Information technology consultation service management method and system

Abstract

The invention relates to the field of natural language processing, in particular to an information technology consultation service management method and system, and the method comprises the steps: enabling consultation voices of salesmen and customers, commodity-related words and sales-related words to generate intention keywords and purchase tendency degrees of the customers through a machine learning-based NLP voice matching model; intercepting a voice segment containing the intention keyword in the consultation voice; extracting the voice intensity and the pitch of the voice of the customer in the voice segment, and converting the voice intensity and the pitch to calculate the intention emotion titer of the customer; the purchase intention level of the customer is calculated through a logistic regression model according to the purchase tendency degree and the intention emotion titer, and the purchase intention level is used for classified management of customer information and judging whether product introduction information is sent to equipment of the customer or not. According to the invention, the cost of classified management of customer intentions by manually reviewing the consultation voice is reduced, and the consultation information processing efficiency is greatly improved.

Description

Technical Field

[0001] The present invention relates to the field of natural language processing technology, and in particular to an information technology consulting service management method and system. Background Art

[0002] The consulting platform is applied to various fields. The user side puts forward consulting needs through voice, and the server side responds to the consulting needs and gives corresponding feedback. At present, with the development of computer technology and Internet technology, it is possible to use NLP natural language processing technology to automatically provide feedback on consulting needs or analyze relevant information such as customer intentions, and provide effective data support for subsequent information push, building user portraits, etc.

[0003] For example, the invention patent with application number 202311337518.X provides a speech recognition method and NLP platform based on machine learning, and discloses the extraction of key information through speech recognition through a machine learning model. However, the extracted key information has a low correlation with the goods being sold, and it is impossible to judge the emotional attitude information contained in the customer's voice.

[0004] Therefore, how to enable the telephone sales system to generate words related to goods and sales through the machine learning model of NLP technology, judge the emotional attitude information contained in the customer's voice, and then comprehensively judge the customer's purchase intention is a technical problem that needs to be solved at present. Summary of the invention

[0005] To this end, the present invention provides an information technology consulting service management method and system, which generates words related to commodities and sales through the machine learning model of NLP technology, judges the emotional attitude information contained in the customer's voice, and then comprehensively judges the customer's purchasing intention, thereby reducing the cost of manually reviewing the consulting voice to classify and manage customer intentions, and greatly improving the efficiency of telephone sales.

[0006] To achieve the above object, the present invention proposes an information technology consulting service management method, comprising:

[0007] The consultation voices of sales staff and customers are matched with product-related words and sales-related words through the NLP voice matching model based on machine learning to generate customers' intended keywords and purchase propensity;

[0008] intercepting a voice segment of the consultation voice containing the intended keyword;

[0009] Extracting the intensity and pitch of the customer's voice in the voice segment, converting the intensity and pitch to calculate the customer's intended emotional valence;

[0010] The purchase propensity and the intention emotional valence are used to calculate the customer's purchase intention level through a logistic regression model. The purchase intention level is used for classification management of customer information and for determining whether to send product introduction information to the customer's device.

[0011] Further, the NLP speech matching model includes a corpus sub-model and a semantic matching sub-model;

[0012] The process of generating the customer's intended keywords and purchase propensity by matching the consultation voice of the salesperson with product-related words and sales-related words through the NLP voice matching model based on machine learning is as follows:

[0013] Acquire the product-related words and the sales-related words on the open source shopping website through crawling technology, and use them to construct the corpus sub-model reflecting the relationship between products and words;

[0014] The consultation speech is preprocessed by data conversion to generate a word list, and the word list is used to find multiple candidate customer intention words in the corpus sub-model through a search engine;

[0015] Calculate the final matching degree of the plurality of candidate customer intention words and the word list through the semantic matching sub-model;

[0016] The candidate customer intention words whose final matching degree exceeds a final matching degree threshold are determined as the intention keywords, and the purchase tendency is calculated according to the intention keywords and their corresponding final matching degrees.

