User marketing transformation prediction method and device, storage medium, equipment and product

Abstract

The invention relates to the technical field of product marketing, and particularly provides a user marketing conversion prediction method and device, a storage medium, equipment and a product, and the method can comprise the steps: obtaining an intention representation vector of each conversation in multi-pass conversations of a user; obtaining an intention interaction vector of each dialogue based on the intention representation vector of the multi-pass dialogue; calculating a user feature vector and the intention interaction vector of each dialogue to obtain a user intention vector; a marketing score is determined according to the user attribute, the telephone dialing statistical characteristics and the user intention vector, and the higher the marketing score is, the better the marketing conversion effect of the user is. According to some embodiments of the invention, accurate prediction of the marketing conversion rate can be realized, and the marketing quality is improved.

Description

Technical Field

[0001] The present application relates to the field of product marketing technology, and specifically, to a method, apparatus, storage medium, device and product for predicting user marketing conversion. Background Art

[0002] With the continuous development of artificial intelligence technology, AI robots are widely used in product marketing scenarios in various fields. The conversation content between AI robots and users contains important information. By extracting and analyzing the conversation content, we can understand the user's true intentions, and then combine the corresponding marketing strategies to improve the marketing conversion effect. At present, when extracting user intentions from historical conversation content, the statistical high-frequency intention tags are mostly used as analysis indicators. Obviously, this extraction method is prone to loss of semantic information, which will lead to a one-sided understanding of the user's true intentions, and thus fail to guarantee the effect of user marketing.

[0003] Therefore, how to provide a technical solution for understanding user intentions in conversations with high accuracy and thereby improving the accuracy of user marketing conversion predictions has become a technical problem that needs to be solved urgently. Summary of the invention

[0004] The purpose of some embodiments of the present application is to provide a method, device, storage medium, equipment and product for predicting user marketing conversion. Through the technical solutions of the embodiments of the present application, accurate prediction of user marketing conversion rate can be achieved, thereby improving user marketing conversion effect and marketing quality.

[0005] In a first aspect, some embodiments of the present application provide a method for predicting user marketing conversions, including: obtaining an intent representation vector for each of multiple user conversations; obtaining an intent interaction vector for each conversation based on the intent representation vectors of the multiple conversations; calculating a user feature vector and the intent interaction vector for each conversation to obtain a user intent vector; determining a marketing score based on user attributes, telephone dialing statistical characteristics, and the user intent vector, wherein the higher the marketing score, the better the marketing conversion effect of the user.

[0006] Some embodiments of the present application analyze each conversation in multiple conversations of the user to obtain the intention interaction vector of each conversation, and then calculate it with the user feature vector to obtain the user intention vector; finally, the marketing score is determined by combining the user attributes, the statistical characteristics of the phone calls and the user intention vector. Some embodiments of the present application can achieve accurate prediction of the user marketing conversion rate and improve the user marketing conversion effect and marketing quality by performing comprehensive analysis on multiple conversations.

[0007] In some embodiments, obtaining the intention representation vector of each conversation among multiple conversations of the user includes: segmenting multiple rounds of conversations in each conversation through a sliding window to obtain multiple conversation segments; obtaining semantic features of each conversation segment among the multiple conversation segments; and constructing the intention representation vector of each conversation based on the semantic features of each conversation segment.

[0008] Some embodiments of the present application segment each conversation through a sliding window to obtain semantic features of each conversation segment, and then construct an intent representation vector through the semantic features, so as to accurately obtain the user intent of each conversation and provide a prediction basis for subsequent marketing conversion prediction.

[0009] In some embodiments, the intention representation vectors based on multiple conversations are used to obtain the intention interaction vector of each conversation, including: inputting the intention representation vector of each conversation into a self-attention model, and outputting the intention interaction vector of each conversation.

[0010] Some embodiments of the present application can obtain the intention interaction vector between each conversation corresponding to the intention representation vector of each conversation through the self-attention model, thereby realizing accurate analysis of the correlation between conversations.

