Recommended verbal skill generation method and device, computer equipment and storage medium

Abstract

The invention belongs to the field of artificial intelligence and finance, and relates to a recommended verbal skill generation method, which comprises the steps of dividing different customer groups based on customer features extracted from historical underwriting data and historical service records, generating customer group portrait tags, obtaining conversation audio data of target customers, and generating a recommendation verbal skill generation result. Matching a target customer group and a target group portrait label of a target customer, preprocessing call audio data, performing voice recognition to obtain text data, obtaining an insurance wish result according to the text data and target customer characteristics, and retrieving a target service verbal skill according to an intention category obtained according to the text data and the target group portrait label; and generating a verbal skill suggestion result according to the target customer features and the target service verbal skill through a verbal skill recommendation model. The invention further provides a recommended verbal skill generation device, computer equipment and a storage medium. In addition, the invention also relates to a block chain technology, and historical underwriting data and historical service records can be stored in a block chain. According to the invention, personalized service verbal skills can be provided.

Description

Technical Field

[0001] The present application relates to the fields of artificial intelligence and financial technology, and in particular to a method, device, computer equipment and storage medium for generating recommendation words. Background Art

[0002] At present, the growth of the automobile market has shown a trend of slowing down. This change has directly affected the performance of related insurance companies, resulting in an increasingly significant decline in their performance. In order to meet this challenge, insurance companies are in urgent need of exploring new business growth points to maintain or improve their performance. Under this circumstance, non-auto property insurance products have gradually become a new focus of insurance companies. Such products include but are not limited to health insurance, pet insurance, travel insurance, frame accident insurance and other accident insurance, etc. They each have a wide audience group and extremely strong market demand.

[0003] However, the diversity of non-auto property insurance products has also brought new problems. Different types of non-auto insurance products target significantly different customer groups. Therefore, during the promotion and sales process, insurance companies cannot simply use the same sales pitch to deal with all types of insurance. Otherwise, not only will they fail to effectively attract customers, but they may also cause customer disgust, thereby reducing the policy closing rate. Summary of the invention

[0004] The purpose of the embodiments of the present application is to propose a method, device, computer equipment and storage medium for generating recommendation words, so as to solve the technical problem in the related art that the unified words may cause customers to be disgusted, thereby reducing the policy promotion rate.

[0005] In order to solve the above technical problems, the present application embodiment provides a method for generating recommendation words, which adopts the following technical solution:

[0006] Obtaining historical underwriting data and historical service records of all customers of the insurance company, extracting customer features from the historical underwriting data and the historical service records, dividing all customers into different customer groups based on the customer features, and generating customer group portrait labels;

[0007] Acquire call audio data of target customers, match corresponding target customer groups according to target customer characteristics of the target customers, and acquire target customer group portrait labels of the target customer groups;

[0008] Preprocessing the call audio data to obtain enhanced audio data, and performing speech recognition on the enhanced audio data to obtain text data;

[0009] Inputting the text data and the target customer characteristics into a trained insurance willingness prediction model to obtain an insurance willingness result;

[0010] When the insurance willingness result is willingness to insure, performing intent recognition on the text data to obtain an intent category;

[0011] According to the target customer group portrait label and the intent category, a matching target service script is retrieved from a pre-built service script library;

[0012] The target customer characteristics and the target service words are input into the trained words recommendation model to obtain the words recommendation results.

[0013] In order to solve the above technical problems, the embodiment of the present application further provides a device for generating recommendation words, which adopts the following technical solution:

[0014] A classification module is used to obtain the historical underwriting data and historical service records of all customers of the insurance company, extract customer features from the historical underwriting data and the historical service records, classify all the customers into different customer groups based on the customer features, and generate customer group portrait labels;

[0015] An acquisition module is used to acquire call audio data of a target customer, match a corresponding target customer group according to the target customer characteristics of the target customer, and acquire a target customer group portrait label of the target customer group;

[0016] A speech recognition module, used for preprocessing the call audio data to obtain enhanced audio data, and performing speech recognition on the enhanced audio data to obtain text data;

[0017] A willingness prediction module, used to input the text data and the target customer characteristics into a trained insurance willingness prediction model to obtain an insurance willingness result;

[0018] An intention recognition module, used for performing intention recognition on the text data to obtain an intention category when the insurance willingness result is willingness to insure;

[0019] A matching module, used to retrieve matching target service words from a pre-built service word library according to the target customer group portrait label and the intent category;

[0020] The speech recommendation module is used to input the target customer characteristics and the target service speech into the trained speech recommendation model to obtain the speech recommendation results.

[0021] In order to solve the above technical problems, the embodiment of the present application further provides a computer device, which adopts the following technical solution:

[0022] The computer device includes a memory and a processor, wherein the memory stores computer-readable instructions, and the processor implements the steps of the recommendation speech generation method as described above when executing the computer-readable instructions.

[0023] In order to solve the above technical problems, the embodiment of the present application further provides a computer-readable storage medium, which adopts the following technical solution:

[0024] The computer-readable storage medium stores computer-readable instructions, and when the computer-readable instructions are executed by a processor, the steps of the method for generating recommendation words as described above are implemented.

[0025] Compared with the prior art, this application has the following beneficial effects:

[0026] The present application provides a method for generating recommendation scripts, which divides all customers into different customer groups based on customer features extracted from historical underwriting data and historical service records, generates customer group portrait labels, obtains call audio data of target customers, matches target customers to corresponding target customer groups to obtain target group portrait labels, and then pre-processes the call audio data for speech recognition to obtain recognized text data, obtains the insurance willingness result based on the text data and target customer features, and when the insurance willingness result is willingness to insure, performs intent recognition on the text data to obtain the intent category, then retrieves the matching target service script from a pre-built service script library based on the target group portrait label and the intent category, and finally processes the target customer features and the target service script through a script recommendation model to obtain the script suggestion result; the present application can intelligently provide personalized service scripts based on the specific features, intent categories and insurance willingness of target customers to improve customer satisfaction and acceptance, which helps to improve the policy promotion rate, and at the same time, improves work efficiency, greatly reduces meaningless communication, and thereby improves the service skills of agents. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the scheme in the present application, a brief introduction is given below to the drawings required for use in the description of the embodiments of the present application. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0028] Figure 1 is an exemplary system architecture diagram to which the present application may be applied;

[0029] Figure 2 is a flow chart of an embodiment of a method for generating recommendation words according to the present application;

[0030] Figure 3 is a structural schematic diagram of an embodiment of a device for generating recommendation words according to the present application;

[0031] Figure 4 It is a structural diagram of an embodiment of a computer device according to the present application. DETAILED DESCRIPTION

[0032] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by technicians in the technical field of the present application; the terms used in the specification of the application herein are only for the purpose of describing specific embodiments and are not intended to limit the present application; the terms "including" and "having" and any variations thereof in the specification and claims of the present application and the above-mentioned drawings are intended to cover non-exclusive inclusions. The terms "first", "second", etc. in the specification and claims of the present application or the above-mentioned drawings are used to distinguish different objects, not to describe a specific order.

