CN116600053B

CN116600053B - Customer service system based on AI large language model

Info

Publication number: CN116600053B
Application number: CN202310869404.3A
Authority: CN
Inventors: 范洪昌
Original assignee: Beijing Renzhong Internet Information Technology Co ltd
Current assignee: Beijing Renzhong Internet Information Technology Co ltd
Priority date: 2023-07-17
Filing date: 2023-07-17
Publication date: 2023-10-27
Anticipated expiration: 2043-07-17
Also published as: CN116600053A

Abstract

The application discloses a customer service system based on an AI large language model, and relates to the technical field of intelligent telephone service. The system comprises an outbound dialing module, a client voice input module, a voice text recognition module, a sentence type recognition module, an electric marketing start-up interaction module, an electric marketing professional interaction module and a voice synthesis output module, and through the functional coordination of the multiple modules, the system can actively dial and carry out start-up communication dialogue and professional communication dialogue for clients based on an artificial intelligence technology, thereby improving professional communication service quality, client satisfaction and electric marketing success rate, and greatly reducing electric marketing cost. In addition, the professional communication link can enable customer service personnel to intervene at any time to perform professional communication dialogue, so that the purposes of man-machine cooperative application, client noninductive experience, screening out intended clients and the like are achieved, and practical application and popularization are facilitated.

Description

Customer service system based on AI large language model

Technical Field

The application belongs to the technical field of intelligent telephone service, and particularly relates to a customer service system based on an AI large language model.

Background

For a long time, telemarketing relies on traditional manual calls, i.e. existing telemarketing systems operate mainly by manual customer service in a telephone outbound manner. It follows that the size of the electric pin center is limited by the number of manual agents, and if more telephone outbound operations are to be performed, the electric pin center needs to hire more staff, but this will lead to problems of large staff recruitment, staff training, scale management, serious talent loss, and increasingly higher cost of hiring manual work.

Currently, with the advent of voice robots, some businesses began to employ artificial intelligence AI (Artificial Intelligence) robots to handle telephone customer service traffic. Compared with manual customer service, the AI robot has higher cost performance, so that the AI robot is adopted to replace the manual customer service in a plurality of industrial fields has become a development trend.

However, the practice finds that the AI robot in the prior art is limited by the accuracy of voice recognition and semantic understanding of the robot, the content quantity of the knowledge base and the current related technical means are not mature enough, so that the problem that professional communication service quality is limited when the AI robot is adopted to replace manual customer service to carry out telephone outbound operation, and further, customer satisfaction and electric sales yield are reduced is caused.

Disclosure of Invention

The invention aims to provide a customer service system based on an AI large language model, which is used for solving the problems that professional communication service quality is limited and customer satisfaction and electric sales success rate are reduced when an AI robot is adopted to replace manual customer service to carry out telephone calling operation.

In order to achieve the above purpose, the present invention adopts the following technical scheme:

the customer service system based on the AI large language model comprises an outbound dialing module, a customer voice input module, a voice text recognition module, a sentence type recognition module, an electric marketing interaction module, an electric marketing professional interaction module and a voice synthesis output module;

The outbound dialing module is used for actively dialing to call the customer telephone equipment according to the customer telephone number and obtaining a call result;

the electric marketing interaction module is in communication connection with the outbound dialing module and is used for generating an electric marketing white text according to an electric marketing preset strategy when the client telephone equipment is successfully called;

the client voice input module is in communication connection with the outbound dialing module and is used for inputting a client voice signal from the client telephone equipment in real time after the client telephone equipment is successfully called;

the voice text recognition module is in communication connection with the client voice input module and is used for converting the client voice signal into a client input text in real time;

the sentence type recognition module is in communication connection with the voice text recognition module and is used for judging corresponding sentence types in real time for the latest obtained sentences in the input text of the client, wherein the sentence types are divided into an open communication sentence type and a professional communication sentence type;

the electric marketing start interaction module is in communication connection with the sentence type identification module and is used for responding to the latest obtained sentence according to the electric marketing start preset strategy when judging that the sentence type of the latest obtained sentence is the start communication sentence type, and generating a start communication reply text corresponding to the latest obtained sentence;

The electronic marketing professional interaction module is in communication connection with the sentence type identification module and is used for inputting the latest obtained sentence into an AI large language model and outputting to obtain a professional communication reply text when judging that the sentence type of the latest obtained sentence is the professional communication sentence type;

the voice synthesis output module is respectively in communication connection with the electric marketing start interaction module and the electric marketing professional interaction module, and is used for synthesizing the electric marketing start white text, the start communication reply text and the professional communication reply text into customer service voice signals and transmitting the customer service voice signals to the customer telephone equipment in real time.

Based on the above summary of the invention, a new smart phone service scheme for performing phone outbound operation based on a voice recognition technology and an automatic communication interaction technology is provided, that is, the smart phone service scheme comprises an outbound dialing module, a client voice input module, a voice text recognition module, a sentence type recognition module, an electric sales start interaction module, an electric sales professional interaction module and a voice synthesis output module, wherein the voice text recognition module is used for converting the client voice signal into a client input text in real time, the sentence type recognition module is used for determining the corresponding sentence type in real time for the latest obtained sentence in the client input text, the electric sales start interaction module is used for generating an electric sales start white text according to an electric sales start preset strategy when a client phone device is successfully called, and when the sentence type is judged to be the type of the open communication sentence, responding and generating an open communication reply text corresponding to the latest obtained sentence according to the electric marketing open preset strategy, wherein the electric marketing professional interaction module is used for inputting the latest obtained sentence into an AI large language model and outputting the latest obtained sentence to obtain a professional communication reply text when the sentence type is judged to be the type of the professional communication sentence, and the voice synthesis output module is used for synthesizing the electric marketing open white text, the open communication reply text and the professional communication reply text into a customer service voice signal and transmitting the customer service voice signal to the customer telephone equipment in real time. Therefore, the system can actively dial to clients based on the artificial intelligence technology and carry out a start communication session and a professional communication session, so that the professional communication service quality, the client satisfaction degree and the electric marketing success rate can be improved, the electric marketing cost can be greatly reduced, and the system is convenient for practical application and popularization.

