WO2019184117A1

WO2019184117A1 - Response model training method, smart chat method, apparatuses, device and medium

Info

Publication number: WO2019184117A1
Application number: PCT/CN2018/094177
Authority: WO
Inventors: 金戈; 徐亮; 肖京
Original assignee: 平安科技（深圳）有限公司
Priority date: 2018-03-26
Filing date: 2018-07-03
Publication date: 2019-10-03
Also published as: CN108632137A

Abstract

Disclosed in the present application are a response model training method, a smart chat method, apparatuses, a device and a medium. The response model training method comprises: acquiring original training text data; preprocessing the original training text data, to acquire target training text data; dividing the target training text data according to a preset proportion, to acquire a training set and a test set; using an RBM model to train the training set, to acquire an original response model; using the test set to test the original response model, to acquire a target response model. The response model training method effectively solves the existing problem that professional consulting problems cannot be automatically responded, reducing labor cost, and improving efficiency.

Description

Response model training method, intelligent chat method, device, device and medium

This patent application is based on the Chinese invention patent application filed on March 26, 2018, with the application number of 201110250162.9, entitled "Response model training method, intelligent chat method, device, device and medium", and requires its priority.

Technical field

The present application relates to the field of artificial intelligence, and in particular, to a response model training method, an intelligent chat method, an apparatus, a device, and a medium.

Background technique

With the development of WeChat, more and more companies choose to adopt WeChat as an important way of business promotion. When financial institutions such as insurance, securities, and banks use WeChat for business promotion, they usually need to manually respond to the problem of customers consulting through WeChat, resulting in high labor costs and low efficiency. The reason is that when the current financial institution conducts business promotion based on WeChat, when the customer uses the voice chat method to consult the financial service, the voice information cannot be automatically recognized. Moreover, even if voice information is recognized, the problem of consulting involves professional issues in the fields of insurance, securities, and banking. Professionals need to respond based on their own professional knowledge. Therefore, it is necessary to equip a large number of human resources to respond to customer consultations, resulting in labor costs. High, and when multiple customers consult the same question, it may be replied by different professionals, resulting in duplication of effort and making it inefficient.

Summary of the invention

The embodiment of the present application provides a response model training method, device, device and medium to train a response model for a professional problem, so as to solve the problem that the automatic answering problem cannot be automatically addressed to a professional consulting problem.

The embodiment of the present invention provides an intelligent chat method, device, device, and medium for implementing voice recognition and automatic response on WeChat, so as to solve the current high labor cost and efficiency of a professional voice replying to a voice consultation service based on a WeChat promotion service. Low problem.

The embodiment of the present application provides a response model training method, including:

Obtain the original training text data;

Performing pre-processing on the original training text data to obtain target training text data;

And dividing the target training text data into a preset ratio to obtain a training set and a test set;

The training set is trained by using an RBM model to obtain an original response model;

The original response model is tested using the test set to obtain a target response model.

The embodiment of the present application provides a response model training apparatus, including:

The original training text data acquiring module is configured to obtain original training text data;

a target training text data acquiring module, configured to preprocess the original training text data to obtain target training text data;

a target training text data dividing module, configured to divide the target training text data according to a preset ratio, and acquire a training set and a test set;

An original response model acquisition module, configured to train the training set by using an RBM model, and obtain an original response model;

The target response model acquisition module is configured to test the original response model by using the test set to obtain a target response model.

An embodiment of the present application provides a computer device including a memory, a processor, and computer readable instructions stored in the memory and executable on the processor, the processor implementing the computer readable instructions The following steps:

Obtain the original training text data;

Embodiments of the present application provide one or more non-volatile readable storage media storing computer readable instructions, when executed by one or more processors, causing the one or more processors Perform the following steps:

Obtain the original training text data;

The embodiment of the present application provides a smart chat method, including:

Calling the information acquisition interface of the WeChat web version to obtain the WeChat message;

If the WeChat message is a voice message, the third party voice model is called to identify the voice message, and the recognized text data is obtained;

If the WeChat message is a text message, directly acquiring the recognized text data;

Inputting the identification text data into the target response model, acquiring corresponding response information, and calling the information sending interface of the WeChat webpage to send the response information;

The target response model is a model obtained by training using the response model training method described in the present application.