[0011] In some embodiments, the user feature vector and the intention interaction vector of each conversation are calculated to obtain the user intention vector, including: taking the feature vector corresponding to the latest user portrait information as the user feature vector; taking the inner product result of the user feature vector and the intention interaction vector of each conversation as the weight value of each conversation; and performing weighted summation of the intention interaction vector and the weight value of each conversation to obtain the user intention vector.

[0012] Some embodiments of the present application determine the user intention vector by taking the feature vector of the latest user portrait information as the user feature vector and combining it with the intention interaction vector. This can capture the user's recent intentions, obtain user intentions with higher accuracy, and provide data support for the accurate prediction of subsequent marketing conversion rates.

[0013] In some embodiments, determining the marketing score based on user attributes, phone call statistical features and the user intention vector includes: concatenating the user attributes, the phone call statistical features and the user intention vector to obtain a concatenated feature vector; inputting the concatenated feature vector into a deep neural network model to output a fused feature vector; and inputting the fused feature vector into a target activation function to obtain the marketing score.

[0014] Some embodiments of the present application achieve effective prediction of marketing conversion rate with high accuracy by concatenating user attributes, phone call statistical features and user intention vectors, fusing them using a neural network to obtain a fused feature vector, and finally inputting the fused feature vector into a target activation function to obtain a marketing score.

[0015] In some embodiments, the user attributes include: user age, user gender and user region information; the telephone dialing statistical characteristics are obtained by collecting statistics on the number of times the user's telephone dialing is connected, the duration of the connection and other information.

[0016] In a second aspect, some embodiments of the present application provide a device for predicting user marketing conversions, including: an intent acquisition module for acquiring an intent representation vector for each of multiple user conversations; an interaction information acquisition module for acquiring an intent interaction vector for each conversation based on the intent representation vectors for multiple conversations; a global intent acquisition module for calculating a user feature vector and the intent interaction vector for each conversation to obtain a user intent vector; and a marketing prediction module for determining a marketing score based on user attributes, telephone dialing statistical characteristics, and the user intent vector, wherein the higher the marketing score, the better the marketing conversion effect of the user.

[0017] In a third aspect, some embodiments of the present application provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, can implement the method described in any embodiment of the first aspect.

[0018] In a fourth aspect, some embodiments of the present application provide an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor can implement a method as described in any embodiment of the first aspect when executing the program.

[0019] In a fifth aspect, some embodiments of the present application provide a computer program product, wherein the computer program product comprises a computer program, wherein the computer program, when executed by a processor, can implement the method described in any embodiment of the first aspect. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions of some embodiments of the present application, the drawings required for use in some embodiments of the present application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0021] Figure 1A system diagram of user marketing conversion prediction provided for some embodiments of the present application;

[0022] Figure 2 One of the method flow charts for user marketing conversion prediction provided in some embodiments of the present application;

[0023] Figure 3 A schematic diagram of the structure of user marketing conversion prediction provided for some embodiments of the present application;

[0024] Figure 4 A schematic diagram of global intent activation provided for some embodiments of the present application;

[0025] Figure 5 A second flow chart of a method for predicting user marketing conversion provided in some embodiments of the present application;

[0026] Figure 6 A block diagram of the device composition for user marketing conversion prediction provided in some embodiments of the present application;

[0027] Figure 7 A schematic diagram of an electronic device is provided for some embodiments of the present application. DETAILED DESCRIPTION

[0028] The technical solutions in some embodiments of the present application will be described below in conjunction with the drawings in some embodiments of the present application.

[0029] It should be noted that similar reference numerals and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. At the same time, in the description of this application, the terms "first", "second", etc. are only used to distinguish the description and cannot be understood as indicating or implying relative importance.