[0033] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0034] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the accompanying drawings.

[0035] like Figure 1 As shown, the system architecture 100 may include a terminal device 101, a network 102 and a server 103. The terminal device 101 may be a laptop 1011, a tablet computer 1012 or a mobile phone 1013. The network 102 is used to provide a medium for a communication link between the terminal device 101 and the server 103. The network 102 may include various connection types, such as wired, wireless communication links or optical fiber cables.

[0036] The user can use the terminal device 101 to interact with the server 103 through the network 102 to receive or send messages, etc. Various communication client applications can be installed on the terminal device 101, such as web browser applications, shopping applications, search applications, instant messaging tools, email clients, social platform software, etc.

[0037] The terminal device 101 can be any electronic device with a display screen and supporting web browsing. In addition to a laptop computer 1011, a tablet computer 1012 or a mobile phone 1013, the terminal device 101 can also be an e-book reader, an MP3 player (Moving Picture Experts Group Audio Layer III), an MP4 (Moving Picture Experts Group Audio Layer IV), a laptop computer, a desktop computer, etc.

[0038] The server 103 may be a server that provides various services, such as a background server that provides support for a web page displayed on the terminal device 101 .

[0039] It should be noted that the recommendation speech generation method provided in the embodiment of the present application is generally executed by a server / terminal device, and accordingly, the recommendation speech generation device is generally set in the server / terminal device.

[0040] It should be understood that Figure 1 The number of terminal devices, networks and servers in the embodiment is only for illustration. Any number of terminal devices, networks and servers may be provided according to implementation requirements.

[0041] Continue to refer Figure 2 , shows a flow chart of an embodiment of a method for generating recommendation words according to the present application, comprising the following steps:

[0042] Step S201, obtain the historical underwriting data and historical service records of all customers of the insurance company, extract customer characteristics from the historical underwriting data and historical service records, divide all customers into different customer groups based on the customer characteristics, and generate customer group portrait labels.

[0043] Among them, historical underwriting data includes customer basic information, policy information, etc.; historical service records include insurance application services, claims services, consulting services, policy change services, and complaint handling services.

[0044] In this embodiment, the electronic device (eg, Figure 1 The server / terminal device shown in the figure) can obtain historical underwriting data and historical service records through a wired connection or a wireless connection. It should be noted that the above wireless connection method may include but is not limited to 3G / 4G / 5G connection, WiFi connection, Bluetooth connection, WiMAX connection, Zigbee connection, UWB (ultra wideband) connection, and other wireless connection methods currently known or to be developed in the future.

[0045] Preprocess each customer's historical underwriting data and historical service records, clean invalid and duplicate data, fill or delete missing values, and normalize numerical features to obtain preprocessed customer data. Use principal component analysis to extract key features from preprocessed customer data, such as customer age, gender, occupation, family situation, vehicle situation, insurance product type, continuous years of underwriting in new channels, historical premiums, number of claims, number of complaints, continuous years of underwriting in new channels, renewal score, number of years of account manager system, underwriting time habits, average advance underwriting days, service cycle, number of effective contacts, etc., as customer characteristics of each customer. Based on the customer characteristics of each customer, all customers are clustered to obtain different cluster categories. Each cluster category is a customer group, such as late underwriting habit customer group, low premium preference customer group, etc.

[0046] In some optional implementations, the step of dividing all the customers into different customer groups based on the customer characteristics includes:

[0047] The optimal number of clusters K is determined by the silhouette coefficient, where K is a positive integer;

[0048] Convert each customer's customer features into a customer clustering vector, and select K customer clustering vectors from all customer clustering vectors as clustering centers;

[0049] Calculate the distance between other clustering vectors and each cluster center, and assign other clustering vectors to the cluster center with the closest distance according to the distance to form K cluster clusters, wherein the other clustering vectors are customer clustering vectors other than the initial cluster center;

[0050] Calculate the centroid of each cluster, use the centroid as the new cluster center, repeatedly calculate the distance between other cluster vectors and each cluster center, obtain K new clusters, and continuously update the cluster center of each cluster until the clustering stop condition is met to obtain the final K customer groups.

[0051] Among them, the silhouette coefficient is an indicator to measure the quality of clustering effect, and its value range is between [-1,1]. The closer the value is to 1, the better the clustering effect is, that is, the points in the same cluster are closer, and the points between different clusters are more scattered; when the silhouette coefficient is close to -1, it means that the clustering effect is poor, that is, the points may be incorrectly assigned to the cluster; when the silhouette coefficient is close to 0, it means that the separation and cohesion between clusters are similar, and clustering may not make much sense.

[0052] In this embodiment, before clustering, customer features are first vectorized and the Word2Vec algorithm and pre-trained BERT can be used to convert customer features into customer clustering vectors, and each customer corresponds to a customer clustering vector.

[0053] In a specific example, the Euclidean distance algorithm is used to calculate the distance between other cluster vectors and each cluster center.

[0054] In this embodiment, the clustering stopping condition may be that no (or a minimum number) objects are reallocated to different clusters, no (or a minimum number) cluster centers change again, the sum of squared errors is locally minimized, etc., so that the clustering vectors within the cluster are connected as closely as possible, and the distances between different clusters are as large as possible.

