CN114254088A

CN114254088A - Method for constructing automatic response model and automatic response method

Info

Publication number: CN114254088A
Application number: CN202111477784.3A
Authority: CN
Inventors: 许胜强; 王岩; 屠邦燕; 胡加学
Original assignee: iFlytek Co Ltd
Current assignee: iFlytek Co Ltd
Priority date: 2021-12-06
Filing date: 2021-12-06
Publication date: 2022-03-29

Abstract

The invention provides a construction method of an automatic response model and the automatic response method, wherein the construction method of the automatic response model comprises the following steps: performing feature extraction on target reply information to generate first identification information corresponding to the target reply information, wherein the target reply information is client reply information under the condition of machine marketing failure; and determining at least one target marketing tactical message from candidate manual marketing tactical messages corresponding to the candidate customer reply messages on the basis of the first identification message, wherein the candidate customer reply messages are all customer reply messages under the condition that manual marketing is successful. The method for constructing the automatic response model can improve the conversion rate of machine marketing.

Description

Method for constructing automatic response model and automatic response method

Technical Field

The invention relates to the technical field of information processing, in particular to a construction method of an automatic response model and an automatic response method.

Background

Currently, the outbound robot is widely used in marketing scenarios. In the related art, a conversation with a client is generally realized by setting fixed speech information for an outbound robot, but the method cannot automatically judge whether the related speech is effective, and still needs manual intervention in actual execution to improve the success rate and the conversion rate of marketing, so that the effectiveness is low, and higher labor cost and time cost are consumed.

Disclosure of Invention

The invention provides a construction method of an automatic response model and an automatic response method, which are used for solving the defect of low conversion rate of machine marketing in the prior art and improving the conversion rate of machine marketing.

The invention provides a method for constructing an automatic response model, which comprises the following steps:

performing feature extraction on target reply information to generate first identification information corresponding to the target reply information, wherein the target reply information is client reply information under the condition of machine marketing failure;

and determining at least one target marketing tactical message from candidate manual marketing tactical messages corresponding to the candidate customer reply messages on the basis of the first identification message, wherein the candidate customer reply messages are all customer reply messages under the condition that manual marketing is successful.

According to the method for constructing the automatic response model, the step of determining at least one target marketing communication information from candidate artificial marketing communication information corresponding to candidate customer reply information based on the first identification information comprises the following steps:

screening the first identification information to generate a set of second identification information;

performing deduplication on target reply information corresponding to the second identification information to generate a set of first client reply information corresponding to the second identification information, wherein each first client reply information in the set of first client reply information has the same meaning and different meanings;

and matching to obtain at least one target marketing tactical information corresponding to the second identification information from the candidate manual marketing tactical information based on the set of the first customer reply information and the above machine marketing tactical information corresponding to the first customer reply information.

According to the method for constructing the automatic response model provided by the invention, the step of obtaining at least one target marketing tactical information corresponding to the second identification information from the candidate artificial marketing tactical information by matching based on the set of the first customer reply information and the above machine marketing tactical information corresponding to the first customer reply information comprises the following steps:

determining first similarity information between the first client reply information and each candidate client reply information;

under the condition that the first similarity information exceeds a first target threshold value, determining candidate client reply information corresponding to the first similarity information as second client reply information;

determining second similarity information of the machine marketing communication information corresponding to the first customer reply information and the manual marketing communication information corresponding to the second customer reply information;

and determining the following manual marketing communication information corresponding to the second similarity information as the target marketing communication information under the condition that the second similarity information exceeds a second target threshold value.

According to the method for constructing the automatic response model provided by the invention, the determining of the first similarity information between the first client response information and each candidate client response information comprises the following steps:

performing feature extraction on the first client reply message to generate a first feature vector code and a first key information set;

performing feature extraction on the candidate client reply information to generate a first target feature vector code and a first target key information set;

determining first similarity information between the first customer reply information and each of the candidate customer reply information based on the first feature vector encoding, the first key information set, the first target feature vector encoding, and the first target key information set;

and/or the presence of a gas in the gas,

the determining second similarity information of the machine marketing tactical information of the upper part corresponding to the first customer reply information and the manual marketing tactical information of the upper part corresponding to the second customer reply information includes:

performing feature extraction on the above machine marketing tactical information corresponding to the first customer reply information to generate a second feature vector code and a second key information set;

extracting the characteristics of the above artificial marketing tactical information corresponding to the second customer reply information to generate a second target characteristic vector code and a second target key information set;

determining second similarity information of the above machine marketing language information and the above artificial marketing language information based on the second feature vector encoding, the second set of key information, the second target feature vector encoding, and the second set of target key information.

According to the method for constructing the automatic response model provided by the invention, the step of extracting the characteristics of the target reply information to generate the first identification information corresponding to the target reply information comprises the following steps:

extracting features of the target reply information to generate third identification information and probability values corresponding to the third identification information;

and under the condition that the probability value exceeds a third target threshold value, determining third identification information corresponding to the probability value as the first identification information.

According to the method for constructing the automatic response model provided by the invention, after the at least one target marketing communication information is determined from the candidate artificial marketing communication information corresponding to the candidate customer reply information based on the first identification information, the method further comprises the following steps:

performing feature extraction on the target marketing tactical information to generate a plurality of types of labels corresponding to the target marketing tactical information;

based on a target conversational structured classification system, performing structured processing on the multiple types of labels to generate structured classification labels corresponding to the target marketing conversational information;

and generating a target marketing strategy corresponding to the target reply information based on the structured classification label.

