TWM641534U - Intelligent marketing system based on natural language processing - Google Patents

Abstract

An intelligent marketing system includes: a processing server, which uses natural language processing technology to segment the text information about customer responses from the digital customer group interaction platform to obtain vocabulary related to the customer and uses the emotional score lookup table to obtain relevant information. The emotional score of the vocabulary; and the marketing server, when it is determined that the emotional score is not less than zero, it uses the pre-built marketing event preference probability estimation model and obtains relevant information about the specific event from the tag database according to the basic information of the customer The tag data, the sentiment score, and attribute data about multiple marketing events are used to obtain estimation results. The estimation result includes a plurality of preference probabilities respectively corresponding to the marketing events, and according to the estimation result, personalized marketing recommendation information corresponding to the customer is generated.

Description

Intelligent marketing system based on natural language processing

The present invention relates to product marketing, in particular to an intelligent marketing system based on natural language processing.

At present, enterprise organizations usually provide customers with a digital customer service platform for digital customer service. By further analyzing the service information obtained by the digital customer service, such as customer voice information or customer text information, the emotional state of the customer when using the service can be determined, and the service quality evaluation results of the relevant digital customer service can be obtained based on the determined emotional state. However, such service quality evaluation results can only be used as a reference for enterprises to improve the quality of digital customer service.

Therefore, how to come up with an intelligent marketing method that provides personalized marketing service information by using the comment information obtained from the digital customer service platform has become one of the issues to be solved in the related technical field.

Therefore, the purpose of the present invention is to provide an intelligent marketing system based on natural language processing, which can overcome at least one shortcoming of the prior art.

Therefore, an intelligent marketing system based on natural language processing provided by the present invention includes a processing server and a marketing server.

The processing server is used to connect to a digital customer group interaction platform, and stores a look-up table of sentiment scores related to a plurality of reference words.

The marketing server is connected to the processing server, and stores a pre-established marketing event preference probability estimation model related to a plurality of marketing events, and a label database. The tag database includes a plurality of tags respectively representing a plurality of events, the corresponding tag contents and corresponding tag representative values, and the events include the marketing events.

When the processing server receives text information from the digital customer group interaction platform about a customer's response to an interaction with a specific object related to at least one of the events, the processing server utilizes natural language processing technology , perform word segmentation processing on the text information to obtain one or more vocabulary related to the customer, and use the sentiment score lookup table to obtain a sentiment score related to the vocabulary(s), and at least send the sentiment score to The marketing server.

When the marketing server determines that the sentiment score from the processing server is not less than zero, it uses the marketing event preference probability estimation model to obtain information about the customer from the tag database at least based on the basic information about the customer. label data for at least one specific event, the sentiment score, and attribute data about the marketing events to obtain an estimation result including a plurality of preference probabilities respectively corresponding to the marketing events, and based on the estimation Based on the test results, personalized marketing recommendation information corresponding to the customer is generated.

In some embodiments, the marketing server is also used to connect to the digital customer group interaction platform, and transmit the personalized marketing recommendation information to the digital customer group interaction platform in real time, so that it can be provided to the customer in real time.

In some embodiments, the marketing server also pre-stores the attribute data. The marketing event preference probability estimation model is a hierarchical Bayesian logic model, and the hierarchical Bayesian logic model includes a preference structure hierarchy and a preference probability hierarchy. The marketing server first uses the Markov Chain Monte Carlo (MCMC) algorithm to iteratively calculate the basic data, the label data, and the sentiment score in the preference structure level to estimate a plurality of respective correspondences about the customer Model preference parameters based on a plurality of event attributes, and then calculate the estimation result in the preference probability level according to the model preference parameters and the attribute data.

In some embodiments, the specific object includes a specific article provided by the digital customer interaction platform.

In some embodiments, each of the marketing events is a marketing service or a marketing product.

In some embodiments, the marketing server also stores historical marketing event data related to the customer, and the marketing server also adjusts or retrains the marketing event preference probability estimation model according to the historical marketing event data.