[0017] Furthermore, the semantic matching sub-model is an improved long short-term memory network model that combines a bidirectional long short-term memory network and an attention fusion mechanism;

[0018] The final matching degree is calculated by using the semantic matching sub-model for the plurality of candidate customer intention words and the word list, specifically:

[0019] The semantic matching sub-model generates a hypothesis word vector from the candidate customer intention word and generates a premise word vector from the word list;

[0020] Calculating the hypothesis hidden state of the hypothesis word vector and the premise hidden state of the premise word vector through the bidirectional long short-term memory network of the semantic matching sub-model;

[0021] The hypothesis hidden state and the premise hidden state are respectively passed through the time recursive long short-term memory network of the semantic matching sub-model to calculate a network output hypothesis vector and a network output premise vector;

[0022] The network output hypothesis vector and the network output premise vector are respectively calculated by the average pooling layer and the maximum pooling layer of the semantic matching sub-model to obtain a hypothesis average pooling vector, a hypothesis maximum pooling vector, a premise average pooling vector and a premise maximum pooling vector, and the basic matching degree is calculated by cosine similarity;

[0023] The basic matching degree, the assumed hidden state, the premise hidden state, the network output hypothesis vector and the network output premise vector are subjected to the attention fusion mechanism of the semantic matching sub-model to generate the final matching degree.

[0024] Furthermore, the process of calculating the basic matching degree of the assumed average pooling vector, the assumed maximum pooling vector, the premise average pooling vector and the premise maximum pooling vector through the cosine similarity algorithm is as follows:

[0025] After the assumed average pooling vector and the premise maximum pooling vector are normalized, a first matching degree is calculated using a cosine similarity formula;

[0026] After the premise average pooling vector and the assumed maximum pooling vector are normalized, a second matching degree is calculated using a cosine similarity formula;

[0027] Calculate a third matching degree between the assumed average pooling vector and the premise average pooling vector by using a cosine similarity formula;

[0028] Calculate a fourth matching degree between the assumed maximum pooling vector and the premise maximum pooling vector by using a cosine similarity formula;

[0029] The basic matching degree is calculated by weighted summing up the first matching degree, the second matching degree, the third matching degree and the fourth matching degree, wherein the weight value of the weighted summing up is calculated and determined according to the number of the candidate customer intention words and the length of the word list.

[0030] Furthermore, the process of generating the final matching degree by the attention fusion mechanism of the semantic matching sub-model through the basic matching degree, the assumed hidden state, the premise hidden state, the network output assumption vector and the network output premise vector is:

[0031] The assumed hidden state and the premise hidden state are weighted to calculate a first difference through an attention fusion mechanism;

[0032] Calculate a second difference by multiplying the first difference by the network output hypothesis vector and the network output premise vector;

[0033] Generate a hidden layer state through an attention fusion mechanism according to the second difference, the assumed hidden state and the premise hidden state;

[0034] The final matching degree is calculated according to the hidden layer state, the basic matching degree and the second difference degree.

[0035] Furthermore, the process of generating a word list by preprocessing the consultation speech through data conversion is as follows:

[0036] The consultation speech is converted into a call text by a speech recognition model based on a gated recurrent unit;

[0037] The call text is segmented and stop words are removed by a segmentation tool to generate the word list. In the above scheme, the NLP voice matching model is implemented to enhance the relevant information of the hypothesis words and premise words through the attention mechanism, thereby enhancing the final matching degree.

[0038] Furthermore, the intensity and pitch of the customer's voice in the voice segment are extracted, and after the intensity and pitch are converted, the process for calculating the customer's intended emotional valence is as follows:

[0039] The standard deviation of the sound intensity of multiple customer voice segments of the first set time length is used as the sound intensity emotional arousal value;

[0040] Performing Fourier transform on multiple customer voice segments of the second set time length to obtain multiple pitch signals, and using the standard deviation of the multiple pitch signals as the pitch attitude positivity;

[0041] The weight coefficient is calculated according to the frequency of occurrence of the intended keywords in the customer's voice and the speaking speed of the customer's voice, and the intended emotional valence is generated by weighted summation of the weight coefficient, the sound intensity emotional arousal value and the pitch attitude positivity.

[0042] Furthermore, the process of calculating the customer's purchase intention level by using the purchase propensity and the intention emotional valence through a logistic regression model, and determining whether to send product introduction information to the customer's device according to the purchase intention level is as follows:

[0043] Generate a multi-classification logistic regression model based on the customer's gender information, age information and their corresponding sample data;

[0044] The purchase propensity and the intention emotional valence are substituted into the logistic regression model corresponding to the gender information and age information. The input data classification probability generated by the logistic regression model corresponds to the grade gradient, and the purchase intention grade is determined.