The invention also provides an automatic response method, which comprises the following steps:

acquiring reply information of a client to be responded;

performing feature extraction on the reply information of the client to be responded to generate target identification information corresponding to the reply information of the client to be responded;

inputting the target identification information into the automatic response model generated by the automatic response model construction method described in any one of the above, and generating target response information.

The invention also provides a device for constructing the automatic response model, which comprises:

the first generation module is used for extracting features of target reply information and generating first identification information corresponding to the target reply information, wherein the target reply information is client reply information under the condition that the machine marketing fails;

and the first determining module is used for determining at least one target marketing tactical message from candidate artificial marketing tactical messages corresponding to the candidate customer reply messages on the basis of the first identification message, wherein the candidate customer reply messages are all customer reply messages under the condition that artificial marketing is successful.

The present invention also provides an automatic answering device, comprising:

the acquisition module is used for acquiring reply information of the client to be responded;

the second generation module is used for extracting the characteristics of the reply information of the client to be responded and generating target identification information corresponding to the reply information of the client to be responded;

and a third generating module, configured to input the target identification information into the automatic response model generated by the automatic response model constructing method according to any one of the above descriptions, and generate target response information.

The invention further provides an electronic device, which comprises a memory, a processor and a computer program stored on the memory and capable of running on the processor, wherein the processor executes the computer program to realize the steps of the method for constructing the automatic response model or the steps of the automatic response method.

The present invention also provides a non-transitory computer readable storage medium having stored thereon a computer program which, when being executed by a processor, implements the steps of the method for constructing an automatic response model or the steps of the automatic response method as described in any of the above.

The present invention provides a computer program product comprising a computer program which, when executed by a processor, implements the steps of the method for constructing an automatic response model or the steps of the automatic response method as described in any of the above.

According to the automatic response model construction method and the automatic response method, the target reply information under the condition that the machine marketing is failed is subjected to feature extraction to generate the first identification information corresponding to the target reply information, and based on the first identification information, at least one target marketing communication information is determined from the candidate artificial marketing communication information corresponding to the customer reply information under the condition that the artificial marketing is successful, so that the incidence relation between the first identification information and the target marketing communication information can be established, the method is beneficial to rapidly matching the corresponding marketing communication information based on the customer reply information in the subsequent application process, and the marketing conversion rate is improved.

Drawings

In order to more clearly illustrate the technical solutions of the present invention or the prior art, the drawings needed for the description of the embodiments or the prior art will be briefly described below, and it is obvious that the drawings in the following description are some embodiments of the present invention, and those skilled in the art can also obtain other drawings according to the drawings without creative efforts.

FIG. 1 is a schematic flow chart of a method for constructing an auto-answer model according to the present invention;

FIG. 2 is a second schematic flow chart of the method for constructing an auto-answer model according to the present invention;

FIG. 3 is a third schematic flow chart of a method for constructing an auto-answer model according to the present invention;

FIG. 4 is a fourth schematic flowchart of a method for constructing an auto-answer model according to the present invention;

FIG. 5 is a schematic flow chart of an automatic answering method provided by the present invention;

FIG. 6 is a schematic structural diagram of an apparatus for constructing an automatic response model according to the present invention;

FIG. 7 is a schematic structural diagram of an automatic answering machine provided by the present invention;

fig. 8 is a schematic structural diagram of an electronic device provided in the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings, and it is obvious that the described embodiments are some, but not all embodiments of the present invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

The method of constructing the auto-answer model of the present invention is described below with reference to fig. 1 to 4.

It should be noted that the execution subject of the method for constructing an automatic response model of the present invention may be a device for constructing an automatic response model, or a server, or a terminal of a user, such as a mobile phone or a computer.

As shown in fig. 1, the method for constructing the automatic response model includes: step 110 and step 120.

110, extracting features of the target reply information to generate first identification information corresponding to the target reply information, wherein the target reply information is client reply information under the condition of machine marketing failure;

in the outbound scenario, the session between the robot service and the client may be a session initiated by the client or a session initiated by the robot service.

The session initiated by the client is called a voice navigation robot.

The session actively initiated by the robot customer service is called an outbound robot.

In the whole process of calling out the robot, the robot uttered by the robot and the reply information replied by the client for each robot are involved.

The current machine outbound scenario is mainly divided into: a notification class, a return visit class, and a marketing class. The notification scene and the return visit scene are relatively simple, and the customer matching degree is high; the marketing scenario is a recommendation of a product to a customer.

It can be appreciated that in an out-of-machine call marketing scenario, the final marketing results achieved include marketing failures and marketing successes. For example, when the outbound robot successfully recommends a certain product to the customer, it indicates that the marketing is successful; when the customer refuses to purchase the recommended product, marketing failure is indicated.

In this step, all call information generated in the marketing process can be converted into text information, and the target reply information is text information corresponding to all client reply information under the condition that the machine marketing fails in the historical marketing process.

It will be appreciated that in each case of machine marketing failure, there is at least one customer reply message.

In the actual execution process, the client reply information under the condition of machine marketing failure in the historical process can be stored in the local database or the cloud database, and the target reply information can be obtained by calling when needed.

The first identification information may be a tag in the form of a word or a phrase, etc. for characterizing the core meaning of the target reply information.

It can be understood that, different target reply messages have different meanings, and the corresponding first identification information is also different.

By performing feature extraction on the plurality of target reply messages, a plurality of first identification messages can be obtained. The same first identification information may exist in the plurality of first identification information, and different first identification information may also exist.

By classifying the plurality of pieces of first identification information, the first identification information in different types can be obtained.

The first identification information for the case of machine marketing failure may include: the system comprises a plurality of large-class labels, a plurality of large-class labels and a plurality of large-class labels, wherein the large-class labels are required by users but do not have handling conditions, the large-class labels are required by users but do not operate, and the large-class labels are required by users but do not have verification and confirmation succeeds.