[0030] In the related technology, the common marketing scenarios based on session sessions include telephone marketing scenarios and enterprise micro-marketing scenarios. When modeling conversion goals in the current marketing scenario, intent tags are extracted based on historical conversation content, and recent high-frequency intent tag hits and conversions are counted. However, when modeling, "only high-frequency intent tags are detected to see if they appear" is used as a binary feature. This method obviously suffers from the loss of semantic information extraction. At the same time, it also loses the semantic sequence information of the conversation context, making it even more difficult to capture the evolution of the user's intention associations in multiple conversations. In the absence of understanding the user's true intentions, the marketing conversion effect is poor.

[0031] In view of this, some embodiments of the present application provide a method for predicting user marketing conversion, which extracts an intent representation vector from each of the user's multiple conversations, and then constructs an intent interaction vector between the multiple conversations; then, the user's latest user feature vector and intent interaction vector are combined to mine the recent user intent vector, and finally, the user attributes and phone call statistical features are combined to predict the marketing score. Some embodiments of the present application analyze the degree of correlation between multiple conversations and the user's latest features to mine the user's real intent in the near future, and determine the marketing score based on this, so as to achieve accurate prediction of the marketing score and improve the user's marketing conversion effect and marketing quality.

[0032] The following is combined with Figure 1 The overall composition structure of the system for predicting user marketing conversion provided by some embodiments of the present application is exemplified.

[0033] like Figure 1 As shown, some embodiments of the present application provide a system for predicting user marketing conversions, which may include: an AI robot 100 and a server 200. The AI ​​robot is used to make calls with the user, and the call records are stored in the server 200. The server 200 can confirm the user's intention for multiple conversations, and then predict the user's marketing score based on other user characteristics. The higher the marketing score, the better the conversion effect, so that accurate marketing for users can be achieved, and marketing quality and user experience can be improved.

[0034] The following is combined with Figure 2 The implementation process of user marketing conversion prediction performed by the server 200 provided in some embodiments of the present application is exemplified.

[0035] Please see attached Figure 2 , Figure 2 A flowchart of a method for predicting user marketing conversion is provided for some embodiments of the present application. The method for predicting user marketing conversion may include:

[0036] S210, obtaining an intention representation vector of each conversation in multiple conversations of the user.

[0037] For example, in some embodiments of the present application, Figure 3 The schematic diagram of the structure of the internal user marketing conversion prediction of the server 200 shown in FIG. 1 is a schematic diagram of the structure of the internal user marketing conversion prediction of the server 200. The intention representation extraction module in the server 200 can extract a single conversation (as a specific example of each conversation) to obtain an intention representation vector. It should be noted that in the actual application scenario, the server 200 may contain the following: Figure 3The multiple intent representation extraction modules shown extract intent representation vectors for single-call conversations respectively. Alternatively, one intent representation extraction module may extract intent representation vectors for single-call conversations respectively. This embodiment of the present application does not make any specific limitation here.

[0038] In some embodiments of the present application, S210 may include: dividing multiple rounds of conversation in each conversation through a sliding window to obtain multiple conversation segments; obtaining semantic features of each conversation segment in the multiple conversation segments; and constructing the intention representation vector of each conversation based on the semantic features of each conversation segment.

[0039] For example, in some embodiments of the present application, the intent representation extraction module uses a sliding window to segment multiple rounds of conversations in a single-pass conversation, and uses LLM (Large Language Model) technology to extract semantic representations of the conversation text from the multiple segmented conversation fragments (as a specific example of semantic features); then uses the LSTM (Long Short-Term Memory) structure to model the intent sequence information corresponding to the single-pass conversation (as a specific example of an intent representation vector).