[0045] In the above scheme, the logistic regression model is used to judge users of different genders and ages separately, making the purchase intention level combined with the emotional valence of the voice and the content information more accurate.

[0046] The present invention also provides a system for applying the information technology consulting service management method, including an NLP voice matching model, a voice processing module and a logistic regression model installed in a cloud call center, wherein:

[0047] The NLP voice matching model is used to calculate the consulting voice with product-related words and sales-related words to generate the customer's intended keywords and purchase propensity;

[0048] The speech processing module is used to extract the sound intensity and pitch of the speech segment containing the intended keyword in the consultation speech, and convert the sound intensity and the pitch to calculate the intended emotional valence of the customer;

[0049] The logistic regression model is used to calculate the customer's purchase intention level based on the purchase propensity and the intention emotional valence. The purchase intention level is used for classification management of customer information and for determining whether to send product introduction information to the customer's device.

[0050] Furthermore, the cloud call center performs classified management and secure access to customer information through smart contracts on the blockchain.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] 1. The machine learning model of NLP technology generates words related to products and sales, judges the emotional attitude information contained in the customer's voice, and then comprehensively judges the customer's purchase intention, thereby reducing the cost of manually reviewing the consultation voice to classify and manage customer intentions, greatly improving the efficiency of telephone sales.

[0053] 2. The NLP speech matching model is implemented through the attention mechanism to enhance the relevant information of the hypothesis words and premise words and thus enhance their final matching degree.

[0054] 3. Through the logistic regression model, users of different genders and ages are judged separately, making the purchase intention level combined with the emotional valence of the voice and the content information more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 A schematic diagram of the general flow of the information technology consulting service management method according to an embodiment of the present invention;

[0056] Figure 2 A schematic diagram of the NLP speech matching model flow of the information technology consulting service management method according to an embodiment of the present invention;

[0057] Figure 3 A schematic diagram of the general structure of the NLP speech matching model of the information technology consulting service management method according to an embodiment of the present invention;

[0058] Figure 4 Schematic diagram of the structure of a system for automatically identifying customer intentions based on NLP according to an embodiment of the present invention. DETAILED DESCRIPTION

[0059] In order to make the objects and advantages of the present invention more clearly understood, the present invention is further described below in conjunction with embodiments; it should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0060] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood by those skilled in the art that these embodiments are only used to explain the technical principles of the present invention and are not intended to limit the protection scope of the present invention.

[0061] It should be noted that, in the description of the present invention, terms such as "up", "down", "left", "right", "inside" and "outside" indicating directions or positional relationships are based on the directions or positional relationships shown in the drawings. This is merely for the convenience of description and does not indicate or imply that the device or element must have a specific orientation, be constructed and operated in a specific orientation. Therefore, it cannot be understood as a limitation on the present invention.

[0062] In addition, it should be noted that in the description of the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For those skilled in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0063] like Figures 1 to 4 As shown, the present invention provides an information technology consulting service management method and system, which generates words related to commodities and sales through the machine learning model of NLP technology, judges the emotional attitude information contained in the customer's voice, and then comprehensively judges the customer's purchase intention, thereby reducing the cost of manually reviewing the consulting voice to classify and manage customer intentions, and greatly improving the efficiency of telephone sales.

[0064] like Figures 1 to 4 As shown, this embodiment proposes an information technology consulting service management method and system, the system applies the information technology consulting service management method, the information technology consulting service management method includes:

[0065] The consulting voice between the salesperson and the customer is matched with product-related words and sales-related words through an NLP voice matching model based on machine learning to generate the customer's intended keywords and purchase propensity; a voice segment containing the intended keywords in the consulting voice is intercepted; the sound intensity and pitch of the customer's voice in the voice segment are extracted, and the sound intensity and the pitch are converted and processed to calculate the customer's intended emotional valence; the purchase propensity and the intended emotional valence are used to calculate the customer's purchase intention level through a logistic regression model, and the purchase intention level is used for classification management of customer information and to determine whether to send product introduction information to the customer's device.