In some embodiments, the marketing server also stores marketing information related to the marketing events, and from the marketing information, the marketing server obtains information about the marketing events having a relatively higher value than the estimation result. Target marketing information of at least one target marketing event corresponding to at least one preference probability is used as the personalized marketing recommendation information.

The function of this new model is: by analyzing the interactive information generated during the interaction between the business side and the customer side on the digital customer group interaction platform, the customer’s evaluation of the product or service can be obtained; especially, the use of pre-built marketing events The preference probability estimation model and the estimation result are obtained according to the emotional score obtained from the customer's message or comment content contained in the interactive information, the basic information of the customer and the label information of the specific marketing event. According to the estimation result, corresponding to the customer Personalized marketing recommendation information, and provide customers with personalized marketing recommendation information based on the estimation results in real time through the digital customer group interaction platform, so as to achieve the purpose of personalized marketing, and then improve digital marketing recommended products or service success rate.

Before the present invention is described in detail, it should be noted that in the following description, similar elements are denoted by the same numerals.

Referring to FIG. 1 , an intelligent marketing system 100 according to an embodiment of the present invention is exemplarily shown, which can be applied to corporate institutions such as but not limited to banks. In this embodiment, the intelligent marketing system 100 can be used together with a digital customer group interaction platform 200 and a customer service server 400 provided by the banking institution. The digital customer group interaction platform 200 can provide various objects related to the financial products and financial services currently marketed by the banking enterprise, for example, news articles, product information and service information, for connecting customers (which can be the banking enterprise) Customers owned by the institution, or customers not owned by the bank's corporate institution), and can collect or record information from the user devices (such as smart phones, tablets, personal computers) used by the connected customers etc.) and interactive information (such as message content of connected customers). The client server 400 is used for performing instant response processing of online customer service. The intelligent marketing system 100 includes a processing server 1 and a marketing server 2, which can constitute a computer system.

The processing server 1 can be connected to the digital customer group interaction platform 200 and the customer service server 400 via a communication network (not shown), and store a look-up table of emotional scores related to a plurality of reference words, which can be determined according to the actual situation Additions. The following Table 1 can exemplarily illustrate the content contained in the emotion score lookup table, but not limited thereto. Table 1 reference vocabulary sentiment score success 1 fall into 0 profit 1 become 2 Exceed 0 beware -1 rebound -2 profit -1 superior to 0 excitation 1 question 0 shock -1 reduce -2 obvious 0

The customer service server 400 can be connected to the digital customer group interaction platform 200 and to the processing server 1 through the communication network. In this embodiment, the marketing server 2 stores a pre-established marketing event preference probability estimation model, a label database, attribute data and a marketing information data related to multiple current marketing events of the banking organization. library. In this embodiment, each marketing event can be, for example, a financial product (for example, US dollar fixed deposit, fund, credit, insurance, etc.) or a financial service (for example, the use of mobile payment, the update of personal basic information, credit card swiping installment application, etc.).

The tag library contains multiple tags representing multiple events and their associated tag content. It is worth noting that these events include the current marketing events. More specifically, each marketing event can be represented by a corresponding tag, and the tag database also includes, for each tag, a corresponding tag representative value predefined according to the relevant tag content. In addition, in this embodiment, each object provided by the digital customer group interaction platform 200 may be related to one or more events (that is, may be related to one or more tags).

行銷類別 例行性行銷 1 非例行性行銷 0 專案類別 商品_排除黑名單 1 通知_不排除黑名單 0 The attribute data includes, for example, a plurality of event attributes related to the marketing events, their corresponding attribute contents, and corresponding attribute representative values. To give a simple example, the attribute data related to these marketing events such as "event category", "marketing category", and "project category" can be shown in Table 2 below. Table 2 event properties attribute content attribute representative value event category Credit card_promote swiping 1 Credit card_do card 0

Marketing category routine marketing 1 non-routine marketing 0 Project category Commodity_exclusion blacklist 1 Notification_Do not exclude blacklist 0