The embodiment of the present application provides a smart chat device, including:

a WeChat message obtaining module, configured to invoke an information obtaining interface of a WeChat webpage to obtain a WeChat message;

a first identification text data obtaining module, configured to: if the WeChat message is a voice message, invoke a third-party voice model to identify the voice message, and obtain the identification text data;

a second identification text data obtaining module, configured to directly obtain the identification text data if the WeChat message is a text message;

a response information obtaining and sending module, configured to input the recognized text data into the target response model, obtain corresponding response information, and invoke an information sending interface of a WeChat webpage to send the response information;

The target response model is a model obtained by training using the response model training method.

Details of one or more embodiments of the present application are set forth in the accompanying drawings and description below. Other features and advantages of the present invention will be apparent from the description, drawings and claims.

DRAWINGS

In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings used in the description of the embodiments of the present application will be briefly described below. It is obvious that the drawings in the following description are only some embodiments of the present application. Other drawings may also be obtained from those of ordinary skill in the art based on these drawings without the inventive labor.

1 is a flowchart of a response model training method provided in Embodiment 1 of the present application;

Figure 2 is a specific schematic view of step S12 of Figure 1;

Figure 3 is a specific schematic view of step S123 of Figure 2;

Figure 4 is a specific schematic view of step S14 of Figure 1;

Figure 5 is a specific schematic view of step S15 of Figure 1;

6 is a schematic block diagram of a response model training apparatus provided in Embodiment 2 of the present application;

7 is a flowchart of a smart chat method provided in Embodiment 3 of the present application;

8 is a schematic block diagram of a smart chat device provided in Embodiment 4 of the present application;

FIG. 9 is a schematic diagram of a computer device 1 provided in Embodiment 6 of the present application.

detailed description

The technical solutions in the embodiments of the present application are clearly and completely described in the following with reference to the drawings in the embodiments of the present application. It is obvious that the described embodiments are a part of the embodiments of the present application, and not all of the embodiments. All other embodiments obtained by a person of ordinary skill in the art based on the embodiments of the present application without departing from the inventive scope are the scope of the present application.

Example 1

Fig. 1 is a flow chart showing a response model training method in this embodiment. The response model training method can be applied to computer equipments of financial institutions such as insurance, securities, and banks or other institutions for training response models to achieve intelligent response purposes. In this embodiment, the response model of the training insurance service is taken as an example for description, so that the trained response model can be applied to the service promotion process of the insurance institution, and the problem of customer consultation is automatically answered, thereby improving the response efficiency. As shown in FIG. 1, the response model training method includes the following steps:

S11: Acquire original training text data.

The original training text data includes, but is not limited to, corpus data in a specific domain corpus. The specific field in this embodiment refers specifically to the field of insurance, and the domain-specific corpus specifically refers to a text library with the theme of insurance business. Among them, the corpus data refers to the linguistic material data that has actually appeared in the actual use of the language. Specifically, the original training text data includes training questions and corresponding training answers, and the training questions and training answers are labeled in advance. For example: Under the theme of growth accident insurance, the age of insurance (training problem): 3-18 years old (training answer). The training response model is obtained based on the acquisition of the original training text data, so that the response model can perform deep learning based on the original training text data, thereby achieving the purpose of the intelligent response.

S12: Pre-processing the original training text data to obtain target training text data.

The pre-processing includes, but is not limited to, Chinese and English recognition, word segmentation processing, and vectorization processing. Chinese and English recognition refers to the distinction between Chinese characters and English characters for word segmentation. Word segmentation refers to the process of segmenting words in a sentence according to a dictionary. Vectorization processing refers to the process of vectorizing representations of sentences. Specifically, when the neural network model trains the text data, the text cannot be directly trained, and the original training text data needs to be preprocessed to obtain the target training text data represented by the vectorization, so as to input the target training text data. Train to the neural network model.