In one possible design, converting the customer speech signal to customer input text in real time includes:

and the intelligent voice recognition cloud service of hundred degrees, ali, tech or scientific big news flight is called by the Internet phone exchange IPPBX through a real-time transmission protocol RTP to analyze the client voice signal, and a text obtained by analysis is returned to the local through a transmission control protocol TCP to obtain the client input text.

In one possible design, determining, in real time, for the most recently obtained sentence in the customer input text, the corresponding sentence type includes:

word segmentation processing is carried out on the latest sentences in the input text of the client to obtain a latest word set;

if the latest word set has the electric sales professional word and the questioning common word, judging that the sentence type of the latest obtained sentence is the professional communication sentence type, otherwise, judging that the sentence type of the latest obtained sentence is the open communication sentence type.

In one possible design, responding to the latest obtained sentence according to the electric marketing offer preset strategy, generating an offer communication reply text corresponding to the latest obtained sentence, including:

Acquiring a plurality of conference communication client sentences and a plurality of conference communication response texts corresponding to the conference communication client sentences one by one from the electric marketing conference preset strategy;

calculating the similarity between the corresponding sentence and the latest obtained sentence for each of the plurality of the open communication guest sentences;

and taking the open communication response text corresponding to the open communication client sentence with the maximum similarity in the plurality of open communication client sentences as the open communication response text corresponding to the latest obtained sentence.

In one possible design, for a certain open communication guest sentence among the plurality of open communication guest sentences, calculating to obtain a similarity between a corresponding sentence and the latest obtained sentence includes:

word segmentation processing is respectively carried out on the sentence of the certain start communication guest and the latest obtained sentence, so as to obtain two word sets;

calculating a first index value, a second index value and a third index value of the two word sets, wherein the first index value is used for representing the similarity degree of the two word sets, which is calculated based on word frequency-inverse document frequency TF-IDF, the second index value is used for representing the difference degree of the two word sets, which is calculated based on an editing distance MED, and the third index value is used for representing the similarity degree of the two word sets, which is calculated based on a Jacquard similarity coefficient;

And taking the maximum value from the product of the first index value and the third index value and the product of the second index value and the third index value, and taking the maximum value as the similarity between the client sentence and the latest obtained sentence in the certain start communication.

In one possible design, the system also comprises a voice playing module and a customer service voice input module;

the voice playing module is in communication connection with the client voice input module and is used for playing the client voice signal to customer service personnel;

the customer service voice input module is in communication connection with the sentence type recognition module and is used for permitting the input of an artificial voice signal which is from the customer service personnel and is used for replying to the latest obtained sentence when the sentence type of the latest obtained sentence is judged to be the professional communication sentence type;

the voice text recognition module is also in communication connection with the customer service voice input module and is used for converting the manual voice signal into manual input text in real time;

the voice synthesis output module is also in communication connection with the voice text recognition module and is used for synthesizing the manual input text into the customer service voice signal and canceling synthesizing the professional communication reply text into the customer service voice signal when the manual input text is earlier than the professional communication reply text.

In one possible design, when the manually entered text occurs before the professional communication reply text, synthesizing the manually entered text into the customer service voice signal includes:

word segmentation processing is carried out on the manual input text to obtain a manual input word set;

judging whether a preset sensitive word exists in the manual input word set, if so, canceling synthesizing the manual input text into the customer service voice signal, waiting for the professional communication reply text, synthesizing the professional communication reply text into the customer service voice signal, and otherwise, synthesizing the manual input text into the customer service voice signal.

In one possible design, the system further comprises a customer sex identification module;

the client gender recognition module is in communication connection with the client voice input module and is used for extracting voiceprint features from the client voice signals, then importing the voiceprint features into a neural network-based gender recognition model which is trained, and outputting to obtain client gender;

the voice synthesis output module is also in communication connection with the customer sex identification module and is used for synthesizing the customer service voice signals with opposite sexes according to the customer sexes.

In one possible design, a customer intent recognition module is also included;

the client intention recognition module is in communication connection with the voice text recognition module and is used for estimating the current electric marketing intention of the client according to the following steps:

collecting all of said customer input text for a plurality of sample customers and recorded during a time period when the customer telephone device was successfully called;

for each sample client in the plurality of sample clients, sequentially splicing all corresponding client input texts according to the sequence from the first to the last of the acquisition time stamps to obtain corresponding character strings;

the character strings of all sample clients are preprocessed respectively to obtain new character strings of all sample clients, and non-characteristic characters are removed, wherein the non-characteristic characters refer to characters which do not reflect the behavior characteristics of clients;

according to all the new character strings, calculating the word frequency of each word in a first word set in each new character string by adopting a word frequency-inverse document frequency TF-IDF statistical algorithm, wherein the first word set is obtained by word segmentation processing of all the new character strings;

the method comprises the following steps of respectively selecting a plurality of sample feature words for each sample client: selecting front Round (w multiplied by M) words belonging to corresponding categories and sequenced from high to low according to the word frequency of sample clients from the first word set as corresponding selected feature words for each word category, and then summarizing the selected feature words of each word category to obtain a plurality of sample feature words, wherein the word frequency of the sample clients refers to the word frequency of words in the new character string corresponding to the sample clients, w refers to the ratio of words of the corresponding category in the first word set, M refers to the total number of feature word selections, and Round () refers to rounding-rounding function;