[0055] The average characteristic value is calculated based on the customer characteristics of each customer in each customer group, and a corresponding label is generated for each customer group based on the average characteristic value as a customer group portrait label. These labels can be descriptive text or numerical labels. For example, customer group A is mainly composed of customers who are between 30 and 45 years old, have stable jobs and higher incomes, purchase insurance frequently and have fewer claims. The generated customer group portrait label can be high-value customers; customer group B is mainly composed of customers who are over 50 years old, have unstable jobs or are retired, purchase insurance infrequently and have more claims. The generated customer group portrait label can be traditional customers.

[0056] By dividing customers into different customer groups, we can accurately locate the groups, formulate targeted marketing strategies and service plans for each customer group, and provide personalized services, which will help improve the policy promotion rate. At the same time, it will help to more accurately assess the risk levels of different customer groups, optimize resource allocation, cut unnecessary costs and expenditures, and improve overall operational efficiency.

[0057] It should be emphasized that in order to further ensure the privacy and security of historical underwriting data and historical service records, the above historical underwriting data and historical service records can also be stored in a blockchain node.

[0058] The blockchain referred to in this application is a new application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanism, encryption algorithm, etc. Blockchain is essentially a decentralized database, a string of data blocks generated by cryptographic methods. Each data block contains a batch of network transaction information, which is used to verify the validity of its information (anti-counterfeiting) and generate the next block. Blockchain can include the underlying blockchain platform, platform product service layer, and application service layer.

[0059] Step S202, obtaining the call audio data of the target customer, matching the corresponding target customer group according to the target customer characteristics of the target customer, and obtaining the target customer group portrait label of the target customer group.

[0060] In this embodiment, the target customer refers to an individual or organization that has potential demand or interest in insurance products and has established contact with an insurance company or its agent or expressed a willingness to purchase; the call audio data is audio data collected during a call with the target customer.

[0061] Extract the target customer features of the target customers, match the target customer features with each customer group through the feature matching algorithm, and select the customer group with the highest matching degree as the target customer group. Among them, the feature matching algorithm includes Jaccard similarity coefficient, cosine similarity, Euclidean distance, etc. For example, the Euclidean distance is used to calculate the distance similarity between the target customer features and the average feature value of the customer features of all customers in each customer group. The smaller the Euclidean distance, the higher the similarity. The customer group with the smallest Euclidean distance is selected as the target customer group of the target customer, and the target group portrait label corresponding to the target customer group is obtained accordingly.

[0062] Step S203, pre-processing the call audio data to obtain enhanced audio data, and performing speech recognition on the enhanced audio data to obtain text data.

[0063] In this embodiment, the call audio data is preprocessed, and the preprocessing includes denoising, framing, endpoint detection, and voice enhancement, etc. Specifically, the call audio data is denoised to obtain denoised audio data; the denoised audio data is framed to obtain voice frames; voice endpoint detection is performed on the voice frames to obtain valid voice frames; and finally, voice enhancement is performed on the valid voice frames to obtain enhanced audio data.

[0064] In this embodiment, spectral subtraction, Wiener filtering and other methods are used to denoise the original speech data in the speech data set. Among them, spectral subtraction is based on the assumption of the stationarity of the short-time spectrum, that is, it is assumed that the spectrum of the speech signal is stable and unchanged in a short time. The principle is to subtract the short-time spectrum of the noise from the short-time spectrum of the noisy speech signal to obtain the spectrum of the pure speech; Wiener filtering is a linear minimum mean square error (LMMSE) estimation, which minimizes the mean square error between the actual signal and the estimated signal by constructing a linear filter.

[0065] The speech signal is a non-steady-state, time-varying signal, but it can be considered that the speech signal is steady-state and time-invariant within a short time range. This short time is generally 10-30ms. Therefore, when processing the speech signal, in order to reduce the non-steady-state and time-varying effects of the overall speech signal, the speech signal is segmented, each segment is called a frame, and the frame length is generally 25ms. In order to make a smooth transition between frames and maintain their continuity, the framing generally adopts the overlapping segmentation method to ensure that two adjacent frames overlap each other. The time difference between the starting positions of two adjacent frames is called frame shift. In the application process, the frame shift can be 10ms. It should be understood that the frame length and frame shift here are only exemplary, and this embodiment is not particularly limited here.

[0066] Speech endpoint detection can use the double threshold method to determine the starting and ending points of the speech signal by setting the energy threshold and the zero-crossing rate threshold. It can also be performed by training a speech endpoint recognition model to separate the speech signal from the non-speech signal and obtain a valid speech frame.

[0067] Speech enhancement includes: performing short-time Fourier transform on effective speech frames to obtain one-dimensional spectrum features, performing time-frequency decomposition on the one-dimensional spectrum features, and transforming them into two-dimensional time-frequency domain speech signals.

[0068] By preprocessing the call audio data, the clarity of the voice signal can be enhanced, the impact of noise and interference on the speech conversion and synthesis performance can be effectively reduced, and the efficiency and accuracy of speech recognition can be improved.

[0069] In some optional implementations, the step of performing speech recognition on the enhanced audio data to obtain text data includes:

[0070] Inputting the enhanced audio data into a trained speech recognition model, wherein the speech recognition model includes an acoustic feature extraction layer and a speech recognition layer;

[0071] The enhanced audio data is subjected to feature extraction through an acoustic feature extraction layer to obtain an acoustic feature vector;

[0072] The acoustic feature vector is input into the speech recognition layer for inference and prediction to obtain text data.

[0073] Among them, audio is a typical time series signal. In this embodiment, the acoustic feature extraction layer can use an LSTM network, specifically, a bidirectional Bi-LSTM, and use sequential and reverse recurrent neural networks for the input enhanced audio data to obtain two independent hidden layer representations, and then perform certain calculations (splicing or adding) on ​​the two hidden layer representations to obtain a final hidden layer representation, that is, an acoustic feature vector, and output the acoustic feature vector for subsequent calculations. The acoustic feature vector represented by this hidden layer contains the speech information from both the previous moment and the next moment.