S13: The target training text data is divided according to a preset ratio, and the training set and the test set are obtained.

The training set is a learning sample data set. The classifier is built by matching some parameters, that is, the target training text data in the training set is used to train the machine learning model to determine the parameters of the machine learning model. The test set is used to test the resolving power of a trained machine learning model, such as recognition rate or accuracy. Specifically, the data is divided by a ten-fold cross-validation method to ensure the accuracy of the response model training. The ten-fold cross-validation method is a commonly used method for testing the accuracy of an algorithm. In this embodiment, the ten-fold cross-validation method is used to divide the data by specifically classifying the target training text data according to a ratio of 9:1, and the target training text data can be divided into 10 groups, and 9 groups of target training are performed. The text data is used as a training set, and the remaining 1 set of target training text data is used as a test set.

S14: The training set is trained by the RBM model to obtain the original response model.

The original response model is a model obtained by training the target training text data in the training set by the RBM model. The RBM (Restricted Boltzmann Machine) model is an undirected graph model consisting of a visible layer and a hidden layer. The RBM model includes several neurons, each of which is a binary unit. That is, the value of each neuron can only be 0 or 1. Moreover, each neuron of the visible layer is connected to each neuron of the hidden layer; but between the neurons of the visible layer, there is no connecting line between the neurons of the hidden layer, that is, between the neurons of the same layer Independent of each other, each visible layer of neurons is only affected by neurons in the hidden layer, and has the advantages of fast convergence and small prediction error. In this embodiment, the RBM model is used to train the original response model, which has the advantages of high training efficiency and high accuracy.

S15: The original response model is tested by using a test set to obtain a target response model.

The target response model is a model that tests the original response model with a test set to make the accuracy of the original response model reach a preset accuracy. Specifically, the original response model is tested by using the target training text data in the test set to obtain a corresponding accuracy rate; if the accuracy reaches the preset accuracy, the original response model is used as the target response model.

In this embodiment, the original training text data is first acquired to preprocess the original training text data, and the target training text data is acquired, so that the target training text data is input to the neural network model for training. Then, the ten-fold cross-validation method is used to divide the target training text, and the training set and test set are obtained to ensure the accuracy of the target response model obtained by the training. Then the RBM model is used to train the training set, and the original response model is obtained, which improves the training efficiency and accuracy of the original response model. Finally, the original response model is tested with the test set to obtain the target response model and improve the accuracy of the response model.

In a specific embodiment, as shown in FIG. 2, in step S12, the original training text data is preprocessed to obtain the target training text data, which specifically includes the following steps:

S121: Perform original Chinese and English recognition on the original training text data to obtain the identification text.

The recognized text refers to the text obtained by distinguishing Chinese characters and English characters in the original training text data. Since Chinese and/or English may appear in the original training text data, the operation of Chinese word segmentation and English word segmentation is different in subsequent segmentation, so it needs to be distinguished.

In this embodiment, the method for performing Chinese and English recognition on the original training text data includes, but is not limited to, a regular expression. Among them, the regular expression describes a pattern of string matching, which can be used to check whether a string contains a certain substring, replace the matched substring, or take a substring conforming to a certain condition from a string. Specifically, the method for recognizing Chinese and English by using regular expressions is as follows: the regular expression matching Chinese characters is [u4e00-u9fa5], and the regular expression matching English characters is [a-zA-Z]. Regular expressions based on Chinese characters and regular expressions of English characters are used to identify the original training text data in Chinese and English to obtain corresponding recognition texts (including Chinese characters and English characters), so that the word segmentation can be quickly performed when the subsequent word segmentation is performed. Operation to improve the efficiency of model training.

Further, the recognized English characters can also be mapped to English characters by using a pre-stored Chinese-English comparison table to obtain converted Chinese characters, thereby improving the generalization ability of the model. At this time, the recognized text includes Chinese characters and converted Chinese characters mapped by English characters.