[0040] Specifically, based on a small amount of scene intent annotation data, the open source LLM is fine-tuned to the scene requirements, and the model is deployed in the server 200 to obtain the LLM intent representation extraction service in the intent representation extraction module. For a single-channel human-computer dialogue (i.e., a single-channel dialogue), assuming that it contains N rounds of dialogue (generally N will not exceed 20, as a specific example of a multi-round dialogue), it is divided according to the sliding window width K, and there is a sequence of dialogue segments with a length of N / K (N / K rounded down) (as a specific example of multiple dialogue segments), which can be expressed as part1, part2,..., partN / K. Among them, N and K are adjustable parameters, and K is a divisor of N as much as possible. N / K rounded down means that if the number of dialogue rounds cut to the end by the sliding window is less than K, the remaining dialogue rounds will be merged into the previous window content.

[0041] After that, the embedding representation (that is, semantic features) of each dialogue segment sequence in the dialogue segment sequence part1, part2, ..., partN / K is obtained based on the above-deployed LLM intent representation extraction service, which is recorded as: part_embedding1, part_embedding2, ..., part_embeddingN / K. After obtaining the embedding representation of each dialogue segment sequence, the intent sequence information of the single-channel dialogue is modeled using the LSTM structure to obtain the intent representation vector of the single-channel dialogue, and the dimension is recorded as 1*k, where k is an adjustable parameter. For example, for a certain business scenario, the number of user historical dialogues in multiple time windows is counted, such as the past month, the past three months, the past six months, the past year, etc. In order to make more than 60% of users have 10 historical dialogues, the past six months is selected as the appropriate time window from the current time. Based on the intent representation extraction module, the intent representation embedding vector (that is, the intent representation vector) of 10 dialogues (as a specific example of multi-channel dialogues) can be obtained, that is, 10 vectors of 1*k dimensions.

[0042] S220, based on the intention representation vectors of multiple conversations, obtain the intention interaction vector of each conversation.

[0043] For example, in some embodiments of the present application, Figure 3 The intention representation interaction module in the server 200 can associate the information between multiple conversations, and obtain the intention interaction information of a single conversation that has learned the association between multiple conversations based on the intention representation embedding vector (as a specific example of the intention interaction vector).

[0044] In some embodiments of the present application, S220 may include: inputting the intention representation vector of each conversation into a self-attention model, and outputting the intention interaction vector of each conversation.

[0045] For example, in some embodiments of the present application, the intention representation interaction module can use the self-attention structure (as a specific example of the self-attention model) to extract the intention interaction information between multiple conversations with a time sequence based on the intention representation vector. For example, when modeling the intention interaction information between multiple conversations with a time sequence based on the self-attention structure, the relative time offset position information is added. The intention interaction vector output by the intention representation interaction module is also 10 1*k dimensional vectors. Compared with the 10 1*k dimensional vectors of the intention representation extraction module, the vector output this time integrates the intention interaction information between multiple conversations. Among them, the relative time offset position information refers to the number of days from the current time of the conversation. The closer the conversation is to the current number of days, the more it can represent the user's recent intentions.

[0046] S230, calculating the user feature vector and the intention interaction vector of each conversation to obtain a user intention vector.

[0047] For example, in some embodiments of the present application, Figure 3 The global intention activation module in the server 200 uses the user feature vector corresponding to the user portrait information updated on a daily basis to activate the intention interaction information sequence (that is, the intention interaction vector), with the aim of extracting the user's more recent intention (as a specific example of the user intention vector).

[0048] In some embodiments of the present application, S230 may include: taking the feature vector corresponding to the latest user portrait information as the user feature vector; taking the inner product result of the user feature vector and the intention interaction vector of each conversation as the weight value of each conversation; and performing weighted summation of the intention interaction vector and the weight value of each conversation to obtain the user intention vector.