[0074] The speech recognition layer is implemented using the Transformer decoding network. The acoustic feature vector is input into the Transformer decoding network for calculation, and the text output probability distribution corresponding to each frame is decoded. Common decoding algorithms include greedy decoding and beam search decoding, etc., to obtain the corresponding text sequence. Finally, the text sequence is spliced ​​together in chronological order to obtain complete speech recognition text data.

[0075] In this embodiment, in order to achieve real-time speech transcription, it is necessary to synchronize the transcribed text data with the customer's call audio data. Usually, a sliding window is used to perform speech transcription every certain period of time (such as 500 milliseconds), and the transcription result is output synchronously with the corresponding audio clip to ensure the alignment of text and audio. Through real-time speech transcription technology, the work efficiency of customer service can be greatly improved, the workload of customer service can be reduced, and the quality of customer service can be improved.

[0076] In this embodiment, the speech recognition model is trained using a large amount of historical corpus data. The training process of the model continuously adjusts the weight parameters of the network through the back propagation algorithm to minimize the difference between the model's prediction results and the true labels, that is, the loss value no longer changes. Among them, the CTC loss function is used to calculate the loss value between the prediction results and the true labels during the training process of the speech recognition model.

[0077] By using a speech recognition model to perform speech recognition on enhanced audio data, the accuracy, speed, and stability of speech recognition can be improved.

[0078] Step S204, input the text data and target customer characteristics into the trained insurance willingness prediction model to obtain the insurance willingness result.

[0079] In this embodiment, the insurance willingness prediction model can be a CatBoost classification model. The trained CatBoost classification model is used to predict the customer insurance probability based on the text data and the target customer characteristics, and the target customer's insurance willingness is determined based on the predicted customer insurance probability, that is, whether the target customer is willing to insure or not.

[0080] The CatBoost classification model is trained using historical customer data sets. Specifically, key customer features are extracted from the historical customer data sets through feature engineering, and the key customer features are standardized to obtain standard customer features. A symmetrical classification and regression tree (CART) decision tree is constructed based on the standard customer features. Then, the ordered enhancement mode is used in the first stage of constructing the next tree to construct the tree structure. The gradient boosting decision tree (GBDT) algorithm is used in the second stage to construct the tree structure. The root nodes of the two symmetrical CART trees are iteratively split according to the minimum Gini index principle. The mathematical expression of the Gini index is: Among them, Gini(p) represents the obtained Gini index value, N is the probability range of customer insurance, and p is i is the probability that the historical customer data set belongs to the i-th category. The iteration is stopped until the Gini index of the historical customer data set in the node of the CART tree is less than the preset threshold, and finally a symmetrical CART decision tree is generated, which is the CatBoost classification model.

[0081] The text features in the text data are extracted, and the text features and target customer features are used as input features and input into the trained CatBoost classification model to obtain the customer insurance probability. When the customer insurance probability is greater than or equal to the preset probability threshold, it means that the customer has a high willingness to insure, and the insurance willingness result is willing to insure. When the customer insurance probability is less than the preset probability threshold, it means that the customer has a low willingness to insure, and the insurance willingness result is unwilling to insure.

[0082] Step S205, when the insurance intention result is willingness to insure, the text data is subjected to intent recognition to obtain an intent category.

[0083] When the target customer is willing to purchase insurance, the text data is used for intent recognition to obtain the target customer's intent category, where the intent category includes but is not limited to consulting products, complaining about services, and applying for claims. If the intent category is consulting products, the product recommendation script suitable for the target customer is matched from the pre-built service script library based on customer characteristics such as the number of years of insurance, premium amount, and insurance habits; if the intent category is complaint services, the sentiment analysis algorithm is used to judge the complaint sentiment based on the complaint content and the customer's historical service records, and the product information, service terms, etc. related to the complaint content are searched in the knowledge graph, and the product recommendation script suitable for the target customer is matched from the pre-built service script library; when the intent category is applying for claims, the knowledge graph technology is used to obtain information such as insurance policies, cases, and compensation amounts related to the target customer's claims, and the corresponding claims guidance script is matched from the pre-built service script library.

[0084] In some optional implementations, the step of performing intent recognition on text data to obtain an intent category includes:

[0085] Input text data into the trained intent recognition model, where the intent recognition model includes a text embedding layer, a multi-scale feature extraction layer, a fully connected layer, and a classification layer;

[0086] Preprocess and vectorize text data through the text embedding layer to obtain text vectors;

[0087] Input the text vector into the multi-scale feature extraction layer for feature extraction to obtain the hidden states of features at different scales;

[0088] The feature hidden states of different scales are input into the fully connected layer for concatenation to obtain the global features of the text;

[0089] The global features of the text are input into the classification layer for classification prediction to obtain the intent category.

[0090] Among them, the text data is preprocessed through the text embedding layer, including word segmentation, part-of-speech tagging, etc., and then the preprocessed text data is converted into a text vector that the model can understand. The multi-scale feature extraction layer includes at least two connected LSTM layers. Since the call audio data is time series data, the corresponding text data obtained by speech recognition is also time series data. Each LSTM layer processes information of different time scales to capture features of different time scales. The input of the next LSTM layer is the output of the previous LSTM layer and the input text vector. The specific number of LSTM layers can be set according to actual needs. The classification layer can use the Softmax classifier for classification prediction.

[0091] In some optional implementations of this embodiment, each LSTM layer includes a forget gate, an input gate, and an output gate. The LSTM layer overcomes the problems of gradient vanishing and gradient exploding during model prediction by using a mechanism constructed using the input gate, the forget gate, and the output gate, and can capture long-term dependencies in the sequence, achieve long-term memory, and improve the accuracy of classification prediction.

[0092] By identifying the intention of target customers through the intention recognition model, it is possible to capture the long-term dependencies between input text data, improve the efficiency and accuracy of time series feature extraction, and improve the efficiency and accuracy of intent category prediction.

[0093] In some optional implementations, after the step of obtaining the insurance willingness result, the method further includes:

[0094] When the insurance willingness result is unwillingness to insure, a preset number of historical insurance failure cases are matched based on the target customer characteristics and text data;

[0095] Input historical insurance failure cases into the attribution analysis model, analyze historical insurance failure cases through the decision tree algorithm, and obtain the key factors for target customers' insurance failure;

[0096] According to the key factors, the corresponding response dialogue is matched from the service script library.