[0066] It is understandable that the intended emotional valence is often used in psychology to describe the degree of pleasure of consumers' emotional experience, which to a certain extent reflects the customer's views and provides motivation for their purchasing behavior. The relevant research and principles of emotion in management are applied by scientifically and objectively judging the intended emotional valence through sound intensity and pitch.

[0067] Furthermore, the NLP voice matching model includes a corpus sub-model and a semantic matching sub-model; the consulting voices of the salesperson and the customer are matched with product-related words and sales-related words through the NLP voice matching model based on machine learning to generate the customer's intended keywords and purchase propensity, specifically:

[0068] The product-related words and the sales-related words are obtained on the open source shopping website through crawling technology, and are used to construct the corpus sub-model that reflects the relationship between products and words; the consulting voice is subjected to data conversion preprocessing to generate a word list, and the word list is used to find multiple candidate customer intention words in the corpus sub-model through a search engine; the multiple candidate customer intention words and the word list are used to calculate the final matching degree through the semantic matching sub-model; the candidate customer intention words whose final matching degree exceeds the final matching degree threshold are determined as the intended keywords, and the purchase propensity is calculated based on the intended keywords and their corresponding final matching degrees.

[0069] Specifically, the crawling technology, such as web crawling, automatically visits open source shopping websites according to set rules by writing programs or scripts and extracts product-related words and sales-related words therein. For example, when the product is a tourist product, the product-related words include transportation, sightseeing, accommodation, catering, shopping, attractions, scenic spots, etc., and the sales-related words include free travel, customized travel, promotion, direct sales, travel agency, off-season, peak season, and flat season.

[0070] Specifically, the amount of data in the corpus sub-model is large, so the MySQL database is used to store the data, and the Django Model ls class is used to build the database model. Preferably, the Model ls class has four data members, namely, int type commodity category, varchar type target word, varchar type commodity related word and varchar type sales related word.

[0071] Furthermore, the consultation voice is subjected to data conversion preprocessing to generate a word list, specifically: the consultation voice is converted into call text through a speech recognition model based on a gated recurrent unit; the call text is segmented and stop words are removed through a word segmentation tool to generate the word list.

[0072] Specifically, in order to obtain words related to the words in the word list, and also to shorten the search running time and improve the logical relevance of the search results, a lightweight search engine is used to obtain alternative customer intended words, and then the answer is made more accurate through the semantic matching sub-model. Preferably, the search engine is implemented by setting relevant parameters of the third-party Python library Whoosh.

[0073] Furthermore, the semantic matching sub-model is an improved long short-term memory network model that combines a bidirectional long short-term memory network and an attention fusion mechanism.

[0074] It can be understood that the Long Short-Term Memory (LSTM) network model is a special recurrent neural network (RNN) that aims to solve the problem of gradient vanishing or gradient exploding when traditional RNN processes long sequence data. The conventional Long Short-Term Memory network model includes an input encoding unit, a decoding reasoning unit, and an reasoning synthesis unit. By introducing the attention fusion mechanism, the model's ability to capture important information can be further enhanced, thereby improving the model's performance and accuracy.

[0075] The final matching degree is calculated by using the semantic matching sub-model for the plurality of candidate customer intention words and the word list, specifically:

[0076] The semantic matching sub-model generates a hypothesis word vector from the candidate customer intention word and generates a premise word vector from the word list; and calculates a hypothesis hidden state of the hypothesis word vector and a premise hidden state of the premise word vector through a bidirectional long short-term memory network of the semantic matching sub-model;

[0077] Specifically, the words in the word list are used to generate elements in the premise word vector, for example, premise word vector a = (a1, a2, ..., a n) represents the premise word vector containing words in multiple word lists, premise word vector b=(b1,b2,…,b n ) represents a hypothetical word vector containing multiple candidate customer intention words, which can be expressed as follows after passing through a bidirectional long short-term memory network:

[0078]

[0079] In the formula, k represents the current epoch value, a and b represent the premise word vector and hypothesis word vector, a k ,b k They represent the premise hidden state and the hypothesis hidden state respectively.