Performing classification training on a machine learning model based on the LightGBM by using the plurality of sample feature words and the electric pin intention labels of each sample client to obtain an electric pin intention prediction model, wherein the electric pin intention labels are used for marking whether electric pin intention exists in the corresponding clients;

sequentially splicing all the input texts of the current client and recorded since the client telephone equipment is successfully called according to the sequence of the time stamps from first to last to obtain a current character string;

performing the preprocessing on the current character string to obtain a current new character string from which the non-characteristic characters are removed;

according to the current new character string and all the new character strings, calculating the word frequency of each word in a second word set in the current new character string by adopting the word frequency-inverse document frequency TF-IDF statistical algorithm, wherein the second word set is obtained by word segmentation processing of the current new character string and all the new character strings;

selecting a plurality of current feature words for the current client in the following manner: firstly, selecting front Round (w multiplied by M) words belonging to corresponding categories and sequenced from high to low according to the current client word frequency from the second word set as corresponding selected feature words aiming at each word category, and then summarizing the selected feature words of each word category to obtain a plurality of current feature words, wherein the current client word frequency refers to the word frequency of a word in the current new character string corresponding to the current client;

Inputting the plurality of current feature words of the current client into the electric marketing trading intention prediction model, and outputting to obtain an electric marketing trading intention label or electric marketing trading probability of the current client.

In one possible design, the display device further comprises a display module;

the display module is respectively in communication connection with the outbound dialing module, the client voice input module, the voice text recognition module, the sentence type recognition module, the electric marketing interaction module, the electric marketing professional interaction module and/or the voice synthesis output module, and is used for displaying the work processing progress and/or the work processing result of each connecting module to customer service personnel.

The beneficial effect of above-mentioned scheme:

(1) The invention creatively provides a new intelligent telephone service scheme for carrying out telephone outbound operation based on a voice recognition technology and an automatic communication interaction technology, which comprises an outbound dialing module, a client voice input module, a voice text recognition module, a sentence type recognition module, an electric marketing open interaction module, an electric marketing professional interaction module and a voice synthesis output module, wherein the voice text recognition module is used for converting a client voice signal into a client input text in real time, the sentence type recognition module is used for judging the corresponding sentence type in real time aiming at the latest obtained sentence in the client input text, the electric marketing open interaction module is used for generating an electric marketing open white text according to an electric marketing open preset strategy when a client telephone device is successfully called, and responding and generating an open communication reply text corresponding to the latest obtained sentence according to the electric marketing open preset strategy when the sentence type is judged as the professional communication sentence type, and outputting the latest obtained sentence input large communication language and the electric marketing reply text to the client speech synthesis module when the sentence type is judged as the professional communication sentence type, and transmitting the electric marketing reply text to the client speech synthesis text and the real-time communication text. Therefore, the system can actively dial to the customer based on the artificial intelligence technology and carry out the start communication dialogue and the professional communication dialogue, thereby improving the professional communication service quality, the customer satisfaction degree and the electric marketing success rate and greatly reducing the electric marketing cost;

(2) The customer service personnel can intervene in the professional communication link at any time to perform professional communication dialogue, so that the purpose of man-machine cooperative application is realized, and the professional communication service quality, the customer satisfaction degree and the electric marketing success rate are ensured;

(3) When the user is in the middle of the user, the manual voice signal is firstly converted into the manual input text, and then the manual input text is synthesized into the customer service voice signal, so that the consistency of the customer service voice signal before and after the customer service voice signal can be ensured, the customers can sound indistinct, and the purpose of customer non-sensing experience is realized;

(4) In the process of carrying out the start communication dialogue and the professional communication dialogue, the electric marketing trading intention of the current customer can be determined in real time so as to achieve the purpose of screening the intention customers, further improve the electric marketing trading efficiency and avoid wasting unnecessary electric marketing cost.

Drawings

In order to more clearly illustrate the embodiments of the invention or the technical solutions in the prior art, the drawings that are required in the embodiments or the description of the prior art will be briefly described, it being obvious that the drawings in the following description are only some embodiments of the invention, and that other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.

Fig. 1 is a schematic structural diagram of a customer service system based on an AI large language model according to an embodiment of the present application.

Detailed Description

In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the present application will be briefly described below with reference to the accompanying drawings and the description of the embodiments or the prior art, and it is obvious that the following description of the structure of the drawings is only some embodiments of the present application, and other drawings can be obtained according to these drawings without inventive effort to a person skilled in the art. It should be noted that the description of these examples is for aiding in understanding the present application, but is not intended to limit the present application.

It should be understood that although the terms first and second, etc. may be used herein to describe various objects, these objects should not be limited by these terms. These terms are only used to distinguish one object from another. For example, a first object may be referred to as a second object, and similarly a second object may be referred to as a first object, without departing from the scope of example embodiments of the application.

It should be understood that for the term "and/or" that may appear herein, it is merely one association relationship that describes an associated object, meaning that there may be three relationships, e.g., a and/or B, may represent: three cases of A alone, B alone or both A and B exist; as another example, A, B and/or C, can represent the presence of any one of A, B and C or any combination thereof; for the term "/and" that may appear herein, which is descriptive of another associative object relationship, it means that there may be two relationships, e.g., a/and B, it may be expressed that: the two cases of A and B exist independently or simultaneously; in addition, for the character "/" that may appear herein, it is generally indicated that the context associated object is an "or" relationship.

Examples:

as shown in fig. 1, the customer service system provided in the first aspect of the present embodiment and based on the AI large language model includes, but is not limited to, an outbound dialing module, a customer voice input module, a voice text recognition module, a sentence type recognition module, an electric sales outlet interaction module, an electric sales professional interaction module and a voice synthesis output module.

The outbound dialing module is used for actively dialing and calling the customer telephone equipment according to the customer telephone number and obtaining a calling result. The foregoing customer phone number may be provided by an electric sales center person/customer service person input, specifically but not limited to a mobile phone number or a landline number (accordingly, the customer phone device may be but not limited to a customer mobile phone or a customer landline, etc.); when a plurality of customer telephone numbers are provided, an active dialing action can be performed one by one. In addition, the call result is specifically that the client telephone device is successfully called or the call fails (such as the case of busy, timeout missed or space number).