[0097] Among them, the attribution analysis model can adopt a decision tree model. For example, the ID3 decision tree model is used to analyze 2,000 historical insurance failure cases to find out the key factors for the target customers' insurance failure, such as "high premiums". According to the key factors, the corresponding response dialogue is matched from the service script library. For example, the response dialogue matched to "high premiums" is "We can provide you with a flexible plan for installment payment."

[0098] Through response techniques, personalized service strategies can be dynamically adjusted and applied to subsequent communications with target customers, thereby increasing the policy promotion rate and thus the underwriting success rate.

[0099] Step S206, according to the target customer group portrait label and intent category, retrieve matching target service scripts from a pre-built service script library.

[0100] The service script library is a personalized marketing and promotion service script library designed by business experts based on the characteristics of insurance products and customer group portrait labels of different customer groups.

[0101] In this embodiment, a text similarity algorithm is used to retrieve words that match the target customer group portrait labels and intent categories from a pre-built service speech library. Specifically, the corresponding words are matched according to the intent categories as a service speech candidate set; the TF-IDF algorithm is used to extract the word keywords in the service speech candidate set, and the cosine similarity is used to calculate the correlation score between the word keywords and the target customer group portrait labels, and a preset number of words with the highest scores are selected as the target service words.

[0102] Step S207, input the target customer characteristics and target service words into the trained words recommendation model to obtain the words recommendation results.

[0103] In this embodiment, the speech recommendation model includes a vector embedding layer, a space expansion layer, an encoding layer and a decoding layer. The target customer characteristics and the target service speech are input into the speech recommendation model. The input data is processed in sequence by the vector embedding layer, the space expansion layer, the encoding layer and the decoding layer to generate personalized speech that is most relevant to the current target customer.

[0104] In some optional implementations of this embodiment, the step of inputting the target customer characteristics and the target service speech into the trained speech recommendation model to obtain the speech recommendation result includes:

[0105] The target customer features and target service words are vector-encoded through the vector embedding layer to obtain the feature embedding vector.

[0106] The feature embedding vector is input into the spatial expansion layer, and the feature embedding vector and the introduced pre-trained prompt vector are concatenated to obtain an expanded feature vector;

[0107] The extended feature vector is contextually understood through the encoding layer to obtain a text semantic feature vector containing context information;

[0108] The text semantic feature vector is autoregressively generated through the decoding layer to obtain the speech suggestion result.

[0109] The vector embedding layer is used to convert the target customer features and the target service words into vectors, including position embedding, word embedding and character embedding of the target customer features and the target service words. In this embodiment, the feature embedding vector includes a position embedding vector, a word embedding vector and a character embedding vector.

[0110] A pre-trained learnable hint vector pool P is introduced in the input space, where the hint vector pool is represented as P = {p k ∈R d |k∈N,1≤k≤S}, there are S prompt vectors in the prompt vector pool. Input the feature embedding vector into the spatial expansion layer, and obtain the preset prompt vector pool through the spatial expansion layer. The prompt vector pool contains S prompt vectors, where S is a positive integer; obtain the pre-trained k prompt vectors from the prompt vector pool, where k is a positive integer and less than or equal to S; concatenate the feature embedding vector with the k prompt vectors to obtain the extended feature vector. By introducing the prompt vector, additional information or constraints are actually introduced on the basis of the original feature space, thereby expanding the feature space, allowing the model to capture more features related to speech generation, thereby improving the accuracy of speech generation; at the same time, the generalization ability of the model is enhanced, the computational efficiency is improved, and the training process of the model is optimized.

[0111] The encoding layer can be composed of multiple stacked encoders, each of which includes a multi-head attention layer and a feedforward neural network layer. The output of each encoder is used as the input of the next encoder, and text information at different levels is transmitted between encoders to better capture the contextual information of the context and enrich the representation capability of the features.

[0112] The decoding layer can be composed of multiple decoders connected in stacks and an output layer connected to the last decoder, and each decoder includes a masked multi-head self-attention layer, an encoder-decoder attention layer, and a feedforward network layer. The input text semantic feature vector is calculated by a multi-head attention mechanism through the masked multi-head self-attention layer to obtain a masked attention feature; the masked attention feature is calculated by a cross-attention mechanism through the encoder-decoder attention layer to obtain a global attention feature; the global attention feature is input into the feedforward network layer for speech generation to obtain a predicted speech text sequence; the predicted speech text sequence is mapped to a longer sequence through the output layer, and the text at each position represents the score of the corresponding text, and then the score is converted into a probability through the softmax function, and the preset number of texts with the highest probability are output to obtain the probability distribution of each predicted speech text. Then, a sequence decoding technology such as Beam Search is used to search and score multiple possible predicted texts, and the text with the highest score is selected as the final output speech suggestion result.

[0113] The recommended speech suggestion results are pushed to the agent's interface in real time, guiding the agent to communicate with the target customers in a targeted manner based on the speech suggestion results, thereby improving customer recognition and purchase intention.

[0114] The speech recommendation model generates speech suggestion results based on target customer characteristics and target service speech, which achieves accurate analysis of target customers and intelligent generation of personalized service speech, improves the efficiency and accuracy of speech generation, greatly reduces meaningless communication, improves customer satisfaction and acceptance, and helps to increase the policy promotion rate.

[0115] In some optional implementation methods, if the target customer is ultimately not covered after receiving the recommended results, the attribution analysis model is used to determine the reason for not covering the insurance, and the analysis results are fed back to the background to optimize the customer group segmentation and service script library.

[0116] This application can improve customer satisfaction and acceptance by intelligently providing personalized service scripts based on the specific characteristics, intention categories and insurance willingness of target customers, which helps to increase the policy promotion rate. At the same time, it improves work efficiency, greatly reduces meaningless communication, and thus improves the service skills of agents.

[0117] In some optional implementations, after the step of obtaining the result of the speech suggestion, the method further includes:

[0118] Continuously monitor the feedback data of target customers and update the target customer characteristics based on the feedback data;

[0119] Dynamically adjust the model parameters of the speech recommendation model based on feedback data to obtain an optimized speech recommendation model;

[0120] The updated target customer profile labels are used as input for the next round of services.