式1 其中

為客戶 i之第 k個變數，其例如可為客戶 i之人口統計變量(demographics，例如包含但不限於性別、年齡、婚姻狀態、學歷等)、與平台/文章/行銷事件有關的標籤、情緒分數等其中一者；

代表迴歸係數；

代表殘差值；及

,

,…代表有關於客戶 i的多個分別對應於多個事件屬性的模型偏好參數。在該偏好機率層級中，該行銷伺服器2根據在該偏好結構層級所估算出的多個模型偏好參數和該屬性資料計算出有關於該輸入資料的估測結果。更具體地，在該偏好機率層級中的演算模型例如可以下式2來表示：

式2 其中

代表客戶 i對於J個行銷事件中的第j個行銷事件的偏好機率；及

為擬推薦客戶 i之第j個行銷事件之屬性資料。 The marketing event preference probability estimation model is, for example, a Hierarchical Bayesian Model, which includes a preference structure hierarchy and a preference probability hierarchy. In the preference structure level, the marketing server 2, for example, uses a Markov Chain Monte Carlo (MCMC) algorithm to perform multiple iterative calculations on the input data to estimate values corresponding to multiple event attributes. Multiple model preference parameters. In this embodiment, the input data may include basic information about the customer (such as but not limited to gender, age, marital status, education, etc.), one or more tags, and emotional scores (detailed below) . More specifically, the calculus model at the preference structure level can be represented by the following formula 1, for example:

Formula 1 where

It is the kth variable of customer i , which can be, for example, customer i ’s demographic variables (demographics, such as including but not limited to gender, age, marital status, education, etc.), tags related to platforms/articles/marketing events, emotions one of scores, etc.;

Represents the regression coefficient;

represents the residual value; and

,

,… represent multiple model preference parameters about customer i corresponding to multiple event attributes. In the preference probability level, the marketing server 2 calculates an estimation result related to the input data according to the plurality of model preference parameters estimated in the preference structure level and the attribute data. More specifically, the calculus model in the preference probability level can be represented by the following formula 2, for example:

Formula 2 where

represents the probability of customer i 's preference for the jth marketing event among the J marketing events; and

It is the attribute data of the jth marketing event of the customer i to be recommended.

The marketing information database stores marketing information related to the marketing events.

As shown in FIG. 1 , when a customer 500 browses a specific object provided by the digital customer group interaction platform 200 through a user device 300 (such as a smart phone) connected to the digital customer group interaction platform 200, the input interaction When replying text information, the specific object can be a report article or a publicity advertisement of a financial product/service and one or more specific events (which can be current marketing events or non-current marketing events) In this case, the digital customer group interaction platform 200 will, for example, transmit the basic information about the customer 500 , the text information and the object data indicating the specific event(s) to the processing server 1 . In this embodiment, the customer 500 is, for example but not limited to, a new customer of the banking institution, and in other embodiments, the customer may also be an old customer of the banking institution.

Hereinafter, referring to FIG. 1 and FIG. 2, when the processing server 1 receives the basic data, the text information and the object data from the digital customer group interaction platform 200, how does the intelligent marketing system 100 A smart marketing program for the client 500 is executed. The smart marketing program includes the following steps S21-S28.

First, in step S21 , the processing server 1 uses known natural language processing technology to segment the received text information to obtain one or more words related to the client 500 .

Then, in step S22, the processing server 1 utilizes the emotional score lookup table stored in it to find out the corresponding emotional score of each of the vocabulary(s), and use the sum of all the corresponding emotional scores found as A sentiment score representing the current emotional response of the customer 500.

Then, in step S23, the processing server 1 determines whether the emotion score is not less than zero. If the determination result is positive, the process will proceed to step S24, otherwise, the process will proceed to step S28.

When the processing pre-server 1 determines that the emotion score is not less than zero, in step S24, the processing server 1 sends the basic data, the object data and the emotion score to the marketing server 2.