And the electric marketing interaction module is in communication connection with the outbound dialing module and is used for generating an electric marketing white text according to an electric marketing preset strategy when the client telephone equipment is successfully called. Because the phone call-out action is performed, the active call-out of the center side of the electric pin is required, and therefore, the electric pin call-out white text is required to be generated according to the preset strategy of the electric pin call-out, for example, the specific content of the electric pin call-out white text can be but is not limited to' hello, i is a medical consultant XXX of the XX company, disturb your work/rest, we company can now do a market research, can ask your help.

The client voice input module is in communication connection with the outbound dialing module and is used for inputting the client voice signal from the client telephone equipment in real time after the client telephone equipment is successfully called.

The voice text recognition module is in communication connection with the client voice input module and is used for converting the client voice signal into a client input text in real time. The specific algorithm for converting the voice signal into the text content is the existing algorithm; in order to fully utilize external powerful cloud computing service resources for rapid conversion purposes, it is preferable to convert the client speech signal into client input text in real time, including but not limited to: the customer speech signal is parsed by an IP pbx (IP-Private Branch eXchange, abbreviated form, which is an IP-based company telephone system that can integrate voice communication completely into a company's data network, thereby establishing a unified voice and data network capable of connecting and distributing over all over the world office sites and employees) with a Real-time transport protocol RTP (Real-time Transport Protocol) invoking intelligent speech recognition cloud services of hundred degrees, ali, tech, or science big news fly, etc., and the parsed text is returned locally by a transport control protocol TCP (Transmission Control Protocol), resulting in customer input text.

The sentence type recognition module is in communication connection with the voice text recognition module and is used for judging corresponding sentence types in real time for the latest obtained sentences in the input text of the client, wherein the sentence types are divided into an open communication sentence type and a professional communication sentence type. Because of the phone call-out action, there is generally an electric marketing process for first making a call and then making a professional communication, and further it is necessary to identify sentence types of different sentences in the text input by the client, so as to start different interactive modules. Specifically, for the latest sentence in the text input by the client, the corresponding sentence type is determined in real time, including but not limited to the following steps S41 to S42: s41, performing word segmentation on the latest sentences in the text input by the client to obtain a latest word set; s42, if the latest word set contains the electric sales professional word and the questioning common word, judging that the sentence type of the latest obtained sentence is the professional communication sentence type, otherwise, judging that the sentence type of the latest obtained sentence is the open communication sentence type. The foregoing tool for performing word segmentation may be, but not limited to, jieba word segmentation tool, where the specific content of the electric pin professional word and the questioning general word may be determined in advance according to the actual situation of the electric pin, for example, when the electric pin insurance product is an electric pin insurance product, some insurance industry terms may be used as the electric pin professional word, and "please ask" or "what" may be used as the questioning general word.

And the electric marketing start interaction module is in communication connection with the sentence type identification module and is used for responding to the latest obtained sentence according to the electric marketing start preset strategy and generating a start communication reply text corresponding to the latest obtained sentence when the sentence type of the latest obtained sentence is judged to be the start communication sentence type. Because the specific situations occurring in the start-up communication link can be expected in advance, the latest obtained sentence can be responded according to the electric marketing start-up preset strategy, specifically, the start-up communication reply text corresponding to the latest obtained sentence is generated according to the electric marketing start-up preset strategy, including but not limited to the following steps S51 to S53.

S51, acquiring a plurality of conference communication client sentences and a plurality of conference communication response texts corresponding to the conference communication client sentences one by one from the electric marketing conference preset strategy.

In the step S51, the preset strategy of electric marketing in the presence of the customer may be preset based on a plurality of different situations occurring in the presence of the customer, i.e. each situation may correspond to a sentence of the customer in the presence of the customer, and the corresponding presence of the customer may be customized in advance, for example, when the sentence of the customer in the presence of the customer is "i consider", the corresponding presence of the customer may be customized in advance, i.e. when you are convenient, i ask you again ", etc.

S52, calculating the similarity between the corresponding sentence and the latest obtained sentence according to each of the plurality of the open communication guest sentences.

In the step S52, specifically, for a certain open communication guest sentence among the plurality of open communication guest sentences, the similarity between the corresponding sentence and the latest obtained sentence is calculated, including but not limited to the following steps S521 to S523.

S521, word segmentation processing is carried out on the sentence of the certain start communication client and the latest obtained sentence respectively, so as to obtain two word sets.

In the step S521, the tool used for the word segmentation process may also be, but not limited to, a jieba word segmentation tool.

S522, calculating a first index value, a second index value and a third index value of the two word sets, wherein the first index value is used for representing the similarity degree of the two word sets calculated based on word Frequency-inverse document Frequency TF-IDF (Term Frequency-Inverse Document Frequency), the second index value is used for representing the similarity degree of the two word sets calculated based on an edit distance MED (Minimum Edit Distance, which is a quantitative measurement of the difference degree of two character strings, the measurement mode is a measurement mode that one character string can be changed into the other character string after looking at least how many times, and the third index value is used for representing the difference degree of the two word sets calculated based on a Jacquard similarity coefficient (Jaccard Similarity Coefficient) used for measuring the similarity between the two sets and defined as the number of elements of an intersection set divided by the number of elements of a union set.

