 is executed by a processor, it is used to implement the method for updating the marketing strategy in the above embodiment.

[0085] The computer-readable storage medium 90 may be a server, a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, or other media that can store program codes.

[0086] In the several embodiments provided in this application, it should be understood that the disclosed methods and devices can be implemented in other ways. For example, the device implementation described above is only illustrative, for example, the division of modules or units is only a logical function division, and there may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed.

[0087] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0088] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0089] The above descriptions are merely embodiments of the present application and are not intended to limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A method for updating a marketing strategy, characterized in that: include: Obtaining data on failed marketing conversations and successful marketing conversations; Using an intention recognition model to perform intention recognition processing on the marketing failure dialogue data and the marketing success dialogue data respectively, to obtain a first dialogue semantic label graph and a second dialogue semantic label graph; Associating the marketing failure dialogue data with the marketing success dialogue data based on the first dialogue semantic label graph and the second dialogue semantic label graph; Analyze and process the marketing success dialogue data associated with the marketing failure dialogue data to generate a marketing success strategy; The successful marketing strategy is added to the marketing strategy of the failed marketing conversation data to update the marketing strategy of the failed marketing conversation data.

2. The method for updating marketing strategy according to claim 1, characterized in that: The intention recognition model includes a label prediction layer and a graph parsing layer, the marketing failure dialogue data includes at least one content data, and the step of performing intent recognition processing on the marketing failure dialogue data using the intention recognition model to obtain the first dialogue semantic label graph includes: Using the label prediction layer to identify the content data to obtain a predicted content label; Processing the predicted content labels using the graph parsing layer to construct a first semantic label graph; Obtaining a second semantic label graph, where the second semantic label graph is a semantic label graph corresponding to each piece of marketing conversation data in a preset database; Matching the first semantic label graph with the second semantic label graph to determine a third semantic label graph, where the third semantic label graph is a second semantic label graph whose matching degree with the first semantic label graph is greater than a preset matching degree; Determining conversation intention information based on the third semantic label graph; The first semantic label graphs corresponding to all the content data are merged to obtain the first conversation semantic label graph.

3. The method for updating marketing strategy according to claim 2, characterized in that: The preset database also includes the intention information corresponding to the marketing dialogue data. The step of determining the dialogue intention information based on the third semantic tag graph includes: Using the intent information corresponding to the third semantic label graph as sub-intent information; All the sub-intention information are merged to obtain the dialogue intention information.

4. The method for updating marketing strategy according to claim 2, characterized in that: The predicted content label includes a first content label and a second content label, the label prediction layer includes a coding layer, a labeling layer and a first self-attention layer, and the step of using the label prediction layer to identify and process the content data to obtain the predicted content label includes: Inputting the content data into the coding layer to obtain content feature data; Inputting the content feature data into the annotation layer to obtain the position information of the first content tag; Based on the position information, extracting the first content tag from the content data; The content feature data is input into the first self-attention layer to obtain the second content label.

5. The method for updating marketing strategy according to claim 2, characterized in that: The step of associating the marketing failure dialogue data with the marketing success dialogue data based on the first dialogue semantic label graph and the second dialogue semantic label graph includes: Based on the first semantic label graph and the second semantic label graph, calculating the similarity between the marketing failure dialogue data and the marketing success dialogue data to obtain data similarity; Determining whether the data similarity is greater than a preset similarity; If so, an association relationship between the marketing success dialogue data and the marketing success dialogue data is established.

6. The method for updating marketing strategy according to claim 5, characterized in that: The step of calculating the similarity between the marketing failure dialogue data and the marketing success dialogue data based on the first semantic label graph and the second semantic label graph to obtain data similarity includes: The marketing failure dialogue data is encoded using a first preset encoding method to obtain first feature data; the first dialogue semantic label graph is encoded using a second preset encoding method to obtain second feature data; The marketing success dialogue data is encoded using the first preset encoding method to obtain third feature data; the second dialogue semantic label graph is encoded using the second preset encoding method to obtain fourth feature data; The data similarity is calculated based on the first feature data, the second feature data, the third feature data, and the fourth feature data.

7. The method for updating marketing strategy according to claim 6, characterized in that: The method further comprises: The first feature data and the second feature data are processed by a second self-attention layer to obtain fifth feature data; the fifth feature data and the first feature data are processed to obtain sixth feature data; the sixth feature data and the first feature data are processed by a pooling layer to obtain seventh feature data; the seventh feature data and the second feature data are fused to obtain first fused feature data; The third feature data and the fourth feature data are processed by the second self-attention layer to obtain eighth feature data; the eighth feature data and the third feature data are processed to obtain ninth feature data; the ninth feature data and the third feature data are processed by a pooling layer to obtain tenth feature data; the tenth feature data and the fourth feature data are fused to obtain second fused feature data; The data similarity is calculated based on the first fused feature data and the second fused feature data.

8. The method for updating marketing strategy according to claim 7, characterized in that: The step of calculating the data similarity based on the first fused feature data and the second fused feature data includes: Performing a weighted summation on the seventh feature data and the second feature data to generate the first fused feature data; Performing a weighted summation on the tenth feature data and the fourth feature data to generate the second fused feature data; The similarity between the first fused feature data and the second fused feature data is calculated to obtain the data similarity.

9. The method for updating marketing strategy according to claim 7, characterized in that: The step of calculating the data similarity based on the first feature data, the second feature data, the third feature data and the fourth feature data comprises: Calculating the similarity between the second feature data and the fourth feature data to obtain a first similarity; Calculating the similarity between the seventh feature data and the tenth feature data to obtain a second similarity; A weighted sum is performed on the first similarity and the second similarity to obtain the data similarity.

10. The method for updating marketing strategy according to claim 7, characterized in that: The method further comprises: multiplying the fifth characteristic data by the first characteristic data to obtain the sixth characteristic data; The eighth characteristic data is multiplied by the third characteristic data to obtain the ninth characteristic data.

11. The method for updating marketing strategy according to claim 2, characterized in that: The method further comprises: Classifying the successful marketing strategies according to the types of the conversation intention information to obtain classification results; Based on the user selection instruction, the classification results are screened to obtain a screened successful marketing strategy; The screened successful strategy is added to the marketing strategy corresponding to the marketing failure dialogue data to obtain an updated marketing strategy.

12. A marketing strategy updating device, characterized in that: It comprises a memory and a processor connected to each other, wherein the memory is used to store a computer program, and when the computer program is executed by the processor, it is used to implement the method for updating the marketing strategy described in any one of claims 1-11.

13. A computer-readable storage medium for storing a computer program, characterized in that: When the computer program is executed by a processor, it is used to implement the method for updating the marketing strategy according to any one of claims 1 to 11.

Images