After the marketing server 2 receives the basic data, the object data and the sentiment score, in step S25, the marketing server 2 can use the marketing event preference probability estimation model to obtain the The database obtains tag data related to the object data and the sentiment score, and obtains an estimation result corresponding to the customer 500 . The estimation result includes a plurality of preference probabilities respectively corresponding to the marketing events. More specifically, the marketing server 2 first utilizes the Markov chain Monte Carlo (MCMC) algorithm in the preference structure level for the basic data, the label data, the sentiment score, and attribute data about the marketing events An iterative calculation is performed to estimate a plurality of model preference parameters corresponding to a plurality of event attributes about the customer, and then the estimation result is calculated in the preference probability level according to the model preference parameters and the attribute data.

，

，

；及在該偏好結構層級的演算模型中例如具有7個變數，其分別為「性別」、「學歷1」、「學歷2」、「文章標籤1」、「文章標籤2」、「文章標籤3」及「情緒分數」。於是，上述式1可表示成以下式3：

式3 其中與「性別」、「學歷1」、「學歷2」等變數有關的代表值可事先定義如下表3，而與「文章標籤1」、「文章標籤2」、「文章標籤3」等變數有關的標籤代表值可事先定義如下表4：   表3 變數 內容 代表值 性別 男性 1 女性 0 學歷1 具有學士學歷 1 不具有學士學歷 0 學歷2 具有碩士學歷 1 不具有碩士學歷 0 表4 變數 標籤內容 標籤代表值 文章標籤1 有基金單筆標籤 1 無基金單筆標籤 0 文章標籤2 有保險標籤 1 無保險標籤 0 文章標籤3 有投資理財標籤 1 無投資理財標籤 0 此外，該客戶500例如為一具有碩士學歷的女性，並所瀏覽一特定物件為例如一有關於例如「保險」和「投資理財」等特定事件的報導文章且該處理伺服器1根據該客戶500的文字資訊所獲得的情緒分數為1。於是，該行銷伺服器2將自表3獲得對應於該客戶500的學歷和性別的代表值資料、自表4獲得對應於「保險」和「投資理財」標籤的標籤代表值資料，以及該情緒分數輸入到如上述式3的演算模型，並且經過例如600次的迭代演算後所估算出的三個模型偏好參數

，

，

如以下式4所示： [

,

,

] = [0.22, 0.85, 1.52]                            式4 另一方面，由於該等行銷事件M1，M2，M3具有3個事件屬性，因此上述式2可表示成以下式5：

式5 於是，根據上述表2可獲得有關於該等行銷事件M1，M2，M3的屬性代表值作為該屬性資料如以下表5所示：   表5 行銷事件 屬性內容 (事件類別，行銷類別，專案類別) 屬性資料(屬性代表值) M1 (信用卡_辦卡，例行性行銷，通知) (0，1，0) M2 (信用卡_促刷，例行性行銷，通知) (1，1，0) M3 (信用卡_促刷，非例行性行銷，商品) (1，0，1) 接著，該行銷伺服器2將上述式4的模型偏好參數

，

，

以及上述表5所示的屬性資料代入上述式5後可獲得對應於每個行銷事件M1/M2/M3的

以及偏好機率(即，

如下表6所示：   表6 行銷事件

偏好機率(

M1 13.33 0.09 M2 36.23 0.24 M3 98.49 0.67 Hereinafter, a simple example is used to describe in detail how the marketing server 2 uses the marketing event preference probability estimation model to obtain the estimation result in step S25. In this example, the banking enterprise currently launches, for example, 3 (J=3) marketing events (which are respectively the routine credit card application notice M1, the routine credit card promotion notice M2 and the non-routine credit card product Promote brushing M3); have for example 3 model preference parameters corresponding to 3 event attributes respectively

,

,

; and there are, for example, 7 variables in the calculus model of the preference structure level, which are respectively "sex", "education 1", "education 2", "article label 1", "article label 2", "article label 3 ” and “Sentiment Score”. Then, the above formula 1 can be expressed as the following formula 3:

In formula 3, the representative values related to variables such as "gender", "education 1" and "education 2" can be defined in advance in the following table 3, and related to "article label 1", "article label 2", "article label 3", etc. Variable-related label representative values can be defined in advance as shown in Table 4: Table 3 variable content representative value gender male 1 female 0 Education 1 have a bachelor's degree 1 without a bachelor's degree 0 Education 2 Have a master's degree 1 Without a master's degree 0 Table 4 variable label content Labels represent values Article Tag 1 There is a fund single label 1 No fund single label 0 Article Tag 2 with insurance label 1 no insurance label 0 Article Tag 3 Has investment and financial management label 1 No investment and financial management label 0 In addition, the client 500 is, for example, a woman with a master's degree, and the specific object browsed is, for example, a report article on specific events such as "insurance" and "investment and financial management" and the processing server 1 according to the client 500 The sentiment score obtained for text messages is 1. Therefore, the marketing server 2 will obtain the representative value data corresponding to the educational background and gender of the customer 500 from Table 3, the representative value data corresponding to the labels of "insurance" and "investment and financial management" from Table 4, and the emotion The score is input into the calculation model as in the above formula 3, and the three model preference parameters estimated after, for example, 600 iterative calculations

,

,

As shown in Equation 4 below: [

,

,

] = [0.22, 0.85, 1.52] Formula 4 On the other hand, since the marketing events M1, M2, M3 have 3 event attributes, the above formula 2 can be expressed as the following formula 5:

Equation 5 Therefore, according to the above table 2, the representative values of attributes related to the marketing events M1, M2, M3 can be obtained as the attribute data as shown in the following table 5: Table 5 marketing event Attribute content (event category, marketing category, project category) Attribute data (attribute representative value) M1 (credit card_applying card, routine marketing, notice) (0,1,0) M2 (credit card_swipe promotion, routine marketing, notification) (1,1,0) M3 (Credit card_ promotion, non-routine marketing, merchandise) (1,0,1) Next, the marketing server 2 uses the model preference parameter of the above formula 4

,

,

And after substituting the attribute data shown in the above Table 5 into the above formula 5, we can obtain the corresponding to each marketing event M1/M2/M3

and preference probabilities (ie,

Table 6 below: Table 6 marketing event

Preference probability (

M1 13.33 0.09 M2 36.23 0.24 M3 98.49 0.67

Afterwards, in step S26, the marketing server generates personalized marketing recommendation information for the customer 500 according to the estimation result. More specifically, in this embodiment, the marketing server 2 will select the marketing event corresponding to the highest preference probability contained in the estimation result as the target marketing event to be recommended to the client 500, and from the marketing information data Target marketing information related to the target marketing event is obtained from the database as the personalized marketing recommendation information.

Then, in step S27, the marketing server 2 transmits the personalized marketing recommendation information to the digital customer group interaction platform 200 in real time, so that it can be provided to the customer 500 in real time, thereby achieving for personalized marketing purposes.

On the other hand, when the processing server 1 determines that the emotional score is less than 0 in step S23, it means that the customer 500 may be in a negative evaluation state for the relevant information provided by the digital customer interaction platform. In step S28, The processing server 1 notifies the customer service server 400 of the customer 500's negative evaluation, so as to facilitate subsequent response processing.

So far, the intelligent marketing program executed for the client 500 is completed. It should be noted that, in other embodiments, if the customer is an old customer of the banking institution and the marketing server 2 also stores historical marketing event data related to the customer, the marketing server 2 can also The marketing event probability estimation model is adjusted or retrained according to the historical marketing event data.

To sum up, by analyzing the interactive information generated during the interaction between the business side and the customer side on the digital customer group interaction platform 200, the customer's evaluation of the product or service can be obtained; especially, by using the pre-built marketing event The preference probability estimation model and the estimation results are obtained according to the emotional scores obtained from the customer messages or comments contained in the interactive information, the basic information of the customers and the label information of specific marketing events, and according to the estimation results, corresponding to the customer The personalized marketing recommendation information of the customer is generated according to the estimation result, and the customer's personalized marketing recommendation information is provided to the customer in real time through the digital customer group interaction platform 200 . In particular, regardless of whether the customer is an old customer or a new customer, the intelligent marketing system 100 can achieve the purpose of personalized marketing, thereby increasing the success rate of digital marketing recommended products or services. Therefore, the intelligent marketing system 100 of the present invention based on natural language processing can indeed achieve the purpose of the present invention.