In the step S522, the first index value and the third index value respectively represent a similarity degree of two word sets different from each other, and the second index value represents a difference degree of two word sets, where the first index value may, but is not limited to, refer to definition of word frequency-inverse document frequency TF-IDF (i.e., statistically evaluate importance degree of a word to one of a document set or a corpus), and use importance degree of all words in one word set, which are statistically obtained based on word frequency-inverse document frequency, to another word set as the first index value; the second index value may be, but is not limited to, a definition of the edit distance MED (i.e., measuring how many times it takes to change one character string to another character string), and how many times it takes to change one word set to another word set based on the edit distance measurement is taken as the second index value; the third index value may be, but not limited to, a result of a calculation of dividing the number of intersecting words of the one word set and the other word set by the number of words of the union is taken as the third index value with reference to a definition of a jekcard similarity coefficient (i.e., a result of counting the number of elements of the intersection of two sets divided by the number of elements of the union).

S523, taking the maximum value from the product of the first index value and the third index value and the product of the second index value and the third index value, and taking the maximum value as the similarity between the customer sentence and the latest obtained sentence in the communication of a certain start.

In the step S523, although the first index value is a cosine value in the word frequency-inverse document frequency dimension, the second index value is normalized in the edit distance dimension and the third index value is normalized in the jaccard similarity coefficient dimension so that the values are all 0,1, it should be noted that the values given in each dimension describe different degrees of similarity/difference, and that even if the same value is given in each dimension, the degree of similarity/difference is also different, and how to aggregate the three index values in different cases is an important problem. Considering the representation meaning of the first index value, the value of the first index value is affected by word frequency, that is, when the two word sets are smaller, the word frequency of each word may be 1, and at this time, the importance degree of each word cannot be well distinguished, so that the first index value is unfavorable for calculating the similarity between sentences which are too short. Whereas the second index value is exactly opposite to the first index value, it is better suited to evaluate the variability between shorter sentences, since shorter sentences cut out words are limited, requiring fewer steps of computation to find the mapping of query words in the word set. And the third index value is an index that is insensitive to sentence length, i.e. regardless of sentence length, it always determines the set difference in terms of intersection size. Therefore, by calculating the final sentence similarity by using the formula max (first index value x third index value, second index value x third index value), the product of the second index value and the third index (generally, the second index value is larger than the first index value) can be highlighted when the sentence is shorter to comprehensively measure the similarity/difference of a pair of sentences, and the product of the first index value and the third index value (generally, the first index value is larger than the second index value) can be highlighted when the sentence is longer to comprehensively measure the similarity/difference of a pair of sentences, that is, the accuracy of the final obtained sentence similarity can be ensured regardless of the length of the sentences.

S53, taking the open communication response text corresponding to the open communication guest sentences which are in the plurality of open communication guest sentences and have the maximum similarity as the open communication response text corresponding to the latest obtained sentence.

And the electronic marketing professional interaction module is in communication connection with the sentence type identification module and is used for inputting the latest obtained sentence into an AI large language model and outputting and obtaining the professional communication reply text when judging that the sentence type of the latest obtained sentence is the professional communication sentence type. The AI large language model refers to a deep learning method based on artificial intelligence technology, and is used for reasoning and generating new language expression with certain semantics and consistency. These models can automatically understand and process natural language and even enable conversations, authoring various text content, answering questions, etc. Specifically, the AI large language models can be divided into two main classes: the language generation model is characterized in that a neural network is trained through learning of large-scale data, and after a section of speech is input, the neural network automatically generates an article, story or other types of text content which is consistent with the original input logic. Since GPT-3 (generating Pre-trained Transformer 3) is the most powerful and advanced language generation model at present, which possesses up to 1750 hundred million parameter quantities, can fit almost any type of text, while also providing near-human-level natural language understanding capabilities, the AI large language model preferably employs the GPT-3 language generation model so that it can be used as a chat robot tool for professional communication replies, resulting in professional communication reply text corresponding to the latest resulting sentence.

The voice synthesis output module is respectively in communication connection with the electric marketing start interaction module and the electric marketing professional interaction module, and is used for synthesizing the electric marketing start white text, the start communication reply text and the professional communication reply text into customer service voice signals and transmitting the customer service voice signals to the customer telephone equipment in real time. The aforementioned specific way of synthesizing text content into speech signals is the existing conventional way. In addition, the transmission period of the customer service voice signal needs to be separated from the receiving period of the customer service voice signal so as to ensure the communication experience of both parties.

The voice text recognition module is used for converting the client voice signal into a client input text in real time, the sentence type recognition module is used for judging the corresponding sentence type in the latest obtained sentence in the client input text in real time, the electric expense on-site interaction module is used for generating an electric expense on-site white text according to an electric expense on-site preset strategy when the client telephone equipment is successfully called, and responding and generating an on-site reply text corresponding to the latest obtained sentence according to the electric expense on-site preset strategy when the sentence type is judged to be the on-site communication sentence type, the electric expense on-site professional interaction module is used for responding and generating the on-site reply text corresponding to the latest obtained sentence when the sentence type is judged to be the communication sentence type, the electric expense on-site voice communication specialized interaction module is used for outputting the latest obtained sentence into the voice communication text by the client input text, and outputting the electric expense on-site reply text by the client voice communication speech interaction module when the latest obtained by the client input text is judged to be the communication sentence type, and the electric expense on-site reply text is synthesized by the client voice communication text response module, and the electric expense on-site reply text is synthesized by the client voice communication text response text is obtained by the aid of the latest obtained by the client voice interaction module, and the client voice communication text is synthesized by the client voice interaction module. Therefore, the system can actively dial to clients based on the artificial intelligence technology and carry out a start communication session and a professional communication session, so that the professional communication service quality, the client satisfaction degree and the electric marketing success rate can be improved, the electric marketing cost can be greatly reduced, and the system is convenient for practical application and popularization.