[0121] Specifically, during the conversation between the agent and the target customer, feedback data from the target customer is continuously collected. The feedback data includes the target customer's emotional inclination and satisfaction with the suggested speech results. The model parameters of the speech recommendation model are dynamically adjusted according to the feedback data to optimize the speech generation strategy. At the same time, the target customer characteristics are updated and the target customer's customer group affiliation is recalculated.

[0122] Use the updated target customer characteristics and the latest conversation audio data as input for the next round of service, re-execute the above steps S202 to S207, continuously optimize the speech recommendation and service strategy, and form a closed-loop optimization mechanism. This continuous optimization ensures the continuous improvement of service quality and customer satisfaction.

[0123] The embodiments of the present application can acquire and process relevant data based on artificial intelligence technology. Among them, artificial intelligence (AI) is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results.

[0124] AI basic technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing technology, operation / interaction systems, mechatronics, etc. AI software technologies mainly include computer vision technology, robotics technology, biometrics technology, speech processing technology, natural language processing technology, and machine learning / deep learning.

[0125] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through computer-readable instructions, and the computer-readable instructions can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, the aforementioned storage medium can be a non-volatile storage medium such as a disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0126] It should be understood that, although the steps in the flowchart of the accompanying drawings are displayed in sequence as indicated by the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least a part of the steps in the flowchart of the accompanying drawings may include multiple sub-steps or multiple stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be executed in turn or alternately with other steps or at least a part of the sub-steps or stages of other steps.

[0127] Further references Figure 3 , as a response to the above Figure 2 The present application provides an embodiment of a device for generating a recommendation speech. Figure 2 Corresponding to the method embodiment shown, the device can be specifically applied to various electronic devices.

[0128] like Figure 3 As shown, the recommended speech generation device 300 described in this embodiment includes: a division module 301, an acquisition module 302, a speech recognition module 303, a willingness prediction module 304, an intention recognition module 305, a matching module 306 and a speech recommendation module 307. Among them:

[0129] The classification module 301 is used to obtain the historical underwriting data and historical service records of all customers of the insurance company, extract customer features from the historical underwriting data and the historical service records, classify all the customers into different customer groups based on the customer features, and generate customer group portrait labels;

[0130] The acquisition module 302 is used to acquire the call audio data of the target customer, match the corresponding target customer group according to the target customer characteristics of the target customer, and acquire the target customer group portrait label of the target customer group;

[0131] The speech recognition module 303 is used to pre-process the call audio data to obtain enhanced audio data, and perform speech recognition on the enhanced audio data to obtain text data;

[0132] The willingness prediction module 304 is used to input the text data and the target customer characteristics into the trained insurance willingness prediction model to obtain the insurance willingness result;

[0133] The intention recognition module 305 is used to perform intention recognition on the text data to obtain an intention category when the insurance willingness result is willingness to insure;

[0134] The matching module 306 is used to retrieve matching target service words from a pre-built service word library according to the target customer group portrait label and the intention category;

[0135] The speech recommendation module 307 is used to input the target customer characteristics and the target service speech into the trained speech recommendation model to obtain speech suggestion results.

[0136] It should be emphasized that in order to further ensure the privacy and security of historical underwriting data and historical service records, the above historical underwriting data and historical service records can also be stored in a blockchain node.

[0137] Based on the above-mentioned recommendation script generation device 300, by intelligently providing personalized service scripts according to the specific characteristics, intention categories and insurance willingness of the target customers, it is possible to improve customer satisfaction and acceptance, help improve the policy promotion rate, and at the same time, improve work efficiency, greatly reduce meaningless communication, and thus improve the service skills of the agents.

[0138] In some optional implementations, the partitioning module 301 is further configured to:

[0139] The optimal number of clusters K is determined by the silhouette coefficient, where K is a positive integer;

[0140] Convert the customer features of each customer into a customer clustering vector, and select K customer clustering vectors from all the customer clustering vectors as clustering centers;

[0141] Calculate the distance between other clustering vectors and each of the clustering centers, and assign other clustering vectors to the clustering center with the closest distance according to the distance to form K cluster clusters, wherein the other clustering vectors are customer clustering vectors other than the clustering center;

[0142] The centroid of each cluster is calculated, and the centroid is used as the new cluster center. The distance between the other cluster vectors and each cluster center is repeatedly calculated to obtain K new clusters. The cluster centers of each cluster are continuously updated until the clustering stop condition is met to obtain the final K customer groups.

[0143] By dividing customers into different customer groups, we can accurately locate the groups, formulate targeted marketing strategies and service plans for each customer group, and provide personalized services, which will help improve the policy promotion rate. At the same time, it will help to more accurately assess the risk levels of different customer groups, optimize resource allocation, cut unnecessary costs and expenditures, and improve overall operational efficiency.

[0144] In some optional implementations, the speech recognition module 303 includes:

[0145] A speech input submodule, used for inputting the enhanced audio data into a trained speech recognition model, wherein the speech recognition model includes an acoustic feature extraction layer and a speech recognition layer;

[0146] A feature extraction submodule, configured to extract features from the enhanced audio data through the acoustic feature extraction layer to obtain an acoustic feature vector;

[0147] The speech recognition submodule is used to input the acoustic feature vector into the speech recognition layer for inference prediction to obtain text data.

[0148] By using a speech recognition model to perform speech recognition on enhanced audio data, the accuracy, speed, and stability of speech recognition can be improved.

[0149] In some optional implementations, the intention identification module 305 includes:

[0150] A text input submodule, used to input the text data into a trained intent recognition model, wherein the intent recognition model includes a text embedding layer, a multi-scale feature extraction layer, a fully connected layer, and a classification layer;

[0151] A text embedding submodule, used for preprocessing and vectorizing the text data through the text embedding layer to obtain a text vector;

[0152] A multi-scale feature extraction submodule, used for inputting the text vector into the multi-scale feature extraction layer for feature extraction to obtain feature hidden states of different scales;

[0153] A splicing submodule, used for inputting the feature hidden states of different scales into the fully connected layer for splicing to obtain the global features of the text;

[0154] The classification submodule is used to input the global features of the text into the classification layer for classification prediction to obtain the intent category.