A plurality of subclass labels can also be included under each of the major class labels.

For example, taking a marketing traffic packet as an example, under a large category of labels where the user has no demand, the marketing traffic packet may include: the user directly hangs up without speaking in the whole process, the user directly rejects without reasons, the user expresses enough flow, and the user is satisfied with the current package.

Under this broad category of labels where user requirements are not certain, this may include: the expression is temporarily not needed, the expression is considered, the expression is busy, the expression is sent out with short messages, the worry that the cancellation is forgotten, the business hall is handled, the suspicion cost is too high, the number is not commonly used and other subclass labels are used.

Under the broad category of labels that the user has a requirement but does not have a handling condition, the following can be included: minor cards, bundled other services, and minor labels such as non-decisioners.

Under the broad category of tags that the user has a need but will not operate, it may include: and the minor labels such as key pressing and the like are avoided.

Under a broad class of tags where the user has a need but the validation is not successful, it may include: agreement but no key press confirmation, agreement but a key press error, etc.

It should be noted that, in the present invention, the first identification information may include, but is not limited to, a plurality of minor tags under a plurality of major tags.

For example, in the case of recommending a flow package to a user, after the call is connected, the outbound robot carries out product recommendation, and the user replies that the message is no longer used and is enough at me, and then the message is the target reply message.

And extracting the characteristics of the target reply information to obtain key information of 'enough flow', wherein the 'enough flow' is the first identifier corresponding to the target reply information.

After the first identification information is generated, the first identification information can be stored in a local database or a cloud database and can be called when needed.

And step 120, determining at least one target marketing tactical message from candidate manual marketing tactical messages corresponding to the candidate customer reply messages based on the first identification message, wherein the candidate customer reply messages are all customer reply messages under the condition that manual marketing is successful.

In this step, the candidate customer reply message is the whole reply message of the user to the manual marketing dialogue under the condition that the manual marketing is successful.

And at least one piece of customer reply information is included for each case that the manual marketing is successful.

The candidate manual marketing tactical information is context manual marketing tactical information corresponding to each sentence of customer reply information under the condition that manual marketing is successful.

It is understood that each candidate customer reply message corresponds to at least one manual marketing tactical message.

It should be noted that, in the actual marketing process, under the manual marketing scenario, after the customer refuses, the customer is rescued by the manual marketing tactical information, and finally persuade the customer, thereby realizing the successful marketing condition; in the machine marketing scenario, after the customer says the same reply information and rejects, the robot fails to successfully retrieve the customer, thereby causing marketing failure.

Table 1 shows partial call contents in two different situations, namely, a machine marketing failure situation and a manual marketing success situation, where the left side is the machine marketing call information corresponding to the customer reply information and the client reply information in the machine marketing failure situation, and the right side is the manual marketing call information corresponding to the client reply information and the client reply information in the manual marketing success situation.

As can be seen from table 1, after the robot and the agent have finished the introduction, the user X and the user Y make responses with the same meaning, and then the robot and the agent have sent out subsequent callback techniques for the response information of the user, and the user X hangs up directly, resulting in the failure of machine marketing, while the user Y agrees to the handling, the successful callback, and the success of manual marketing.

For table 1, the reply information of the user X is the target reply information, and the tag extracted based on the reply information of the user X is the first identification information corresponding to the reply information of the user X; the reply information of the user Y is the reply information of the candidate client, the dialect information of the agent is the candidate manual marketing dialect information corresponding to the reply information of the user Y, and the actual meanings of the first identification information corresponding to the reply information of the user X and the reply information of the user Y are the same or similar.

TABLE 1

In this step, the first identification information and the corresponding target reply information are processed to match and obtain the target marketing communication information corresponding to the first identification information, so that the association relationship between the first identification information and the target marketing communication information can be established.

According to the method for constructing the automatic response model, provided by the embodiment of the invention, the first identification information corresponding to the target response information is generated by extracting the characteristics of the target response information under the condition that the machine marketing is failed, and at least one target marketing communication information is determined from the candidate artificial marketing communication information corresponding to the customer response information under the condition that the artificial marketing is successful on the basis of the first identification information, so that the incidence relation between the first identification information and the target marketing communication information can be established, the method is beneficial to quickly matching the corresponding marketing communication information on the basis of the customer response information in the subsequent application process, and the marketing conversion rate is improved; in addition to this, a closed loop optimization from human-computer interaction analysis to analysis by human-human dialogue can be achieved.

In some embodiments, step 110 comprises:

In this embodiment, the third identification information may be a tag in the form of a word or a phrase, etc. for characterizing the meaning that may be contained in the target reply information.

It is understood that one or more third identification information may be extracted for the same target reply message.

And the probability value corresponding to the third identification information is used for representing the similarity between the core meanings contained in the third identification information and the target reply information.

The higher the probability value is, the closer the third identification information is to the core meaning contained in the target reply information is.

The third target threshold is an evaluation criterion used for judging whether the third identification information can be used for representing the target reply information. The third target threshold may be user-defined, such as set to 80% or 70%.

And under the condition that the probability value corresponding to the third identification information exceeds a third target threshold value, determining the third identification information as the first identification information by approximately considering that the third identification information can be used for representing the meaning to be expressed by the target reply information.

In the actual execution process, a sort layer Softmax function can be added by using a BERT model in natural language processing to predict first identification information corresponding to the target reply information.

Among them, BERT (bidirectional Encoder retrieval from transformations) is a pre-trained language characterization model used to generate deep bidirectional language characterization.

The Softmax function, also called normalized exponential function, is used to show the result of multi-classification in the form of probability.