[0049] For example, in some embodiments of the present application, since the user portrait information updated on a daily basis (as a specific example of the latest user portrait information) can reflect the user's recent status, the present application mainly uses the characterization vector profile_embedding (as a specific example of a feature vector) of the user attribute keywords of the user portrait information updated on a daily basis, with a dimension of 1*k (the value of k is the same as above). Using the target-attention structure (i.e., the target attention mechanism), profile_embedding is used as the target vector (as a specific example of a user feature vector). Based on the 10 1*k dimensional vectors obtained by the intention characterization interaction module, the inner products of the target vector and the 10 1*k dimensional vectors corresponding to the 10 conversations are calculated as weighted weights. Figure 4 The diagram of global intent activation is shown in the figure. The daily updated user profile profile_embedding is the target vector. The 1st conversation intent embedding to the 10th conversation intent embedding are 10 1*k dimensional vectors output by the intent representation interaction module. The weights val1, val2, ..., val10 can be obtained by inner product calculation. Then, the 10 conversation intent embeddings and weighted weights are weighted summed to output a 1*k dimensional user intent vector. The user intent vector is obtained by global intent activation embedding based on the user's recent status, which can represent the user's recent real intention.

[0050] S240, determining a marketing score according to user attributes, telephone dialing statistical characteristics and the user intention vector, wherein the higher the marketing score is, the better the marketing conversion effect of the user is.

[0051] For example, in some embodiments of the present application, Figure 3 The fusion estimation module in the algorithm will comprehensively analyze the user's static attributes (as a specific example of user attributes), call statistical features (as a specific example of phone call statistical features) and the user's recent intention representation (that is, the user intention vector obtained above) to determine the marketing score.

[0052] In some embodiments of the present application, S240 may include: concatenating the user attributes, the telephone dialing statistical features and the user intention vector to obtain a concatenated feature vector; inputting the concatenated feature vector into a deep neural network model to output a fused feature vector; inputting the fused feature vector into a target activation function to obtain the marketing score. Among them, user attributes include: user age, user gender and user regional information; telephone dialing statistical features are obtained by at least counting the number of times the user's telephone dialing is connected and the duration of the connection.

[0053] For example, in some embodiments of the present application, user static attributes include: user age, gender, location (as a specific example of user location information) and other information. The dialing statistical feature is based on the recent dialing contact with the user, that is, the number of calls, the duration of the call and other information are aggregated and statistically obtained. The user's recent intention representation is based on the 1*k dimensional user intention vector obtained by the global intention activation module. Figure 3 As shown in the figure, the above three types of features are concat spliced ​​to obtain a spliced ​​feature vector. Then the spliced ​​feature vector is passed into the n-layer deep neural network DNN model for fusion, and the fused feature vector is output. Among them, n is an adjustable parameter, usually with a value of 3. Finally, the fused feature vector is predicted by sigmoid (as a specific example of the target activation function) to obtain the probability that the user will be converted by marketing, that is, the marketing score. Generally, the higher the marketing score of the user, the better the conversion effect.

[0054] It should be understood that the models, functions and values ​​of related parameters involved in the above implementation process can be flexibly selected according to the actual application scenario, and the embodiments of the present application are not limited to this.

[0055] The following is combined with Figure 5 The specific process of user marketing conversion prediction provided by some embodiments of the present application is exemplified.

[0056] Please see attached Figure 5 , Figure 5A flow chart of a method for predicting user marketing conversions provided for some embodiments of the present application.

[0057] The above process is explained below as an example.

[0058] S510, dividing multiple rounds of dialogue in each of the multiple dialogues by using a sliding window to obtain multiple dialogue segments.

[0059] S520, obtaining semantic features of each of the multiple conversation segments.

[0060] S530: construct an intention representation vector for each conversation based on the semantic features of each conversation segment.

[0061] S540: Input the intention representation vectors of multiple conversations into the self-attention model, and output the intention interaction vector of each conversation.

[0062] S550: Use the feature vector corresponding to the latest user portrait information as the user feature vector.

[0063] S560: taking the inner product results of the user feature vector and the intention interaction vector of each conversation as the weight value of each conversation.

[0064] S570, performing weighted summation on the intention interaction vector and weight value of each conversation to obtain a user intention vector.

[0065] S580, the user attributes, the telephone dialing statistical features and the user intention vector are spliced ​​and input into the deep neural network model, and a fused feature vector is output.