But the above-mentioned person is only the embodiment of the present invention, and should not limit the scope of implementation of the present invention with this, and all simple equivalent changes and modifications made according to the patent scope of the present application and the content of the patent specification are still within the scope of the present invention. Within the scope covered by this patent.

100: Intelligent marketing system 1: Handle the server 2: Marketing server 200: digital customer group interaction platform 300: user device 400: customer service server 500: customer S21~S28: Steps

Other features and functions of the present invention will be clearly presented in the implementation manner with reference to the drawings, wherein: FIG. A digital customer group interaction platform and a customer service server of the intelligent marketing system; and FIG. 2 is a flow chart illustrating how the embodiment executes an intelligent marketing program.

100: Intelligent marketing system

1: Handle the server

2: Marketing server

200: digital customer group interaction platform

300: user device

400: customer service server

500: customer

Claims (7)

An intelligent marketing system based on natural language processing, comprising: a processing server, used to connect to a digital customer group interaction platform, and store a sentiment score look-up table about multiple reference words; and a marketing server, connected to the Processing the server, and storing a pre-established marketing event preference probability estimation model related to a plurality of marketing events, and a tag database, the tag database includes a plurality of tags representing a plurality of events and corresponding tags Content and corresponding tag representative values, these events include these marketing events; Wherein, when the processing server receives from the digital customer group interaction platform and a customer responds to a specific event related to at least one of these events When the text information of the interactive reply of a specific object is used, the processing server uses natural language processing technology to segment the text information to obtain one or more words related to the customer, and uses the sentiment score lookup table to obtain a has a sentiment score for the word(s), and at least transmits the sentiment score to the marketing server; and wherein the marketing server utilizes the sentiment score when determining that the sentiment score from the processing server is not less than zero A marketing event preference probability estimation model obtained at least based on the basic information about the customer, the tag data related to the at least one specific event obtained from the tag database, the sentiment score, and attribute data about the marketing events An estimation result, the estimation result includes a plurality of preference probabilities respectively corresponding to the marketing events, and according to the estimation result, personalized marketing recommendation information corresponding to the customer is generated. The intelligent marketing system based on natural language processing as described in claim 1, wherein the marketing server is also used to connect to the digital customer group interaction platform, and transmit the personalized marketing recommendation information to the digital customer group interaction in real time platform for it to provide immediately to that customer. The intelligent marketing system based on natural language processing as described in claim 1, wherein: the marketing server also pre-stores the attribute data; the marketing event preference probability estimation model is a hierarchical Bayesian logic model, and the hierarchical Bayesian The logic model includes a preference structure level and a preference probability level; and the marketing server first uses the Markov Chain Monte Carlo (MCMC) algorithm in the preference structure level to perform the basic data, the label data and the emotional score The iterative calculation is performed to estimate a plurality of model preference parameters corresponding to a plurality of event attributes about the customer, and then the estimation result is calculated in the preference probability level according to the model preference parameters and the attribute data. The intelligent marketing system based on natural language processing as described in Claim 1, wherein the specific object includes a specific article provided by the digital customer group interaction platform. The intelligent marketing system based on natural language processing as described in Claim 1, wherein each of the marketing events is a marketing service or a marketing commodity. The intelligent marketing system based on natural language processing as described in claim 1, wherein: the marketing server also stores historical marketing event data related to the customer; and the marketing server also adjusts or The marketing event preference probability estimation model is trained again. The intelligent marketing system based on natural language processing as described in claim 1, wherein: the marketing server also stores marketing information related to the marketing events; and the marketing server obtains information about the marketing from the marketing information Target marketing information of at least one target marketing event corresponding to at least one target marketing event with a relatively high preference probability in the estimation result is used as the personalized marketing recommendation information.

Images