The specific network structure is shown in fig. 2. Wi represents the ith character corresponding to the text of the target reply message, [ CLS ] and [ SEP ] are special identifiers added at the beginning and the end of the text of the target reply message, Ei is the initial code corresponding to the ith character in the text of the target reply message, and Hi is the code of the ith character after being processed by a BERT model.

The core idea of the whole network structure is as follows: firstly, obtaining the characteristic representation corresponding to the target reply information by using a BERT model, and then obtaining the probability values corresponding to all label categories by the characteristic representation of the user text CLS position corresponding to each sentence of target reply information through a full connection layer and a Softmax function, namely obtaining the third identification information and the probability values corresponding to all the third identification information.

Comparing the probability value with a third target threshold value, if the probability value of the label category is not lower than the preset third target threshold value, indicating that the probability value corresponding to the label category is larger, and outputting a corresponding label, wherein the label is the first identification information.

In some embodiments, in the case that the probability value corresponding to the third identification information is lower than the third target threshold, the unknown tag is output.

According to the method for constructing the automatic response model, provided by the embodiment of the invention, the target reply information is subjected to feature extraction, the first identification information corresponding to each target reply information is obtained through screening based on the probability value, the extraction precision is high, and the accuracy of the finally obtained first identification information is high.

In some embodiments, step 120 comprises:

the target reply information corresponding to the second identification information is subjected to duplication elimination, a set of first client reply information corresponding to the second identification information is generated, and all first client reply information in the set of the first client reply information has the same meaning and different meanings;

and matching to obtain at least one target marketing tactical information corresponding to the second identification information from the candidate artificial marketing tactical information based on the set of the first customer reply information and the above machine marketing tactical information corresponding to the first customer reply information.

In this embodiment, the second identification information is a tag with higher user percentage or stronger association with the user in the first identification information.

The number of the second identification information should not exceed the number of the first identification information.

In an actual execution process, the first identification information may be screened based on a user proportion corresponding to each first identification information or an association relationship between each first identification information and a user, the first identification information with a higher user proportion and/or a larger influence on the user is selected as the second identification information, and a set including a plurality of second identification information is generated.

It can be understood that each type of second identification information corresponds to at least one target reply information, and the target reply information corresponding to the same type of second identification information has the same meaning but may be said to be different.

Table 2 shows two types of second identification information and target reply information corresponding to the second identification information.

TABLE 2

In the actual execution process, a plurality of target reply messages corresponding to each type of second identification information are respectively screened, the descriptions which are basically similar are deleted, the user descriptions which have the same expression meaning but have larger differences are reserved and determined as first client reply messages, and a plurality of first client reply messages corresponding to each type of second identification information form a set of client reply messages corresponding to the type of second identification information.

After the first customer reply message is obtained, the machine marketing tactics message corresponding to the first customer reply message is matched based on the first customer reply message.

It is understood that each first customer reply message corresponds to at least one of the above machine marketing tactics messages.

After the first customer reply information and the corresponding upper machine marketing communication information are obtained, the candidate artificial marketing communication information which is obtained by matching the candidate artificial marketing communication information and has higher relevance with the first customer reply information is used as the target marketing communication information corresponding to the second identification information based on the similarity between the first customer reply information and the candidate customer reply information and the similarity between the upper machine marketing communication information corresponding to the first customer reply information and the upper artificial marketing communication information corresponding to the candidate customer reply information.

It should be noted that one or more target marketing tactics information corresponding to the second identification information may be provided.

In this embodiment, by preprocessing the first identification information and the target reply information, the data volume in the subsequent data processing process can be effectively reduced, thereby improving the data processing rate.

The following describes the determination process of similarity in this application with specific embodiments.

In some embodiments, the matching in step 120 to obtain at least one target marketing tactical information corresponding to the second identification information from the candidate manual marketing tactical information based on the set of first customer reply information and the above machine marketing tactical information corresponding to the first customer reply information includes:

determining second similarity information of the machine marketing tactical information corresponding to the first customer reply information and the manual marketing tactical information corresponding to the second customer reply information;

and under the condition that the second similarity information exceeds a second target threshold value, determining the following manual marketing tactical information corresponding to the second similarity information as the target marketing tactical information.

In this embodiment, the first similarity information is used to characterize the degree of similarity between the first client reply message and the candidate client reply message.

The first target threshold may be user-defined, and is not limited in this application.

And the second customer reply message is a customer reply message with higher similarity degree with the first customer reply message under the condition that the manual marketing is successful.

After the first similarity information is obtained, comparing the first similarity information with a first target threshold, and determining candidate client reply information corresponding to the first similarity information as second client reply information if the similarity between the first client reply information corresponding to the first similarity information and the candidate client reply information is approximately considered to be higher under the condition that the first similarity information exceeds the first target threshold.

It is understood that the same type of second identification information may correspond to a plurality of first client reply messages, and each first client reply message may correspond to a plurality of second client reply messages.

Wherein, the first client replies the message and the second client replies the message with the same or similar language; under the same application scene, the meanings of the two expressions can be the same; however, in different application scenarios, the two actually expressed meanings may be different.

As shown in table 3, the client replies with the same language but with different meanings.

TABLE 3

In an embodiment of the present application, after obtaining the reply information of the second client, the reply information of the second client may be further filtered, which may specifically be represented as:

The second similarity information is used for representing the similarity degree between the machine marketing communication information corresponding to the first customer reply information and the manual marketing communication information corresponding to the second customer reply information.

And the target marketing tactical information is the following manual marketing tactical information corresponding to the reply information of the second customer under the condition that the manual marketing is successful.

The second target threshold may be user-defined, and is not limited in this application.