[0080] Specifically, the words in the word list are generated into elements of the premise word vector through the term frequency - inverse document frequency (TF-IDF): the corpus sub-model is generated through the complete consultation speech of this call. The corpus sub-model contains multiple word lists. The frequency of occurrence of a word in the word list in which it is located is taken as the word frequency, the frequency of occurrence of the word in the premise word vector is taken as the inverse document frequency, and the product of the word frequency and the inverse document frequency is taken as the word frequency - inverse document frequency, thereby generating the premise word vector.

[0081] TF-IDF is used to evaluate the importance of a word for one of the word lists in the corpus. It combines word frequency (TF) and inverse document frequency (IDF) to suppress the influence of common words and highlight important words. Similarly, the frequency of occurrence of a single candidate customer intent word in the candidate customer intent word as a whole in the corpus sub-model is used as the inverse document frequency, and the product of the word frequency and the inverse document frequency is used as its word frequency-inverse document frequency to generate a hypothetical word vector.

[0082] The hypothesis hidden state and the premise hidden state are respectively passed through the time recursive long short-term memory network of the semantic matching sub-model to calculate the network output hypothesis vector and the network output premise vector; specifically, the time recursive long short-term memory network can be expressed as:

[0083]

[0084] In the formula, k represents the current epoch value, Represents two bias parameters, a t ,b t Represent the network output premise vector and network output hypothesis vector respectively, They represent the premise hidden state and the hypothesis hidden state respectively.

[0085] The network output hypothesis vector and the network output premise vector are respectively calculated through the average pooling layer and the maximum pooling layer of the semantic matching sub-model to obtain the hypothesis average pooling vector, the hypothesis maximum pooling vector, the premise average pooling vector and the premise maximum pooling vector, and the basic matching degree is calculated through cosine similarity; specifically, the average pooling layer and the maximum pooling layer are conventional pooling processes and are not repeated here.

[0086] The basic matching degree, the assumed hidden state, the premise hidden state, the network output hypothesis vector and the network output premise vector are subjected to the attention fusion mechanism of the semantic matching sub-model to generate the final matching degree.

[0087] Furthermore, the basic matching degree of the assumed average pooling vector, the assumed maximum pooling vector, the premise average pooling vector and the premise maximum pooling vector is calculated by the cosine similarity algorithm, which is specifically:

[0088] After the assumed average pooling vector and the premise maximum pooling vector are normalized, a first matching degree is calculated using a cosine similarity formula;

[0089] After the premise average pooling vector and the assumed maximum pooling vector are normalized, a second matching degree is calculated using a cosine similarity formula;

[0090] Calculate a third matching degree between the assumed average pooling vector and the premise average pooling vector by using a cosine similarity formula;

[0091] Calculate a fourth matching degree between the assumed maximum pooling vector and the premise maximum pooling vector by using a cosine similarity formula;

[0092] The first matching degree, the second matching degree, the third matching degree and the fourth matching degree are weighted and summed to calculate the basic matching degree, wherein the weight value of the weighted sum is calculated and determined according to the number of the candidate customer intention words and the length of the word list, and the specific process is as follows:

[0093] v=k(v cos1 +v cos2 )+l(v cos3 +v cos4 )

[0094] In the formula, v is the basic matching degree, v cos1 ,v cos2 ,v cos3 ,v cos4They are the first matching degree, the second matching degree, the third matching degree and the fourth matching degree respectively, k and l are two weight values, if the number of candidate customer intended words is 0.2 to 0.8 times the length of the word list / the number of words, the corresponding k is 0.3 to 0.5, and the corresponding l is 0.7 to 0.5, for example, if the multiple is 0.3, then k = (0.3 / 0.6) 0.2 + 0.3 = 0.4, l = 1-k = 0.6. It can be understood that since the word list does not exclude repeated words, the length of the word list is longer than the number of candidate customer intended words.

[0095] It can be understood that the above-mentioned normalization process is achieved by dividing each element by the vector modulus. After the average pooling vector and the maximum pooling vector are normalized, the final matching degree can be calculated in the same unit.

[0096] Furthermore, the basic matching degree, the assumed hidden state, the premise hidden state, the network output assumption vector and the network output premise vector are subjected to the attention fusion mechanism of the semantic matching sub-model to generate the final matching degree, specifically:

[0097] The first difference between the assumed hidden state and the premise hidden state is weighted and calculated through the attention fusion mechanism, and the process can be specifically expressed as follows:

[0098]

[0099] In the formula, e i,j is the first difference base value, V T represents the parameter matrix, W1, W2 represent two training parameters, They represent the premise hidden state and the hypothesis hidden state respectively.