S122: Perform word segmentation on the recognized text to obtain at least one word.

Among them, the word is the word element obtained after the word segmentation is performed. In this embodiment, the method for segmenting the recognized text includes, but is not limited to, using the staging word segmentation tool to segment the Chinese characters of the recognized text. The stuttering word segmentation tool is a commonly used Chinese analysis tool, which can effectively extract the words in the sentence one by one, and has the advantages of high accuracy and high efficiency. In this embodiment, since the staging word segmentation tool is a tool for segmenting Chinese characters, the English characters recognized in step S121 can be mapped to English characters by using a pre-stored Chinese-English comparison table to obtain Chinese characters, and then adopted. The stuttering word segmentation tool performs word segmentation to improve the generalization ability of the model.

S123: Perform vectorization processing on at least one word to obtain target training text data.

The target training text data is text data obtained by performing vectorization processing on at least one word. Specifically, the TDF-IF algorithm is used to calculate the weight of each word in the original training text data, and is used as a dimension of the vector to realize vectorized representation of the sentence for at least one word, and obtain the target. Train text data to facilitate the training process of the model and speed up the training of the model.

In this embodiment, the regular expression is used to distinguish between Chinese and English, and the recognition text is obtained, so that the word segmentation tool is used to segment the recognized text and obtain the word order, so as to improve the accuracy and training efficiency of the model. Before the word segmentation, the Chinese and English comparison tables can be used to map the recognized English characters, and the Chinese characters can be converted, so that the Chinese characters can be segmented by using the staging word segmentation tool to improve the generalization ability of the model. Finally, the at least one word is vectorized to obtain the target training text data, which provides convenience for the input of the subsequent response model training.

In a specific embodiment, as shown in FIG. 3, in step S123, the vectorization processing is performed on at least one word to obtain the target training text data, which specifically includes the following steps:

S1231: Perform at least one word operation by using the TF-IDF algorithm to obtain a word frequency corresponding to each word.

Among them, the TF-IDF (term frequency–inverse document frequency) algorithm is a commonly used weighting algorithm for information retrieval and data mining, which has the advantages of simple calculation and high efficiency. Specifically, each word is operated by using the TF-IDF algorithm to obtain the number of occurrences of each word in the original training text data, that is, the word frequency. The calculation formula of the TF-IDF algorithm is

Where u is the number of occurrences of the word in the original training text data, U is the total number of words in the original training text data, and T is the word frequency. In this embodiment, the TF-IDF algorithm is used to calculate at least one word, and the word frequency corresponding to each word is obtained, and the calculation process is simple, which is beneficial to improving the training efficiency of the response model.

S1232: The word frequency corresponding to each word is used as the dimension of the vector, and the target training text data represented by the vector is obtained.

Specifically, the word frequency corresponding to each word is taken as one dimension of the vector, and the target training text data represented by the vector is acquired. For example, the original training text data is “insurance term (training problem)-1 year (training answer)”, and the word obtained after segmentation of the original training text data is “insurance”, “term”, “1 year”, hypothesis The word frequency of each word calculated by step S1231 is 0.2, 0.3, and 0.4, and the target training text data obtained by vectorizing the word is (0.2, 0.3, 0.4), so as to facilitate the input model for training. Thereby improving the training efficiency of the response model. Among them, the training questions and training answers are pre-marked.

In this embodiment, the TF-IDF algorithm is first used to calculate each word order to obtain the number of occurrences of each word in the original training text data, that is, the word frequency, and the calculation process is simple, which is beneficial to improving the training efficiency of the response model. Then, the word frequency corresponding to each word is taken as a dimension of the vector, and the target training text data represented by the vector is obtained, so as to input the model for training, thereby improving the training efficiency of the response model.

In a specific implementation, as shown in FIG. 4, in step S14, the RBM model is used to train the training set to obtain the original response model, which specifically includes the following steps:

S141: Initialize model parameters.