Preferably, the system also comprises a voice playing module and a customer service voice input module; the voice playing module is in communication connection with the client voice input module and is used for playing the client voice signal to customer service personnel; the customer service voice input module is in communication connection with the sentence type recognition module and is used for permitting the input of an artificial voice signal which is from the customer service personnel and is used for replying to the latest obtained sentence when the sentence type of the latest obtained sentence is judged to be the professional communication sentence type; the voice text recognition module is also in communication connection with the customer service voice input module and is used for converting the manual voice signal into manual input text in real time; the voice synthesis output module is also in communication connection with the voice text recognition module and is used for synthesizing the manual input text into the customer service voice signal and canceling synthesizing the professional communication reply text into the customer service voice signal when the manual input text is earlier than the professional communication reply text. Through the cooperation of the multiple modules, not only can customer service personnel of an electric sales seat monitor the speaking of customers, but also can intervene in the professional communication link at any time to carry out the professional communication dialogue, thereby realizing the purpose of man-machine cooperative application and ensuring the professional communication service quality, customer satisfaction and electric sales success rate. In addition, when the user is in the middle of the user, the manual voice signal is firstly converted into the manual input text, and then the manual input text is synthesized into the customer service voice signal, so that the consistency of the customer service voice signal can be ensured, the customers can sound indistinct, and the purpose of customer noninductive experience is realized.

Further preferably, when the manual input text occurs before the professional communication reply text, the manual input text is synthesized into the customer service voice signal, including but not limited to: firstly, word segmentation processing is carried out on the manual input text to obtain a manual input word set; judging whether a preset sensitive word exists in the manual input word set, if so, canceling synthesizing the manual input text into the customer service voice signal, waiting for the professional communication reply text, synthesizing the professional communication reply text into the customer service voice signal, and otherwise, synthesizing the manual input text into the customer service voice signal. The tool used for the word segmentation process can also be, but is not limited to, a jieba word segmentation tool. The specific content of the sensitive words can be predetermined according to the electric marketing technique, such as some words which are not civilized and words which are illegally used. Therefore, whether the manual reply is qualified or not can be automatically checked through the steps, and the manual reply is automatically switched into the machine reply when the non-compliance is found, so that professional communication service quality, customer satisfaction and electric marketing success rate are further ensured.

Preferably, the system also comprises a customer sex identification module; the client gender recognition module is in communication connection with the client voice input module and is used for extracting voiceprint features from the client voice signals, then importing the voiceprint features into a neural network-based gender recognition model which is trained, and outputting to obtain client gender; the voice synthesis output module is also in communication connection with the customer sex identification module and is used for synthesizing the customer service voice signals with opposite sexes according to the customer sexes. The specific extraction mode of the voiceprint features is an existing conventional mode. The training process of the gender identification model is also the conventional model training process, for example, voiceprint features and gender labels of a plurality of sample personnel are led into a neural network model for training, and the gender identification model is obtained. Through the cooperation of the two modules, the sex of the customer can be identified, and the specific customer service voice synthesis is performed based on the sex identification result of the customer, so that the affinity of electric marketing communication is improved, and the professional communication service quality, customer satisfaction and electric marketing success rate are further ensured.

Preferably, the system also comprises a client intention recognition module; the customer intention recognition module is in communication connection with the voice text recognition module and is used for estimating the current customer' S trading intention of electric marketing according to the following steps S901-S911.

S901. collect all of said customer input text for a plurality of sample customers and recorded during the time period that the customer phone device was successfully called.

In the step S901, the sample client is a history dial-up call client marked with an electric sales intention label. The electric pin intention labels are used for marking whether corresponding clients have electric pin intention, and can be marked manually or according to final historical electric pin intention records, for example, historical dialing call clients with the history of electric pin intention are marked as electric pin intention, and other historical dialing call clients are marked as no electric pin intention.

S902, for each sample client in the plurality of sample clients, all corresponding client input texts are spliced in sequence according to the sequence of the time stamps from first to last, and corresponding character strings are obtained.

In the step S902, the obtained timestamp is a timestamp obtained by converting the text input by the client.

S903, respectively preprocessing the character strings of each sample client to obtain new character strings of each sample client, wherein the non-characteristic characters are characters which do not reflect the behavior characteristics of the client, and the non-characteristic characters are removed.

In the step S903, the preprocessing can prevent the non-feature characters in the original character string from participating in the subsequent feature extraction process, so as to ensure the correctness of the subsequent feature extraction result and the model training result. In addition, the non-characteristic characters may be exemplified by punctuation marks, and mood aid words.

S904, according to all the new character strings, calculating the word frequency of each word in a first word set in each new character string by adopting a word frequency-inverse document frequency TF-IDF statistical algorithm, wherein the first word set is obtained by word segmentation processing of all the new character strings.

In the step S904, the word Frequency-inverse document Frequency TF-IDF (Term Frequency-Inverse Document Frequency) statistical algorithm is an existing algorithm. Furthermore, the word segmentation process may also be, but is not limited to, employing a jieba word segmentation tool.

S905, respectively selecting a plurality of sample feature words for each sample client in the following manner: selecting the first Round (w multiplied by M) words belonging to the corresponding category and sequenced from high to low according to the word frequency of the sample client from the first word set as corresponding selected feature words for each word category, and then summarizing the selected feature words of each word category to obtain the plurality of sample feature words, wherein the word frequency of the sample client refers to the word frequency of the word in the new character string corresponding to the sample client, w refers to the ratio of the word of the corresponding category in the first word set, M refers to the total number of feature word selections, and Round () refers to rounding-rounding function.