[0100] The first difference degree is calculated by normalizing the first difference degree base value.

[0101] The second difference is calculated by multiplying the first difference, the network output hypothesis vector, and the network output premise vector, and the sum is calculated. The process can be specifically expressed as follows:

[0102]

[0103] In the formula, e j represents the second degree of difference, represents the first difference after normalization, a t ,b t Represent the network output premise vector and network output hypothesis vector respectively.

[0104] The process of generating a hidden layer state through an attention fusion mechanism according to the second difference, the assumed hidden state and the premise hidden state can be specifically expressed as follows:

[0105]

[0106] In the formula, c i represents the hidden layer state at the i-th moment, They represent the premise hidden state and the assumed hidden state at the previous moment, respectively, and e j Represents the second difference at the jth moment.

[0107] The final matching degree is calculated according to the hidden layer state, the basic matching degree and the second difference degree, and the process can be specifically expressed as follows:

[0108] P(v i |v1,…,v i-1 )=g(v i-1 ,c i ,e j )

[0109] In the formula, v i represents the basic matching degree at the i-th moment, c i represents the hidden state at the i-th moment, e j The above formula represents the final matching degree at the current moment calculated by fitting the curve.

[0110] In the above scheme, the NLP speech matching model is implemented to enhance the relevant information of the hypothesis words and premise words through the attention mechanism, thereby enhancing their final matching degree.

[0111] Furthermore, the intensity and pitch of the customer's voice in the voice segment are extracted, and the intensity and pitch are converted and processed to calculate the customer's intended emotional valence, specifically:

[0112] The standard deviation of the sound intensity of multiple customer voice segments of the first set time length is used as the sound intensity emotional arousal value;

[0113] Performing Fourier transform on multiple customer voice segments of the second set time length to obtain multiple pitch signals, and using the standard deviation of the multiple pitch signals as the pitch attitude positivity;

[0114] The weight coefficient is calculated according to the frequency of occurrence of the intended keywords in the customer's voice and the speaking speed of the customer's voice, and the intended emotional valence is generated by weighted summation of the weight coefficient, the sound intensity emotional arousal value and the pitch attitude positivity.

[0115] Specifically, Fourier transform is used to convert the audio signal from the time domain to the frequency domain to analyze the pitch. According to experiments, when the standard deviation of the pitch signal is within the range of 270 Hz to 590 Hz, it indicates that the pitch fluctuation of the customer is small.

[0116] Specifically, the standard deviation is a conventional calculation method, which can represent the degree of discreteness of data. After experiments, when the standard deviation of the sound intensity signal is in the range of 0.03dB to 0.09dB, it indicates that the customer's emotions are stable and there is no extreme emotional fluctuation.

[0117] Specifically, the weight coefficient of the tone intensity emotional arousal value is proportional to the speech speed and the frequency of occurrence of the intended keywords, and the weight coefficient of the pitch attitude positivity is inversely proportional to the speech speed and the frequency of occurrence of the intended keywords.

[0118] Furthermore, the purchase propensity and the intention emotional valence are used to calculate the customer's purchase intention level through a logistic regression model, and whether to send product introduction information to the customer's device is determined according to the purchase intention level, specifically:

[0119] Generate a multi-classification logistic regression model based on the customer's gender information, age information and their corresponding sample data;

[0120] Substitute the purchase propensity and the intention emotional valence into the logistic regression model corresponding to the gender information and age information, and the input data classification probability generated by the logistic regression model corresponds to the grade gradient to determine the purchase intention grade. Specifically, the multi-classification logistic regression model uses Softmax as the activation function, and its process can be expressed as:

[0121] P = softmax(W1x1+W2x2+B)

[0122] Where P is the purchase intention level output by the logistic regression model, W1, W2, and B are three training parameters, and x1 and x2 represent the purchase tendency and the emotional valence of the intention after normalization.

[0123] In the above scheme, the logistic regression model is used to judge users of different genders and ages separately, making the purchase intention level combined with the emotional valence of the voice and the content information more accurate.