As mentioned above, the RBM model is an undirected graph model consisting of a visible layer and a hidden layer. Initializing the model parameters in step S141 specifically refers to initializing the model parameters associated with the visible layer and the hidden layer in the RBM model. The model parameters include the model cycle, the weight matrix of the visible layer to the hidden layer, the offset of the visible layer, the offset of the hidden layer, the number of neurons in the visible layer, the number of neurons in the hidden layer, the learning rate, and Compare the number of iterations corresponding to the divergence algorithm.

S142: Optimize the model parameters by using the contrast divergence algorithm to obtain the original response model; wherein the formula of the contrast divergence algorithm is: CDK(k, S, W, a, b; ΔW, Δa, Δb); The number of iterations of the divergence algorithm; S is the training set; W is the weight matrix; a is the offset vector of the visible layer; b is the offset vector of the hidden layer; ΔW is the rate of change of the weight matrix; Δa is the visible layer bias The rate of change of the vector, Δb, is the rate of change of the hidden layer offset vector.

In this embodiment, the RBM model is implemented by using a contrast divergence algorithm. Among them, the contrast divergence (CD) algorithm proposed by Hinton can effectively perform RBM learning, and can avoid the trouble of obtaining log-likelihood function gradient. Therefore, it is widely used in depth model based on RBM construction, usually It only needs to be iterated once to get an optimized model, which improves the training efficiency of the model.

Specifically, the CDK (k, S, W, a, b; ΔW, Δa, Δb) is used to train the target training text data in the training set to obtain the change rate of the weight matrix and the visible layer bias. The rate of change of the vector and the rate of change of the hidden layer offset vector.

Specifically, it is assumed that the training problem vector X=(x ₁ , x ₂ ... x _m ) in the target training text data is calculated to obtain the training answer vector Y=(y ₁ , y ₂ ... y _n ), which can be understood. For the RBM model, an m-dimensional training problem vector is mapped to an n-dimensional training answer vector, and the calculation formula for obtaining the hidden unit value of 1 probability includes

Where v is the input of the visible layer (ie vector X), a _i represents the offset of the i-th visible element, and the sigmoid function is a function of the S-type common in biology. In information science, due to its single increase As well as the inverse function of the inverse function, the Sigmoid function is often used as a threshold function of the neural network, mapping the variables between 0, 1. Then, the random function is used to generate a random number of [0, 1]. If the random number is less than P(h _j =1|v), the hidden unit takes a value of 1. Then use the formula

Calculate the probability that the visible unit takes a value of 1, and then reconstruct the visible layer; where h _i is Y in y _n and b _j represents the offset of the j-th hidden neuron. Then, using the formula

The reconstructed RBM model is optimized to obtain the optimized model parameters, and then the target response model is obtained; where W is the weight matrix; a is the offset vector of the visible layer; b is the offset vector of the hidden layer; ΔW is the weight The rate of change of the value matrix; Δa is the rate of change of the visible layer offset vector, Δb is the rate of change of the hidden layer offset vector, and η is the learning rate. The visible unit in this embodiment refers to a neuron in the visible layer, and the hidden unit refers to a neuron in the hidden layer.

In this embodiment, the model parameters are initialized first, and the model parameters include a model cycle period, a weight matrix of the visible layer to the hidden layer, a bias of the visible layer, a bias of the hidden layer, a number of neurons in the visible layer, and a hidden layer. The number of neurons in the middle, the learning rate, and the number of iterations corresponding to the contrast divergence algorithm. Then, the contrast divergence algorithm is used to obtain the probability that each visible unit in the visible layer takes a value of 1, and then the reconstruction of the visible layer is obtained, the calculation amount is reduced, and the training efficiency is improved. Moreover, since the contrast divergence algorithm usually only needs to be iterated once to obtain an optimized model, the training efficiency of the original response model is improved.

In a specific implementation, as shown in FIG. 5, in step S15, the original response model is tested by using the test set, and the target response model is obtained, which specifically includes the following steps:

S151: Testing the original response model with a test set to obtain test accuracy.