[0066] S590, input the fused feature vector into the target activation function to obtain a marketing score.

[0067] It can be understood that the specific implementation process of S510 to S590 can refer to the method embodiment provided above, and in order to avoid repetition, the detailed description is appropriately omitted here.

[0068] Through some of the above-mentioned embodiments of the present application, it can be known that the marketing prediction method based on LLM extraction of multi-session sequence representations provided by the present application includes four modules, namely: an intention representation extraction module based on the content of a single conversation; an intention representation interaction module based on the content of multiple conversations; a global intention activation module based on user portraits and a fusion estimation module based on session intentions (that is, user intention vectors), user static attributes, calling statistics and other features. The nested dependencies of each module play a role, which improves the joint representation effect of multi-pass and multi-round user conversation intentions and improves the marketing conversion effect.

[0069] Please refer to Figure 6 , Figure 6The following is a block diagram of the components of the device for predicting user marketing conversions provided by some embodiments of the present application. It should be understood that the device for predicting user marketing conversions corresponds to the above method embodiment and can execute each step involved in the above method embodiment. The specific functions of the device for predicting user marketing conversions can be found in the above description. To avoid repetition, the detailed description is appropriately omitted here.

[0070] Figure 6 The device for predicting user marketing conversion includes at least one software function module that can be stored in a memory in the form of software or firmware or solidified in the device for predicting user marketing conversion, and the device for predicting user marketing conversion includes: an intention acquisition module 610, used to obtain the intention representation vector of each conversation in multiple user conversations; an interaction information acquisition module 620, used to obtain the intention interaction vector of each conversation based on the intention representation vector of multiple conversations; a global intention acquisition module 630, used to calculate the user feature vector and the intention interaction vector of each conversation to obtain the user intention vector; a marketing prediction module 640, used to determine the marketing score according to user attributes, telephone dialing statistical characteristics and the user intention vector, wherein the higher the marketing score, the better the marketing conversion effect of the user.

[0071] In some embodiments of the present application, the intent acquisition module 610 is used to segment multiple rounds of conversations in each conversation through a sliding window to obtain multiple conversation segments; obtain semantic features of each of the multiple conversation segments; and construct the intent representation vector of each conversation based on the semantic features of each conversation segment.

[0072] In some embodiments of the present application, the interaction information acquisition module 620 is used to input the intention representation vector of each conversation into the self-attention model and output the intention interaction vector of each conversation.

[0073] In some embodiments of the present application, the global intent acquisition module 630 is used to use the feature vector corresponding to the latest user portrait information as the user feature vector; use the inner product result of the user feature vector and the intention interaction vector of each conversation as the weight value of each conversation; and perform weighted summation on the intention interaction vector and the weight value of each conversation to obtain the user intention vector.

[0074] In some embodiments of the present application, the marketing prediction module 640 is used to splice the user attributes, the telephone dialing statistical features and the user intention vector to obtain a spliced ​​feature vector; input the spliced ​​feature vector into a deep neural network model to output a fused feature vector; input the fused feature vector into a target activation function to obtain the marketing score.

[0075] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the device described above can refer to the corresponding process in the aforementioned method, and will not be described in detail here.

[0076] Some embodiments of the present application further provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, can implement the operations of the method corresponding to any of the above methods provided in the above embodiments.

[0077] Some embodiments of the present application further provide a computer program product, which includes a computer program, wherein when the computer program is executed by a processor, it can implement the operations corresponding to any of the above methods provided in the above embodiments.

[0078] like Figure 7 As shown, some embodiments of the present application provide an electronic device 700, which includes: a memory 710, a processor 720, and a computer program stored in the memory 710 and executable on the processor 720, wherein the processor 720 can implement a method as described in any of the above embodiments when reading the program from the memory 710 through a bus 730 and executing the program.