After the second similarity information is obtained, comparing the second similarity information with a second target threshold, and under the condition that the second similarity information exceeds the second target threshold, approximately considering that the similarity between the above machine marketing communication corresponding to the first customer reply information corresponding to the second similarity information and the above manual marketing communication corresponding to the second customer reply information is higher, namely the application scenes between the first customer reply information and the second customer reply information are approximately the same, and determining the below manual marketing communication information corresponding to the above manual marketing communication information corresponding to the second similarity information as the target marketing communication information.

And finally, mining and obtaining a user similarity method under the condition of successful manual marketing based on the target reply information corresponding to the second identification information and the machine marketing communication information above the target reply information, and determining the user similarity method and the manual marketing communication information below the user similarity method as the corresponding target marketing communication information corresponding to the second identification information.

The following describes steps of generating the first similarity information and the second similarity information by using a specific embodiment.

In some embodiments, determining first similarity information between the first client reply message and each candidate client reply message comprises:

and determining first similarity information between the first client reply information and each candidate client reply information based on the first feature vector code, the first key information set, the first target feature vector code and the first target key information set.

In the embodiment, the first feature vector code corresponding to the first client reply message is a feature vector code at a text word level of the first client reply message;

the first key information set corresponding to the first customer reply information is a word expressing sentence key information or semantic fragments in the text of the first customer reply information.

In the actual execution process, the feature vector extraction of the text word level can be carried out on the first customer reply message through the BERT model, and a first feature vector code is generated.

And performing key token extraction on the first customer reply information through a BERT + ATT model to generate a first key information set. The Token category is generally classified into a skill category, a business category, a decoration category and a sentence category.

The skill class is operation words which represent related services, such as query, handling, change and the like; the service class is a specific service object in the field, such as flow, telephone charge, flow packet and the like; the decoration class is attributes contained in a service product, such as specific amount, specific flow, excess, last month, next month, main card, auxiliary card, sharing and the like; sentence category, which is a category for expressing sentences, such as statements and questions.

For example, enter first customer reply message 1: if the flow package is not processed by me, the Token extracted correspondingly is as follows: handling the question of the # flow packet #; input first customer reply message 2: looking up how much traffic I still has, the Token extracted correspondingly is: query # traffic # query.

As shown in fig. 3, taking the first customer reply message as the user utterance X as an example, extracting feature vectors of a text word level for the user utterance X through a BERT model to generate a first feature vector code Ex; and performing key token extraction through a BERT + ATT model user utterance X to generate a first key information set Tx.

According to the same method, feature extraction may be performed on the candidate client reply information to generate a first target feature vector code and a first target key information set corresponding to the candidate client reply information, which are not described herein again.

The first target characteristic vector code is a characteristic vector code of a text word level of the candidate client reply message;

the first target key information set is a word expressing sentence key information or semantic fragments in the text of the candidate client reply information.

After generating the first feature vector code, the first key information set, the first target feature vector code and the first target key information set, based on the first feature vector code and the first target feature vector code, first sub-similarity information of the first customer reply information and each candidate customer reply information on a sentence text vector level may be generated; second sub-similarity information of the first client reply information and each candidate client reply information on a Token level can be generated based on the first key information set and the first target key information set.

In actual implementation, the following formula can be used:

and generating first sub-similarity information, wherein Ex1 is the first feature vector code of the first client reply information, Ex2 is the first target feature vector code of the candidate client reply information, and Sim _ vector (Ex1, Ex2) is the first sub-similarity information.

It is understood that the smaller the value of the first sub-similarity information, the smaller the similarity is described.

By the formula:

generating second sub-similarity information, wherein Tx1 is a first key information set of first client reply information, Tx2 is a first target key information set of candidate client reply information, beta is a harmonic factor, beta is a value between 0 and 1, a Levenshtein distance refers to the minimum number of editing operations required for converting one string into the other string between the two strings, len is the word length corresponding to a key token fragment, and max is the maximum value; sim _ token (Tx1, Tx2) is the second sub-similarity information.

Similarly, the smaller the value of the second sub-similarity information, the smaller the similarity is described.

The first sub-similarity information and the second sub-similarity information are calculated by a weighted fusion algorithm, and the first similarity information can be generated.

In actual implementation, the following formula can be used:

and generating first similarity information, wherein X1 is first customer reply information, X2 is candidate customer reply information, and S (X1, X2) is a weighted fusion score of vector measurement and key token measurement, namely the first similarity information, alpha is a harmonic factor, and alpha takes a value between 0 and 1.

It should be noted that, when the weighted fusion algorithm is performed on the first sub-similarity information and the second sub-similarity information, the weight α of the first sub-similarity and the second sub-similarity may be customized by the user, or a default value is adopted, or may be obtained by continuous optimization in the subsequent calculation process.

For example, in the initial calculation, the weight value may be set to 1:1, and in the subsequent training process, the weight value adopted in the previous training is continuously optimized and adjusted based on the importance levels of the text vector level and Token level of the sentence, or based on the output result, and finally the optimal weight value is determined.

In some embodiments, determining second similarity information for the machine marketing verbal information of the first customer corresponding to the reply message and the manual marketing verbal information of the second customer corresponding to the reply message comprises:

extracting the characteristics of the machine marketing tactical information corresponding to the first customer reply information to generate a second characteristic vector code and a second key information set;

and determining second similarity information of the above machine marketing tactical information and the above artificial marketing tactical information based on the second feature vector code, the second key information set, the second target feature vector code and the second target key information set.