In the step S905, for example, if there are five word classes: a word class a (the ratio of the class of words in the first word set is 5%), a word class B (the ratio of the class of words in the first word set is 25%), a word class C (the ratio of the class of words in the first word set is 35%), a word class D (the ratio of the class of words in the first word set is 15%), and a word class E (the ratio of the class of words in the first word set is 20%), while there is a total number of feature word choices M, and there is a new character string a corresponding to a sample client a, the first 150 words belonging to the corresponding class and ordered in the order of sample client word frequency (i.e., word frequency of words in the new character string a) can be selected from the first word set as corresponding selected feature words for the word class a; for the word class B, the first 750 words belonging to the corresponding class and sequenced from high to low according to the word frequency of the sample client can be selected from the first word set to serve as corresponding selected feature words; for the word category C, the first 1050 words belonging to the corresponding category and sequenced from high to low according to the word frequency of the sample client can be selected from the first word set to serve as corresponding selected feature words; for the word class D, the first 450 words belonging to the corresponding class and sequenced from high to low according to the word frequency of the sample client can be selected from the first word set to serve as corresponding selected feature words; for word class E, the first 600 words belonging to the corresponding class and ordered in the order of top to bottom according to the sample client word frequency may be selected from the first word set as corresponding selected feature words. Finally, the first 150 words, the first 750 words, the first 1050 words, the first 450 words, and the first 600 words are summarized to obtain the plurality of sample feature words of the sample client a (i.e., 1000 words, where the word overlapping condition is not considered). Through the weight-based feature word selection mode, the following problems of the traditional selection mode (namely, directly selecting the first M feature words sequenced from high to low according to word frequency) can be avoided: it may result in feature words being selected more from the majority classes of the word set when the data set is unbalanced, such that the minority classes are not typically sufficient in the feature matrix returned based on TF-IDF.

S906, performing classification training on the machine learning model based on the LightGBM by using the sample feature words and the electric pin achievement intention labels of the sample clients to obtain an electric pin achievement intention prediction model, wherein the electric pin achievement intention labels are used for marking whether electric pin achievement intention exists in the corresponding clients.

In the step S906, the LightGBM (Light Gradient Boosting Machine) is a framework for implementing the GBDT (Gradient Boosting Decision Tree, gradient-enhanced decision tree) algorithm, which can support efficient parallel training, and has the advantages of faster training speed, lower memory consumption, better accuracy, support for distribution, and capability of rapidly processing massive data, so that the machine learning model can be built based on the existing knowledge. Meanwhile, since the plurality of sample feature words of each sample client can form a D×M feature matrix so as to serve as model input data, and the electric pin achievement designation label of each sample client can form a D×M verification matrix so as to serve as model output data, the electric pin achievement designation prediction model can be obtained through a conventional two-class training mode, wherein D represents the total number of sample clients. Preferably, in the two-classification training process of the machine learning model, a Bayesian optimization algorithm based on a tree structure can be used for optimizing model parameters, but is not limited to.

S907, all the input texts of the current client and recorded since the client telephone equipment is successfully called are spliced in sequence from the beginning to the end according to the acquired time stamps, so that the current character string is obtained.

S908, preprocessing the current character string to obtain a current new character string from which the non-characteristic characters are removed.

S909, according to the current new character string and all the new character strings, calculating the word frequency of each word in a second word set in the current new character string by adopting the word frequency-inverse document frequency TF-IDF statistical algorithm, wherein the second word set is obtained by word segmentation processing of the current new character string and all the new character strings.

S910, selecting a plurality of current feature words for the current client in the following mode: first, selecting front Round (w×m) words belonging to the corresponding category and sequenced from high to low according to the current customer word frequency from the second word set as corresponding selected feature words for each word category, and then summarizing the selected feature words of each word category to obtain the plurality of current feature words, wherein the current customer word frequency refers to the word frequency of the word in the current new character string corresponding to the current customer.

Specific technical details of the steps S907 to S910 can be derived with reference to the steps S902 to S905, and are not described herein.

S911, inputting the plurality of current feature words of the current client into the electric pin achievement intention prediction model, and outputting to obtain electric pin achievement intention labels or electric pin achievement probabilities of the current client.

Through the functional design of the client intention recognition module, the intention of the current client to be in electric pin deal can be determined in real time in the process of carrying out the session of communication and professional communication, so that the purpose of screening out the intention clients is realized, the efficiency of electric pin deal is further improved, and unnecessary electric pin cost can be avoided, for example: and the electric marketing interaction module is also in communication connection with the customer intention recognition module and is also used for triggering and generating an electric marketing end text according to the preset electric marketing strategy if the current electric marketing intention of the customer or the electric marketing probability of the current customer exceeds a preset probability threshold after the customer telephone equipment is successfully called, so that the voice synthesis output module also synthesizes the electric marketing end text into the customer service voice signal and transmits the customer service voice signal to the customer telephone equipment in real time, and then actively hangs up the telephone with the current customer.

Preferably, the display module is further included; the display module is respectively in communication connection with the outbound dialing module, the client voice input module, the voice text recognition module, the sentence type recognition module, the electric marketing interaction module, the electric marketing professional interaction module and/or the voice synthesis output module and the like, and is used for displaying the work processing progress and/or the work processing result of each connecting module to customer service personnel. Through the configuration of the display module, the conversation between the robot and the client can be monitored/monitored in real time, so that customer service personnel can grasp the conversation progress, intervene or end the conversation at any time, and the practicability of the system is improved. In addition, the display module can be respectively connected with the customer gender identification module and/or the customer intention identification module in a communication way, so that the work processing progress and/or the work processing result of the connecting module can be displayed to the customer service personnel, for example, the gender identification result or the electric marketing intention identification result of the current customer is displayed to the customer service personnel, and the practicability is further improved.