[0079] Processor 720 can process digital signals and can include various computing structures, such as complex instruction set computer structure, reduced instruction set computer structure, or a structure that implements a combination of multiple instruction sets. In some examples, processor 720 can be a microprocessor.

[0080] The memory 710 may be used to store instructions executed by the processor 720 or data related to the execution of instructions. These instructions and / or data may include codes for implementing some or all functions of one or more modules described in the embodiments of the present application. The processor 720 of the disclosed embodiment may be used to execute instructions in the memory 710 to implement the method shown above. The memory 710 includes a dynamic random access memory, a static random access memory, a flash memory, an optical memory, or other memory known to those skilled in the art.

[0081] The above description is only an embodiment of the present application and is not intended to limit the scope of protection of the present application. For those skilled in the art, the present application may have various changes and variations. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application should be included in the scope of protection of the present application. It should be noted that similar reference numerals and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in the subsequent drawings.

[0082] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

[0083] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

Claims

1. A method for predicting user marketing conversion, characterized in that: include: Obtain the intent representation vector of each conversation among multiple conversations of the user; Based on the intention representation vectors of multiple conversations, obtaining the intention interaction vector of each conversation; Calculating the user feature vector and the intention interaction vector of each conversation to obtain a user intention vector; A marketing score is determined based on user attributes, telephone dialing statistical features and the user intention vector, wherein the higher the marketing score is, the better the marketing conversion effect of the user is.

2. The method according to claim 1, characterized in that The obtaining of the intention representation vector of each conversation in the user's multiple conversations includes: Use sliding windows to segment multiple rounds of conversations in each conversation to obtain multiple conversation segments; Acquire the semantic features of each of the multiple conversation segments; The intention representation vector of each conversation is constructed based on the semantic features of each conversation segment.

3. The method according to claim 1 or 2, characterized in that The acquiring the intention interaction vector of each conversation based on the intention representation vector of multiple conversations includes: The intention representation vector of each conversation is input into the self-attention model, and the intention interaction vector of each conversation is output.

4. The method according to claim 1 or 2, characterized in that: The calculating the user feature vector and the intention interaction vector of each conversation to obtain the user intention vector includes: Using the feature vector corresponding to the latest user portrait information as the user feature vector; Taking the inner product results of the user feature vector and the intention interaction vector of each conversation as the weight value of each conversation; The intention interaction vector and the weight value of each conversation are weightedly summed to obtain the user intention vector.

5. The method according to claim 1 or 2, characterized in that: The determining of the marketing score according to the user attributes, the telephone dialing statistical characteristics and the user intention vector includes: Concatenate the user attributes, the telephone dialing statistical features and the user intention vector to obtain a concatenated feature vector; Inputting the concatenated feature vector into a deep neural network model and outputting a fused feature vector; The fused feature vector is input into the target activation function to obtain the marketing score.

6. The method according to claim 5, characterized in that The user attributes include: user age, user gender and user region information; the telephone dialing statistical characteristics are obtained by collecting statistics on the number of times the user's telephone dialing is connected, the connection time and other information.

7. A device for predicting user marketing conversion, characterized in that: include: An intent acquisition module is used to obtain the intent representation vector of each conversation in multiple conversations of the user; An interaction information acquisition module, configured to acquire an intention interaction vector of each conversation based on the intention representation vectors of multiple conversations; A global intention acquisition module, used to calculate the user feature vector and the intention interaction vector of each conversation to obtain a user intention vector; The marketing prediction module is used to determine the marketing score according to the user attributes, the telephone dialing statistical characteristics and the user intention vector, wherein the higher the marketing score, the better the marketing conversion effect of the user.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, wherein the computer program executes the method according to any one of claims 1 to 6 when executed by a processor.

9. An electronic device, characterized in that: The method comprises a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the computer program executes the method according to any one of claims 1 to 6 when being run by the processor.

10. A computer program product, characterized in that The computer program product comprises a computer program, wherein the computer program executes the method according to any one of claims 1 to 6 when executed by a processor.

Images