Specifically, each target training text data in the training set is input to the RBM model for iterative training to obtain a corresponding original response model, and then the obtained original response model is tested by using the target training text data in the test set to obtain a corresponding Test accuracy.

S152: Acquire a target response model if the test accuracy is not less than the preset accuracy.

If the test accuracy is not less than the preset accuracy, the training is stopped, and the target response model is obtained. If the test accuracy is less than the preset accuracy, the steps S14-S15 are continued until the test accuracy corresponding to the original response model reaches the preset. Accuracy up to the accuracy of the target response model.

In this embodiment, the original response model obtained by iteratively training each group of target training text data by using the RMB model is tested by using the target training text data in the test set, and the test accuracy is obtained, if the test accuracy is not less than the pre-predetermined If the accuracy is set, the training is stopped, and the target response model is obtained. If the test accuracy is less than the preset accuracy, the steps S14-S15 are continued, until the test accuracy corresponding to the original response model reaches the preset accuracy, so as to improve the target response. The accuracy of the model.

It should be understood that the size of the sequence of the steps in the above embodiments does not mean that the order of execution is performed. The order of execution of each process should be determined by its function and internal logic, and should not be construed as limiting the implementation process of the embodiments of the present application.

Example 2

Fig. 6 is a block diagram showing the principle of the response model training device corresponding to the response model training method of the first embodiment. As shown in FIG. 6, the response model training device includes an original training text data acquisition module 11, a target training text data acquisition module 12, a target training text data division module 13, an original response model acquisition module 14, and a target response model acquisition module 15. The implementation function of the original training text data acquisition module 11, the target training text data acquisition module 12, the target training text data division module 13, the original response model acquisition module 14 and the target response model acquisition module 15 and the response model training method in the embodiment Corresponding steps correspond one-to-one, and in order to avoid redundancy, the present embodiment will not be described in detail.

The original training text data obtaining module 11 is configured to acquire original training text data.

The target training text data obtaining module 12 is configured to preprocess the original training text data to obtain the target training text data.

The target training text data dividing module 13 is configured to divide the target training text data according to a preset ratio to obtain a training set and a test set.

The original response model acquisition module 14 is configured to train the training set by using the RBM model to obtain the original response model.

The target response model acquisition module 15 is configured to test the original response model by using the test set to obtain a target response model.

Preferably, the target training text data acquisition module 12 includes an identification text acquisition unit 121, a word acquisition unit 122, and a target training text data acquisition unit 123.

The identification text obtaining unit 121 is configured to perform the Chinese and English recognition on the original training text data to obtain the identification text.

The word acquisition unit 122 is configured to perform word segmentation on the recognized text to obtain at least one word.

The target training text data acquiring unit 123 is configured to perform vectorization processing on at least one word to obtain target training text data.

Preferably, the target training text data acquisition unit 123 includes a word frequency acquisition sub-unit 1231 and a target training text data acquisition sub-unit 1232.

The word frequency acquisition sub-unit 1231 is configured to perform at least one word operation by using the TF-IDF algorithm to obtain a word frequency corresponding to each word.

The target training text data obtaining sub-unit 1232 is configured to obtain the target training text data represented by the vector form by using the word frequency corresponding to each word as the dimension of the vector.

Preferably, the original response model acquisition module 14 includes a parameter initialization unit 141 and an original response model acquisition unit 142.

The parameter initialization unit 141 is configured to initialize the model parameters.

The original response model obtaining unit 142 is configured to optimize the model parameters by using a contrast divergence algorithm to obtain an original response model.

Preferably, the target response model acquisition module 15 includes a test accuracy acquisition unit 151 and a target response model acquisition unit 152.

The test accuracy obtaining unit 151 is configured to test the original response model by using a test set to obtain test accuracy.

The target response model obtaining unit 152 is configured to acquire the target response model if the test accuracy is not less than the preset accuracy.