In the embodiment, feature extraction is carried out on the above machine marketing tactical information corresponding to the first customer reply information, and a second feature vector code and a second key information set corresponding to the above machine marketing tactical information are generated; and performing feature extraction on the above artificial marketing tactical information corresponding to the second customer reply information to generate a second target feature vector code and a second target key information set corresponding to the above artificial marketing tactical information.

The second feature vector code is a feature vector code of a text word level of the above machine marketing tactical information corresponding to the first customer reply information;

the second key information set is a word expressing sentence key information or semantic fragments in the text of the above machine marketing tactical information corresponding to the first customer reply information.

The second target characteristic vector code is a characteristic vector code of a text word level of the above artificial marketing tactical information corresponding to the second customer reply information;

and the second target key information set dimension expresses words of sentence key information or semantic fragments in the text of the above manual marketing tactical information corresponding to the second customer reply information.

The specific feature extraction method is the same as the above embodiment, and may be implemented by using a BERT model and a BERT + ATT model, which are not described herein again.

After generating the second feature vector code, the second key information set, the second target feature vector code and the second target key information set, based on the second feature vector code and the second target feature vector code, third sub-similarity information of the above machine marketing tactical information corresponding to the first customer reply information and the above manual marketing tactical information corresponding to each second customer reply information on the sentence text vector level can be generated; based on the second key information set and the second target key information set, fourth sub-similarity information of the above machine marketing tactical information corresponding to the first customer reply information and the above artificial marketing tactical information corresponding to each second customer reply information on the Token level can be generated.

The calculation method of the third sub-similarity and the fourth sub-similarity is the same as the above embodiment, and is not repeated herein.

The second similarity information may be generated by performing a weighted fusion algorithm calculation on the third sub-similarity information and the fourth sub-similarity information.

The calculation method of the second similarity information is the same as the above embodiment, and is not described herein.

It should be noted that, when the weighted fusion algorithm is performed on the third sub-similarity information and the fourth sub-similarity information, the weights of the third sub-similarity and the fourth sub-similarity may be customized by a user, or default values may be adopted, or may be obtained by continuous optimization in a subsequent calculation process.

According to the method for constructing the automatic response model, the candidate client reply information similar to the target reply information is determined through the first similarity, the upper artificial marketing language information similar to the upper machine marketing language information corresponding to the target reply information is determined from the upper artificial marketing language information corresponding to the candidate client reply information through the second similarity, and the obtained similar users can be ensured to say that the corresponding upper marketing scenes and strategies are basically consistent; on the basis, the incidence relation between the target reply information with similar client reply information and the marketing tactic information and the manual marketing tactic information corresponding to the candidate reply information based on the second identification information is established, and the accuracy and the precision of the result are effectively improved.

As shown in fig. 4, after step 120, the method further comprises, in accordance with some embodiments of the present invention:

based on a target language structured classification system, carrying out structured processing on the multiple types of labels to generate structured classification labels corresponding to target marketing language information;

In this embodiment, the type tag may be a keyword or phrase in the targeted marketing tactical information for characterizing the core meaning of the targeted marketing tactical information.

Table 4 shows a corresponding table of several target reply messages and their corresponding type tags.

As shown in table 4, for the targeted marketing tactical information 1, a plurality of type tags corresponding thereto may be generated, such as expensive traffic overflow, cost-effective marketing campaign, traffic usage scenario introduction, and experience first.

And then, based on a target language structured classification system, carrying out structured processing on the plurality of types of labels to generate the sequence and the logical relationship among the types of labels, thereby generating the structured classification label.

In the actual execution process, a predefined tactical structured classification system can be adopted, based on a BERT + ATT model in natural language processing, each sentence content in the target marketing tactical information corresponding to each second identification information is assigned to a category by using a sliding window thought, and then a plurality of type labels corresponding to each target marketing tactical information after being structured are obtained.

After the structured classification label is generated, a target marketing strategy corresponding to the target reply information can be generated based on the structured classification label.

It will be appreciated that each targeted marketing tactical information corresponds to a plurality of structured classification tags.

On the basis, induction statistics is carried out on a plurality of structured classification labels corresponding to each target marketing strategy information, and then specific strategies corresponding to different marketing strategies can be obtained.

TABLE 4

For example, strategy 1: experience first price # s; strategy 2: traffic overflow is cost effective for # marketing campaigns, each strategy also has diversified conversational content.

In addition, aiming at each target marketing tactical information obtained by mining, combining a plurality of rounds of interactive record information of the upper and lower contexts, summarizing and analyzing the marketing path, not only can excellent marketing strategies and tactical contents corresponding to the second identification information which is not successfully marketed currently be obtained, but also the corresponding combined marketing strategies and tactical contents can be obtained.

In the application, the excellent marketing dialect and marketing strategy of gold medal seat personnel can be mined based on the historical interactive data of the machine unsuccessful marketing user, and in the subsequent process of applying the marketing dialect and the marketing strategy to the machine outbound marketing scene, the interactive experience of the user can be improved, the marketing conversion rate of the machine can be improved, and the revenue is increased.

According to the method for constructing the automatic response model, the target marketing tactic information is subjected to feature extraction to generate the structured classification label, the target marketing strategy is generated based on the structured classification label, and the excellent marketing strategy can be further mined on the basis of mining the excellent marketing tactic, so that the excellent marketing strategy and the excellent marketing tactic information corresponding to the first identification information of the user reply information which is not successfully marketed currently can be obtained, and the corresponding combined marketing strategy and tactic content can be obtained, so that the method is beneficial to finding new strategies which are difficult to think of artificial experience and brain holes, and effectively solves the problems that the marketing strategy is difficult to find, the tactic optimization is lack of pertinence, the marketing is not flexible and the like.

The following describes the automatic response model construction apparatus provided by the present invention, and the automatic response model construction apparatus described below and the automatic response model construction method described above may be referred to in correspondence with each other.