[0155] By identifying the intention of target customers through the intention recognition model, it is possible to capture the long-term dependencies between input text data, improve the efficiency and accuracy of time series feature extraction, and improve the efficiency and accuracy of intent category prediction.

[0156] In some optional implementations, the recommendation speech generating device 300 further includes an attribution analysis module, which is used to:

[0157] When the insurance willingness result is unwillingness to insure, matching a preset number of historical insurance failure cases based on the target customer characteristics and the text data;

[0158] Input the historical insurance failure cases into the attribution analysis model, analyze the historical insurance failure cases through a decision tree algorithm, and obtain the key factors of the target customer's insurance failure;

[0159] According to the key factors, corresponding response dialogues are matched from the service dialogue library.

[0160] Through response techniques, personalized service strategies can be dynamically adjusted and applied to subsequent communications with target customers, thereby increasing the policy promotion rate and thus the underwriting success rate.

[0161] In some optional implementations, the speech recommendation model includes a vector embedding layer, a spatial expansion layer, an encoding layer, and a decoding layer, and the speech recommendation module 307 includes:

[0162] A vector embedding submodule, used to perform vector encoding on the target customer features and the target service words through the vector embedding layer to obtain a feature embedding vector;

[0163] An expansion submodule, used for inputting the feature embedding vector into the spatial expansion layer, concatenating the feature embedding vector and the introduced pre-trained prompt vector to obtain an expanded feature vector;

[0164] An encoding submodule, used for performing context understanding on the extended feature vector through the encoding layer to obtain a text semantic feature vector containing context information;

[0165] The decoding submodule is used to generate the text semantic feature vector through the decoding layer through autoregression to obtain the speech suggestion result.

[0166] The speech recommendation model generates speech suggestion results based on target customer characteristics and target service speech, which achieves accurate analysis of target customers and intelligent generation of personalized service speech, improves the efficiency and accuracy of speech generation, greatly reduces meaningless communication, improves customer satisfaction and acceptance, and helps to increase the policy promotion rate.

[0167] In some optional implementations, the recommendation speech generation device 300 further includes a feedback optimization module, which is used to:

[0168] Continuously monitoring the feedback data of the target customers, and updating the target customer characteristics according to the feedback data;

[0169] Dynamically adjusting the model parameters of the speech recommendation model based on the feedback data to obtain an optimized speech recommendation model;

[0170] The updated target customer characteristics are used as input for the next round of services.

[0171] By continuously optimizing the recommended words and service strategies, a closed-loop optimization mechanism is formed. This continuous optimization ensures the continuous improvement of service quality and customer satisfaction.

[0172] To solve the above technical problems, the present application also provides a computer device. Figure 4 , Figure 4 This is a basic structural block diagram of the computer device in this embodiment.

[0173] The computer device 4 includes a memory 41, a processor 42, and a network interface 43 that are interconnected and communicated through a system bus. It should be noted that the figure only shows a computer device 4 with a memory 41, a processor 42, and a network interface 43, but it should be understood that it is not required to implement all the components shown, and more or fewer components can be implemented instead. Among them, those skilled in the art can understand that the computer device here is a device that can automatically perform numerical calculations and / or information processing according to pre-set or stored instructions, and its hardware includes but is not limited to microprocessors, application specific integrated circuits (Application Specific Integrated Circuit, ASIC), programmable gate arrays (Field-Programmable Gate Array, FPGA), digital processors (Digital Signal Processor, DSP), embedded devices, etc.

[0174] The computer device may be a computing device such as a desktop computer, a notebook, a PDA, a cloud server, etc. The computer device may interact with a user through a keyboard, a mouse, a remote controller, a touch pad, or a voice control device.

[0175] The memory 41 includes at least one type of readable storage medium, and the readable storage medium includes flash memory, hard disk, multimedia card, card-type memory (for example, SD or DX memory, etc.), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, disk, optical disk, etc. In some embodiments, the memory 41 can be an internal storage unit of the computer device 4, such as a hard disk or memory of the computer device 4. In other embodiments, the memory 41 can also be an external storage device of the computer device 4, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (FlashCard), etc. equipped on the computer device 4. Of course, the memory 41 can also include both the internal storage unit of the computer device 4 and its external storage device. In this embodiment, the memory 41 is generally used to store the operating system and various application software installed on the computer device 4, such as computer-readable instructions of the recommendation speech generation method, etc. In addition, the memory 41 can also be used to temporarily store various types of data that have been output or are to be output.

[0176] The processor 42 may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip in some embodiments. The processor 42 is generally used to control the overall operation of the computer device 4. In this embodiment, the processor 42 is used to run computer-readable instructions stored in the memory 41 or process data, such as computer-readable instructions for running the recommendation speech generation method.

[0177] The network interface 43 may include a wireless network interface or a wired network interface. The network interface 43 is generally used to establish a communication connection between the computer device 4 and other electronic devices.

[0178] By intelligently providing personalized service scripts based on the specific characteristics, intent categories and insurance willingness of target customers, it is possible to improve customer satisfaction and acceptance, help increase the policy promotion rate, and at the same time, improve work efficiency, greatly reduce meaningless communication, and thus improve the service skills of agents.

[0179] The present application also provides another embodiment, namely, providing a computer-readable storage medium, wherein the computer-readable storage medium stores computer-readable instructions, and the computer-readable instructions can be executed by at least one processor to enable the at least one processor to perform the steps of the recommendation speech generation method as described above.

[0180] By intelligently providing personalized service scripts based on the specific characteristics, intent categories and insurance willingness of target customers, it is possible to improve customer satisfaction and acceptance, help increase the policy promotion rate, and at the same time, improve work efficiency, greatly reduce meaningless communication, and thus improve the service skills of agents.

[0181] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes a number of instructions for a terminal device (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in each embodiment of the present application.

[0182] Obviously, the embodiments described above are only some embodiments of the present application, rather than all embodiments. The preferred embodiments of the present application are given in the accompanying drawings, but they do not limit the patent scope of the present application. The present application can be implemented in many different forms. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present application more thorough and comprehensive. Although the present application is described in detail with reference to the aforementioned embodiments, for those skilled in the art, it is still possible to modify the technical solutions recorded in the aforementioned specific implementation methods, or to perform equivalent replacement of some of the technical features therein. Any equivalent structure made using the contents of the specification and drawings of this application, directly or indirectly used in other related technical fields, is similarly within the scope of patent protection of this application.