In summary, the customer service system based on the AI large language model provided by the embodiment has the following technical effects:

(1) The embodiment provides a new smart phone service scheme for performing phone outbound operation based on a voice recognition technology and an automatic communication interaction technology, which comprises an outbound dialing module, a client voice input module, a voice text recognition module, a sentence type recognition module, an electric marketing open interaction module, an electric marketing professional interaction module and a voice synthesis output module, wherein the voice text recognition module is used for converting a client voice signal into a client input text in real time, the sentence type recognition module is used for judging a corresponding sentence type in real time for a latest obtained sentence in the client input text, the electric marketing open interaction module is used for generating an electric marketing open white text according to an electric marketing open preset strategy when a client telephone device is successfully called, and responding and generating an open communication reply text corresponding to the latest obtained sentence according to the electric marketing open preset strategy when the sentence type is judged as an open communication sentence type, the electric marketing interaction module is used for inputting a large communication language for inputting the latest obtained sentence, outputting the latest obtained sentence as a speech communication language, and outputting the electric marketing reply text to the client speech communication text, and transmitting the electric marketing reply text to the client speech communication text to the client telephone device in real time. Therefore, the system can actively dial to the customer based on the artificial intelligence technology and carry out the start communication dialogue and the professional communication dialogue, thereby improving the professional communication service quality, the customer satisfaction degree and the electric marketing success rate and greatly reducing the electric marketing cost;

Finally, it should be noted that: the foregoing description is only of the preferred embodiments of the invention and is not intended to limit the scope of the invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. The customer service system based on the AI large language model is characterized by comprising an outbound dialing module, a customer voice input module, a customer gender recognition module, a voice text recognition module, a customer intention recognition module, a sentence type recognition module, an electric sales start interaction module, an electric sales professional interaction module and a voice synthesis output module;

the sentence type recognition module is in communication connection with the voice text recognition module, and is used for determining the corresponding sentence type in real time for the latest obtained sentence in the input text of the client, and specifically comprises the following steps: word segmentation processing is carried out on the latest sentences in the input text of the client to obtain a latest word set; if the latest word set has the electric sales professional word and the questioning common word, judging that the sentence type of the latest obtained sentence is the professional communication sentence type, otherwise, judging that the sentence type of the latest obtained sentence is the open communication sentence type;

the voice synthesis output module is respectively in communication connection with the customer gender identification module, the electric marketing field interaction module and the electric marketing professional interaction module, and is used for synthesizing the electric marketing field white text, the field communication reply text and the professional communication reply text into customer service voice signals with opposite sexes according to the customer gender, and transmitting the customer service voice signals to the customer telephone equipment in real time;

the client intention recognition module is in communication connection with the voice text recognition module and is used for estimating the current electric marketing intention of the client according to the following steps: collecting all of said customer input text for a plurality of sample customers and recorded during a time period when the customer telephone device was successfully called; for each sample client in the plurality of sample clients, sequentially splicing all corresponding client input texts according to the sequence from the first to the last of the acquisition time stamps to obtain corresponding character strings; the character strings of all sample clients are preprocessed respectively to obtain new character strings of all sample clients, and non-characteristic characters are removed, wherein the non-characteristic characters refer to characters which do not reflect the behavior characteristics of clients; according to all the new character strings, calculating the word frequency of each word in a first word set in each new character string by adopting a word frequency-inverse document frequency TF-IDF statistical algorithm, wherein the first word set is obtained by word segmentation processing of all the new character strings; the method comprises the following steps of respectively selecting a plurality of sample feature words for each sample client: selecting front Round (w multiplied by M) words belonging to corresponding categories and sequenced from high to low according to the word frequency of sample clients from the first word set as corresponding selected feature words for each word category, and then summarizing the selected feature words of each word category to obtain a plurality of sample feature words, wherein the word frequency of the sample clients refers to the word frequency of words in the new character string corresponding to the sample clients, w refers to the ratio of words of the corresponding category in the first word set, M refers to the total number of feature word selections, and Round () refers to rounding-rounding function; performing classification training on a machine learning model based on the LightGBM by using the plurality of sample feature words and the electric pin intention labels of each sample client to obtain an electric pin intention prediction model, wherein the electric pin intention labels are used for marking whether electric pin intention exists in the corresponding clients; sequentially splicing all the input texts of the current client and recorded since the client telephone equipment is successfully called according to the sequence of the time stamps from first to last to obtain a current character string; performing the preprocessing on the current character string to obtain a current new character string from which the non-characteristic characters are removed; according to the current new character string and all the new character strings, calculating the word frequency of each word in a second word set in the current new character string by adopting the word frequency-inverse document frequency TF-IDF statistical algorithm, wherein the second word set is obtained by word segmentation processing of the current new character string and all the new character strings; selecting a plurality of current feature words for the current client in the following manner: firstly, selecting front Round (w multiplied by M) words belonging to corresponding categories and sequenced from high to low according to the current client word frequency from the second word set as corresponding selected feature words aiming at each word category, and then summarizing the selected feature words of each word category to obtain a plurality of current feature words, wherein the current client word frequency refers to the word frequency of a word in the current new character string corresponding to the current client; inputting the plurality of current feature words of the current client into the electric marketing trading intention prediction model, and outputting to obtain an electric marketing trading intention label or electric marketing trading probability of the current client;

And the electric marketing interaction module is also in communication connection with the customer intention recognition module and is also used for triggering and generating an electric marketing end text according to the preset electric marketing strategy if the current electric marketing intention of the customer or the electric marketing probability of the current customer exceeds a preset probability threshold after the customer telephone equipment is successfully called, so that the voice synthesis output module also synthesizes the electric marketing end text into the customer service voice signal and transmits the customer service voice signal to the customer telephone equipment in real time, and then actively hangs up the telephone with the current customer.

2. The customer service system of claim 1, wherein converting the customer speech signal in real time to customer input text comprises:

3. The customer service system of claim 1, wherein generating an open communication reply text corresponding to the latest resultant sentence in response to the latest resultant sentence according to the electric sales open preset strategy comprises:

4. A customer service system as claimed in claim 3, wherein calculating a similarity of a corresponding sentence to the latest obtained sentence for a specific one of the plurality of open communication customer sentences comprises:

5. The customer service system of claim 1, further comprising a voice playing module and a customer service voice input module;

6. The customer service system of claim 5, wherein synthesizing the manually entered text into the customer service voice signal when the manually entered text occurs before the professional communication reply text comprises:

7. The customer service system of claim 1, further comprising a display module;