As shown in fig. 6, the apparatus for constructing an automatic response model includes: a first generation module 610 and a first determination module 620.

The first generating module 610 is configured to perform feature extraction on the target reply information, and generate first identification information corresponding to the target reply information, where the target reply information is a customer reply information in the case of machine marketing failure;

and a first determining module 620, configured to determine, based on the first identification information, at least one target marketing tactical information from candidate manual marketing tactical information corresponding to the candidate customer reply information, where the candidate customer reply information is all customer reply information in a case where manual marketing is successful.

According to the automatic response model construction device provided by the embodiment of the invention, the first identification information corresponding to the target response information is generated by performing feature extraction on the target response information under the condition that the machine marketing is failed, and based on the first identification information, at least one target marketing communication information is determined from candidate artificial marketing communication information corresponding to the customer response information under the condition that the artificial marketing is successful, so that the association relationship between the first identification information and the target marketing communication information can be established, the automatic response model construction device is beneficial to rapidly matching the corresponding marketing communication information based on the customer response information in the subsequent application process, and the marketing conversion rate is improved.

In some embodiments, the first determining module 620 is further configured to:

In some embodiments, the first generation module 610 is further configured to

In some embodiments, the apparatus further comprises:

the fourth generation module is used for performing feature extraction on the target marketing tactical information after determining at least one target marketing tactical information from the candidate manual marketing tactical information corresponding to the candidate customer reply information based on the first identification information, and generating a plurality of types of labels corresponding to the target marketing tactical information;

the fifth generation module is used for carrying out structural processing on the multiple types of labels based on the target marketing tactic structured classification system to generate structured classification labels corresponding to the target marketing tactic information;

and the sixth generating module is used for generating a target marketing strategy corresponding to the target reply information based on the structured classification label.

The following describes the automatic response method provided by the present invention, and the automatic response method described below and the method for constructing the automatic response model described above may be referred to correspondingly.

It should be noted that the execution subject of the automatic answering method may be an automatic answering device, or a server, or may also be a terminal of a user, such as a mobile phone or a computer.

As shown in fig. 5, the automatic answering method includes: step 510, step 520 and step 530.

Step 510, acquiring reply information of a client to be responded;

in the step, the customer reply information to be responded is the customer reply information collected in real time in the machine marketing process.

Step 520, performing feature extraction on the reply information of the client to be responded to, and generating target identification information corresponding to the reply information of the client to be responded;

in this step, the target identification information may be a tag in the form of a word or a phrase, etc. for characterizing the core meaning of the reply information of the client to be answered

After the real-time reply information of the client to be responded is obtained, feature extraction can be firstly carried out on the reply information of the client to be responded, and a plurality of pieces of identification information to be selected and probability values corresponding to the identification information to be selected are generated; and then determining the to-be-selected identification information with higher probability value as the target identification information based on the probability value.

The specific implementation manner is the same as that in the above embodiment, for example, a BERT model in natural language processing may be used to add a classification layer Softmax function to predict target identification information corresponding to the reply information of the client to be answered, which is not described herein again.

Step 530, inputting the target identification information into the automatic response model generated by the method for constructing the automatic response model, and generating the target response information.

In this step, the automatic response model is a model generated by a method for constructing the automatic response model, and is used to generate target response information corresponding to the target identification information based on the target identification information.

And the target response information is the following manual marketing tactical information corresponding to the customer reply information similar to the target response information under the condition that the manual marketing is successful.

The target identification information generated in step 520 is input to the automatic response model, and the target response information corresponding to the target identification information can be obtained through matching.

According to the automatic response method provided by the embodiment of the invention, the target identification information is generated by extracting the characteristics of the reply information of the client to be responded under the condition of machine marketing, then the target identification information is matched based on the automatic response model, and the following manual marketing communication information corresponding to the target identification information under the condition of successful manual marketing is obtained through matching, so that the optimal marketing communication can be generated based on the real-time reply information of the client, the matching rate is high, the accuracy of the matching result is high, and the conversion rate of machine marketing is favorably improved.

The following describes the automatic answering device provided by the present invention, and the automatic answering device described below and the automatic answering method described above can be referred to correspondingly.

As shown in fig. 7, the automatic answering device includes: an acquisition module 710, a second generation module 720, and a third generation module 730.

An obtaining module 710, configured to obtain reply information of the client to be answered;

the second generating module 720 is configured to perform feature extraction on the reply information of the client to be responded, and generate target identification information corresponding to the reply information of the client to be responded;

a third generating module 730, configured to input the target identification information into the automatic response model generated by the automatic response model constructing method according to any one of the above embodiments, and generate the target response information.

According to the automatic response device provided by the embodiment of the invention, the target identification information is generated by extracting the characteristics of the reply information of the client to be responded under the machine marketing condition, then the target identification information is matched based on the automatic response model, and the following manual marketing communication information corresponding to the target identification information under the condition of successful manual marketing is obtained through matching, so that the optimal marketing communication can be generated based on the real-time reply information of the client, the matching speed is high, the accuracy of the matching result is high, and the conversion rate of machine marketing is favorably improved.