Claims

1. A method for generating recommendation words, characterized in that: The steps include: Obtaining historical underwriting data and historical service records of all customers of the insurance company, extracting customer features from the historical underwriting data and the historical service records, dividing all customers into different customer groups based on the customer features, and generating customer group portrait labels; Acquire call audio data of target customers, match corresponding target customer groups according to target customer characteristics of the target customers, and acquire target customer group portrait labels of the target customer groups; Preprocessing the call audio data to obtain enhanced audio data, and performing speech recognition on the enhanced audio data to obtain text data; Inputting the text data and the target customer characteristics into a trained insurance willingness prediction model to obtain an insurance willingness result; When the insurance willingness result is willingness to insure, performing intent recognition on the text data to obtain an intent category; According to the target customer group portrait label and the intent category, a matching target service script is retrieved from a pre-built service script library; The target customer characteristics and the target service words are input into the trained words recommendation model to obtain the words recommendation results.

2. The method for generating recommendation words according to claim 1, characterized in that: The step of dividing all the customers into different customer groups based on the customer characteristics comprises: The optimal number of clusters K is determined by the silhouette coefficient, where K is a positive integer; Convert the customer features of each customer into a customer clustering vector, and select K customer clustering vectors from all the customer clustering vectors as clustering centers; Calculate the distance between other clustering vectors and each of the clustering centers, and assign other clustering vectors to the clustering center with the closest distance according to the distance to form K cluster clusters, wherein the other clustering vectors are customer clustering vectors other than the clustering center; The centroid of each cluster is calculated, and the centroid is used as the new cluster center. The distance between the other cluster vectors and each cluster center is repeatedly calculated to obtain K new clusters. The cluster centers of each cluster are continuously updated until the clustering stop condition is met to obtain the final K customer groups.

3. The method for generating recommendation words according to claim 1, characterized in that: The step of performing speech recognition on the enhanced audio data to obtain text data comprises: Inputting the enhanced audio data into a trained speech recognition model, wherein the speech recognition model includes an acoustic feature extraction layer and a speech recognition layer; Extracting features from the enhanced audio data through the acoustic feature extraction layer to obtain an acoustic feature vector; The acoustic feature vector is input into the speech recognition layer for inference prediction to obtain text data.

4. The method for generating recommendation words according to claim 1, characterized in that: The step of performing intent recognition on the text data to obtain the intent category comprises: Inputting the text data into a trained intent recognition model, wherein the intent recognition model includes a text embedding layer, a multi-scale feature extraction layer, a fully connected layer, and a classification layer; Preprocessing and vectorizing the text data through the text embedding layer to obtain a text vector; Inputting the text vector into the multi-scale feature extraction layer for feature extraction to obtain feature hidden states of different scales; Inputting the feature hidden states of different scales into the fully connected layer for splicing to obtain the global features of the text; The global features of the text are input into the classification layer for classification prediction to obtain the intent category.

5. The method for generating recommendation words according to claim 1, characterized in that: After the step of obtaining the insurance willingness result, the method further includes: When the insurance willingness result is unwillingness to insure, matching a preset number of historical insurance failure cases based on the target customer characteristics and the text data; Input the historical insurance failure cases into the attribution analysis model, analyze the historical insurance failure cases through a decision tree algorithm, and obtain the key factors of the target customer's insurance failure; According to the key factors, corresponding response dialogues are matched from the service dialogue library.

6. The method for generating recommendation words according to claim 1, characterized in that: The speech recommendation model includes a vector embedding layer, a space expansion layer, an encoding layer and a decoding layer. The step of inputting the target customer characteristics and the target service speech into the trained speech recommendation model to obtain the speech suggestion result includes: Performing vector encoding on the target customer features and the target service words through the vector embedding layer to obtain a feature embedding vector; Inputting the feature embedding vector into the spatial expansion layer, concatenating the feature embedding vector and the introduced pre-trained prompt vector to obtain an expanded feature vector; Performing context understanding on the extended feature vector through the encoding layer to obtain a text semantic feature vector containing context information; The text semantic feature vector is autoregressively generated through the decoding layer to obtain a speech suggestion result.

7. The method for generating recommendation words according to claim 1, characterized in that: After the step of obtaining the result of the speech suggestion, the method further includes: Continuously monitoring the feedback data of the target customers, and updating the target customer characteristics according to the feedback data; Dynamically adjusting the model parameters of the speech recommendation model based on the feedback data to obtain an optimized speech recommendation model; The updated target customer characteristics are used as input for the next round of services.

8. A device for generating recommendation words, characterized in that: include: A classification module is used to obtain the historical underwriting data and historical service records of all customers of the insurance company, extract customer features from the historical underwriting data and the historical service records, classify all the customers into different customer groups based on the customer features, and generate customer group portrait labels; An acquisition module is used to acquire call audio data of a target customer, match a corresponding target customer group according to the target customer characteristics of the target customer, and acquire a target customer group portrait label of the target customer group; A speech recognition module, used for preprocessing the call audio data to obtain enhanced audio data, and performing speech recognition on the enhanced audio data to obtain text data; A willingness prediction module, used to input the text data and the target customer characteristics into a trained insurance willingness prediction model to obtain an insurance willingness result; An intention recognition module, used for performing intention recognition on the text data to obtain an intention category when the insurance willingness result is willingness to insure; A matching module, used to retrieve matching target service words from a pre-built service word library according to the target customer group portrait label and the intent category; The speech recommendation module is used to input the target customer characteristics and the target service speech into the trained speech recommendation model to obtain the speech recommendation results.

9. A computer device, characterized in that: It includes a memory and a processor, wherein the memory stores computer-readable instructions, and when the processor executes the computer-readable instructions, the steps of the method for generating recommendation words as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-readable instructions, and when the computer-readable instructions are executed by a processor, the steps of the method for generating recommendation words as described in any one of claims 1 to 7 are implemented.

Images