[0124] Furthermore, the system for automatically identifying customer intentions based on NLP includes an NLP voice matching model, a voice processing module and a logistic regression model installed in a cloud call center, wherein the NLP voice matching model is used to calculate the customer's intention keywords and purchase propensity by computing the consulting voice with product-related words and sales-related words;

[0125] The speech processing module is used to extract the sound intensity and pitch of the speech segment containing the intended keyword in the consultation speech, and convert the sound intensity and the pitch to calculate the intended emotional valence of the customer;

[0126] The logistic regression model is used to calculate the customer's purchase intention level based on the purchase propensity and the intention emotional valence. The purchase intention level is used for classification management of customer information and for determining whether to send product introduction information to the customer's device.

[0127] Furthermore, the cloud call center uses blockchain smart contracts to classify and securely access customer information. It is understandable that smart contracts are an automatically executed protocol on the blockchain that ensures the privacy and immutability of call recordings.

[0128] It can be understood that this embodiment uses the machine learning model of NLP technology to generate words related to goods and sales, judge the emotional attitude information contained in the customer's voice, and then comprehensively judge the customer's purchase intention, thereby reducing the cost of manually reviewing the consultation voice to classify and manage the customer's intention, greatly improving the efficiency of telephone sales. The NLP voice matching model uses the attention mechanism to enhance the relevant information of the hypothesis vocabulary and premise vocabulary and thus enhance its final matching degree. Through the logistic regression model, users of different genders and ages are judged separately, making the purchase intention level combined with the emotional valence and content information of the voice more accurate.

[0129] So far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it is easy for those skilled in the art to understand that the protection scope of the present invention is obviously not limited to these specific embodiments. Without departing from the principle of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will fall within the protection scope of the present invention.

[0130] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An information technology consulting service management method, characterized in that: include: The consultation voices of sales staff and customers are matched with product-related words and sales-related words through the NLP voice matching model based on machine learning to generate customers' intended keywords and purchase propensity; intercepting a voice segment of the consultation voice containing the intended keyword; Extracting the intensity and pitch of the customer's voice in the voice segment, converting the intensity and pitch to calculate the customer's intended emotional valence; The purchase propensity and the intention emotional valence are used to calculate the customer's purchase intention level through a logistic regression model. The purchase intention level is used for classification management of customer information and for determining whether to send product introduction information to the customer's device.

2. The information technology consulting service management method according to claim 1, characterized in that: The NLP speech matching model includes a corpus sub-model and a semantic matching sub-model; The process of generating the customer's intended keywords and purchase propensity by matching the consultation voice of the salesperson with product-related words and sales-related words through the NLP voice matching model based on machine learning is as follows: Acquire the product-related words and the sales-related words on the open source shopping website through crawling technology, and use them to construct the corpus sub-model reflecting the relationship between products and words; The consultation speech is preprocessed by data conversion to generate a word list, and the word list is used to find multiple candidate customer intention words in the corpus sub-model through a search engine; Calculate the final matching degree of the plurality of candidate customer intention words and the word list through the semantic matching sub-model; The candidate customer intention words whose final matching degree exceeds a final matching degree threshold are determined as the intention keywords, and the purchase tendency is calculated according to the intention keywords and their corresponding final matching degrees.

3. The information technology consulting service management method according to claim 2, characterized in that: The semantic matching sub-model is an improved long short-term memory network model that combines a bidirectional long short-term memory network and an attention fusion mechanism; The process of calculating the final matching degree of the plurality of candidate customer intention words and the word list through the semantic matching sub-model is as follows: The semantic matching sub-model generates a hypothesis word vector from the candidate customer intention word and generates a premise word vector from the word list; Calculating the hypothesis hidden state of the hypothesis word vector and the premise hidden state of the premise word vector through the bidirectional long short-term memory network of the semantic matching sub-model; The hypothesis hidden state and the premise hidden state are respectively passed through the time recursive long short-term memory network of the semantic matching sub-model to calculate a network output hypothesis vector and a network output premise vector; The network output hypothesis vector and the network output premise vector are respectively calculated by the average pooling layer and the maximum pooling layer of the semantic matching sub-model to obtain a hypothesis average pooling vector, a hypothesis maximum pooling vector, a premise average pooling vector and a premise maximum pooling vector, and the basic matching degree is calculated by cosine similarity; The basic matching degree, the assumed hidden state, the premise hidden state, the network output hypothesis vector and the network output premise vector are subjected to the attention fusion mechanism of the semantic matching sub-model to generate the final matching degree.