Example 3

FIG. 7 is a flowchart of the smart chat method in the embodiment. The smart chat method can be applied to computer equipments of financial institutions such as insurance, securities, and banks, or other institutions, for implementing smart chat and for consulting problems of customers. Auto-answer, which improves response efficiency for business promotion. As shown in FIG. 7, the smart chat method includes the following steps:

S21: Calling the information acquisition interface of the WeChat webpage to obtain a WeChat message.

The information acquisition interface is an interface for receiving information that is open on the WeChat webpage. In this embodiment, the program corresponding to the WeChat webpage is installed and run on the computer device, so that the information acquisition interface of the WeChat webpage can be invoked, and the WeChat information fed back by the customer is obtained in real time. In this embodiment, the WeChat information specifically refers to a problem that a customer consults with a financial institution or other institution through a WeChat client, such as an insurance problem.

Specifically, the computer device first obtains the smart chat request, so as to connect to the WeChat web server based on the smart chat request, after the server starts the program, the server automatically generates a two-dimensional code, which can be scanned by the WeChat client on the mobile phone. After the QR code is authorized to log in, the micro-signal on the mobile phone will be converted into an intelligent robot, and the WeChat webpage information acquisition interface will be used to obtain the WeChat message. The intelligent robot will automatically start chatting after receiving the WeChat message to achieve the personal micro-signal based. The purpose of intelligent chat is conducive to the promotion and use of smart chat technology. Among them, the intelligent robot is specifically an application plug-in applied on the WeChat side.

S22: If the WeChat message is a text message, the recognized text data is directly obtained.

The identification text data refers to the non-voice WeChat message data received in step S21. Specifically, the information obtaining interface of the WeChat webpage returns the message type of the WeChat message while returning the WeChat message; if the message type of the returned WeChat message is a text message, directly acquiring the recognized text data for inputting the model to respond In order to facilitate subsequent input of the recognized text data into the target response model for response.

S23: If the WeChat message is a voice message, the third party voice model is called to identify the voice message, and the recognized text data is obtained.

Specifically, the information obtaining interface of the WeChat webpage returns the message type of the WeChat message while returning the WeChat message; if the message type of the returned WeChat message is a voice message, the third party voice model is called to identify the voice message, Obtaining the identification text data, realizes the purpose of the smart robot based on the WeChat end to recognize the voice, and promotes the development of the intelligent chat technology. Among them, the third-party voice model can refer to the voice model developed by Turing Robot Company. The model is mature, and the voice recognition is more accurate. By directly calling, it is beneficial to save development cost.

S24: Input the recognized text data into the target response model, obtain corresponding response information, and call the information sending interface of the WeChat webpage to send the response information.

The target response model is a model obtained by training using the response model training method in Embodiment 1. Specifically, the target response model is used as an intelligent chat driver of the intelligent robot and compiled into the kernel. When the intelligent robot acquires the recognized text data, it will call the smart chat driver in the kernel, that is, start the target response model to respond, obtain the corresponding response information, and call the information sending interface of the WeChat webpage to send the response information to the corresponding The customer's WeChat client to implement smart chat.

Among them, the kernel is the internal core program of the operating system, which provides the core management call to the computer device to the outside. A driver is generally referred to as a device driver (Device Driver), a special program that allows a computer to communicate with a device. The user can be grounded. After the client authorizes the login to the WeChat account, the user can call the WeChat webpage information acquisition interface to obtain the WeChat message. When the identification text data is obtained, the user can directly call the target response model compiled into the kernel to implement the smart chat. In this embodiment, by compiling the target response model as a driver into the kernel, the process of repeating the training model is eliminated, so that the subsequent target response model is directly called when the chat is performed, so as to achieve the WeChat personal WeChat end. The purpose of intelligent chat is conducive to the promotion and use of smart chat.