Fig. 8 illustrates a physical structure diagram of an electronic device, and as shown in fig. 8, the electronic device may include: a processor (processor)810, a communication Interface 820, a memory 830 and a communication bus 840, wherein the processor 810, the communication Interface 820 and the memory 830 communicate with each other via the communication bus 840. Processor 810 may invoke logic instructions in memory 830 to perform a method of building an auto-answer model, the method comprising: performing feature extraction on target reply information to generate first identification information corresponding to the target reply information, wherein the target reply information is client reply information under the condition of machine marketing failure; determining at least one target marketing tactical message from candidate manual marketing tactical messages corresponding to candidate customer reply messages on the basis of the first identification message, wherein the candidate customer reply messages are all customer reply messages under the condition that manual marketing is successful; or, an automatic response method is performed, the method comprising: acquiring reply information of a client to be responded; performing feature extraction on the reply information of the client to be responded to generate target identification information corresponding to the reply information of the client to be responded; the target identification information is input to the automatic response model generated by the method for constructing an automatic response model as described above, and target response information is generated.

In addition, the logic instructions in the memory 830 may be implemented in software functional units and stored in a computer readable storage medium when the logic instructions are sold or used as independent products. Based on such understanding, the technical solution of the present invention may be embodied in the form of a software product, which is stored in a storage medium and includes instructions for causing a computer device (which may be a personal computer, a server, or a network device) to execute all or part of the steps of the method according to the embodiments of the present invention. And the aforementioned storage medium includes: a U-disk, a removable hard disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), a magnetic disk or an optical disk, and other various media capable of storing program codes.

In another aspect, the present invention also provides a computer program product comprising a computer program stored on a non-transitory computer-readable storage medium, the computer program comprising program instructions which, when executed by a computer, enable the computer to perform the method for constructing an automatic response model provided by the above methods, the method comprising: performing feature extraction on target reply information to generate first identification information corresponding to the target reply information, wherein the target reply information is client reply information under the condition of machine marketing failure; determining at least one target marketing tactical message from candidate manual marketing tactical messages corresponding to candidate customer reply messages on the basis of the first identification message, wherein the candidate customer reply messages are all customer reply messages under the condition that manual marketing is successful; or, an automatic response method is performed, the method comprising: acquiring reply information of a client to be responded; performing feature extraction on the reply information of the client to be responded to generate target identification information corresponding to the reply information of the client to be responded; the target identification information is input to the automatic response model generated by the method for constructing an automatic response model as described above, and target response information is generated.

In yet another aspect, the present invention also provides a non-transitory computer-readable storage medium having stored thereon a computer program, which when executed by a processor, implements a method for constructing the automatic response model provided above, the method including: performing feature extraction on target reply information to generate first identification information corresponding to the target reply information, wherein the target reply information is client reply information under the condition of machine marketing failure; determining at least one target marketing tactical message from candidate manual marketing tactical messages corresponding to candidate customer reply messages on the basis of the first identification message, wherein the candidate customer reply messages are all customer reply messages under the condition that manual marketing is successful; or, an automatic response method is performed, the method comprising: acquiring reply information of a client to be responded; performing feature extraction on the reply information of the client to be responded to generate target identification information corresponding to the reply information of the client to be responded; the target identification information is input to the automatic response model generated by the method for constructing an automatic response model as described above, and target response information is generated.

The above-described embodiments of the apparatus are merely illustrative, and the units described as separate parts may or may not be physically separate, and parts displayed as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of the present embodiment. One of ordinary skill in the art can understand and implement it without inventive effort.

Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented by software plus a necessary general hardware platform, and certainly can also be implemented by hardware. With this understanding in mind, the above-described technical solutions may be embodied in the form of a software product, which can be stored in a computer-readable storage medium such as ROM/RAM, magnetic disk, optical disk, etc., and includes instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute the methods described in the embodiments or some parts of the embodiments.

Finally, it should be noted that: the above examples are only intended to illustrate the technical solution of the present invention, but not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some technical features may be equivalently replaced; and such modifications or substitutions do not depart from the spirit and scope of the corresponding technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing an automatic response model is characterized by comprising the following steps:

2. The method for constructing an automated response model according to claim 1, wherein the determining at least one target marketing communication information from the candidate artificial marketing communication information corresponding to the candidate customer reply information based on the first identification information comprises:

3. The method for constructing the automated response model according to claim 2, wherein the matching of the candidate manual marketing tactics information to obtain at least one target marketing tactics information corresponding to the second identification information based on the set of the first customer response information and the above machine marketing tactics information corresponding to the first customer response information comprises:

4. The method of constructing an automatic response model according to claim 3,

the determining first similarity information between the first client reply message and each of the candidate client reply messages includes:

and/or the presence of a gas in the gas,

5. The method for constructing an automatic response model according to any one of claims 1 to 4, wherein the performing feature extraction on the target reply message to generate the first identification information corresponding to the target reply message includes:

6. The method for constructing an automated response model according to any one of claims 1 to 4, wherein after determining at least one target marketing communication information from the candidate artificial marketing communication information corresponding to the candidate customer response information based on the first identification information, the method further comprises:

7. An automatic answering method, comprising:

acquiring reply information of a client to be responded;

inputting the target identification information into an automatic response model generated by the method of constructing an automatic response model according to any one of claims 1 to 5, generating target response information.

8. An apparatus for constructing an automatic response model, comprising:

9. An automatic answering device, comprising:

a third generating module, configured to input the target identification information into the automatic response model generated by the method for constructing an automatic response model according to any one of claims 1 to 5, and generate target response information.

10. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the steps of the method of constructing an auto-answer model according to any one of claims 1 to 6 or the steps of the auto-answer method according to claim 7 when executing the program.

11. A non-transitory computer-readable storage medium, on which a computer program is stored, which, when being executed by a processor, carries out the steps of the method of constructing an automatic response model according to any one of claims 1 to 6 or the steps of the automatic response method according to claim 7.

12. A computer program product comprising a computer program, characterized in that the computer program, when being executed by a processor, carries out the steps of the method of constructing an automatic response model according to any one of claims 1 to 6 or the steps of the automatic response method according to claim 7.