4. The information technology consulting service management method according to claim 3, characterized in that: The process of calculating the basic matching degree of the assumed average pooling vector, the assumed maximum pooling vector, the premise average pooling vector and the premise maximum pooling vector through the cosine similarity algorithm is as follows: After the assumed average pooling vector and the premise maximum pooling vector are normalized, a first matching degree is calculated using a cosine similarity formula; After the premise average pooling vector and the assumed maximum pooling vector are normalized, a second matching degree is calculated using a cosine similarity formula; Calculate a third matching degree between the assumed average pooling vector and the premise average pooling vector by using a cosine similarity formula; Calculate a fourth matching degree between the assumed maximum pooling vector and the premise maximum pooling vector by using a cosine similarity formula; The basic matching degree is calculated by weighted summing up the first matching degree, the second matching degree, the third matching degree and the fourth matching degree, wherein the weight value of the weighted summing up is calculated and determined according to the number of the candidate customer intention words and the length of the word list.

5. The information technology consulting service management method according to claim 3, characterized in that: The process of generating the final matching degree by the attention fusion mechanism of the semantic matching sub-model through the basic matching degree, the assumed hidden state, the premise hidden state, the network output hypothesis vector and the network output premise vector is as follows: The assumed hidden state and the premise hidden state are weighted to calculate a first difference through an attention fusion mechanism; Calculate a second difference by multiplying the first difference by the network output hypothesis vector and the network output premise vector; Generate a hidden layer state through an attention fusion mechanism according to the second difference, the assumed hidden state and the premise hidden state; The final matching degree is calculated according to the hidden layer state, the basic matching degree and the second difference degree.

6. The information technology consulting service management method according to claim 2, characterized in that: The process of generating a word list by preprocessing the consultation speech through data conversion is as follows: The consultation speech is converted into a call text by a speech recognition model based on a gated recurrent unit; The call text is segmented and stop words are removed by a segmentation tool to generate the word list.

7. The information technology consulting service management method according to any one of claims 1 to 6, characterized in that: The process of calculating the customer's intentional emotional valence is: The standard deviation of the sound intensity of multiple customer voice segments of the first set time length is used as the sound intensity emotional arousal value; Performing Fourier transform on multiple customer voice segments of the second set time length to obtain multiple pitch signals, and using the standard deviation of the multiple pitch signals as the pitch attitude positivity; The weight coefficient is calculated according to the frequency of occurrence of the intended keywords in the customer's voice and the speaking speed of the customer's voice, and the intended emotional valence is generated by weighted summation of the weight coefficient, the sound intensity emotional arousal value and the pitch attitude positivity.

8. The information technology consulting service management method according to any one of claims 1 to 6, characterized in that: The process of calculating the customer's purchase intention level by using the purchase propensity and the intention emotional valence through a logistic regression model is as follows: Generate a multi-classification logistic regression model based on the customer's gender information, age information and their corresponding sample data; The purchase propensity and the intention emotional valence are substituted into the logistic regression model corresponding to the gender information and age information. The input data classification probability generated by the logistic regression model corresponds to the grade gradient, and the purchase intention grade is determined.

9. A system using the information technology consulting service management method according to any one of claims 1 to 8, characterized in that: It includes the NLP voice matching model, voice processing module and logistic regression model installed in the cloud call center, among which: The NLP voice matching model is used to calculate the consulting voice with product-related words and sales-related words to generate the customer's intended keywords and purchase propensity; The speech processing module is used to extract the sound intensity and pitch of the speech segment containing the intended keyword in the consultation speech, and convert the sound intensity and the pitch to calculate the intended emotional valence of the customer; The logistic regression model is used to calculate the customer's purchase intention level based on the purchase propensity and the intention emotional valence. The purchase intention level is used for classification management of customer information and for determining whether to send product introduction information to the customer's device.

10. A system using the information technology consulting service management method according to any one of claims 1 to 9, characterized in that: one or more processors; Memory; and one or more programs, The one or more programs are configured to be executed by one or more processors to execute the information technology consulting service management method.

Images