In this embodiment, the information acquisition interface of the WeChat webpage is first invoked, and the WeChat message is obtained, which provides technical support for subsequent smart chat. If the WeChat message is a voice message, the third-party voice model is called to identify the voice message, and the recognition text data is acquired, so as to realize the purpose that the intelligent robot based on the personal WeChat can recognize the voice, and promote the development of the smart chat technology. If the WeChat message is a text message, the recognized text data is directly obtained. Finally, the recognition text data is input into the target response model, the corresponding response information is obtained, and the information transmission interface of the WeChat webpage is sent to send the response information, so as to achieve the purpose of smart chat based on WeChat personal WeChat, which is beneficial to the promotion of smart chat. use.

In this embodiment, the intelligent robot also has the ability to remember according to the actual situation. Specifically, before the automatic response information, the response information to be sent is compared with the historical response information of the same client. If the same response information has been answered, the response information is not sent, and if the same response is not answered. In response to the message, the response message is sent to reflect the memory capability of the intelligent robot and improve the practicality of the smart chat. For example, when the customer first sends “What are you calling” to the intelligent robot, the intelligent robot will automatically answer “I call XX”. At this time, the “My name is XY” that the intelligent robot responds to the customer will be recorded. During the same session, when the customer re-issues a similar utterance such as "I ask you to name the name", the intelligent robot will automatically answer the "I call XX" response and the intelligent robot has sent out before the automatic response. The responses are compared. If there is the same, the intelligent robot does not respond at this time, which reflects the memory ability of the robot and improves the practicality of the smart chat.

In this embodiment, the developer sets the rules in advance so that the intelligent robot has the ability to end the session as appropriate. Specifically, if the customer sends a disagreement, the intelligent robot will automatically answer a similar dissent. If the customer chats again after the robot sends a disagreement, the robot will not automatically respond, so that the robot can end the session as appropriate. Ability to promote the promotion of intelligent robots. For example, when the customer sends a disguise to the intelligent robot (such as "thank you" or "goodbye!"), indicating that the conversation should end, the intelligent robot will automatically answer the disagreement. At this time, if the customer re-partitions the robot In response to the language, the robot will not respond, so that the intelligent robot has the ability to end the session as appropriate, which is conducive to the promotion of intelligent robots.

In this embodiment, the information acquisition interface of the WeChat webpage is first invoked, and the WeChat message is obtained, which provides technical support for subsequent smart chat. Judging the message type of the WeChat message, if the WeChat message is a voice message, the third party voice model is called to identify the voice message, and the recognition text data is acquired, so that the intelligent voice based on the personal WeChat can recognize the voice and promote the smart chat technology. development of. If the WeChat message is a text message, the recognized text data is directly obtained. Then, by compiling the target response model as a driver into the kernel, the process of repeating the training model is eliminated, so that the trained target response model is directly called to respond in the subsequent chat to realize the smart chat. Finally, the recognized text data is input to the target response model, the corresponding response information is obtained, and the information transmission interface of the WeChat webpage is called to send the response information to achieve the target of the intelligent response. Moreover, before the response message is sent, the response information to be sent is compared with the history response information of the same client. If the same response message has been answered, the response message is not sent, and if the same response is not answered. The information is sent to the response message, which reflects the memory ability of the intelligent robot and improves the practicality of the smart chat. Further, if the customer sends a disagreement, the intelligent robot will respond to a similar dissociation. If the customer chats again after the robot leaves the disagreement, the robot no longer responds, so that the robot has the ability to end the session as appropriate. Conducive to the promotion and use of intelligent robots.

Example 4

FIG. 8 is a schematic block diagram showing a smart chat device corresponding to the smart chat method in the third embodiment. As shown in FIG. 8, the smart chat device includes a WeChat message acquisition module 21, a first identification text data acquisition module 22, a second identification text data acquisition module 23, and a response information acquisition and transmission module 24. The implementation functions of the WeChat message acquisition module 21, the first identification text data acquisition module 22, the second recognition text data acquisition module 23, and the response information acquisition and transmission module 24 are in one-to-one correspondence with the steps corresponding to the smart chat method in the embodiment. In order to avoid redundancy, the present embodiment will not be described in detail.

The WeChat message obtaining module 21 is configured to invoke an information obtaining interface of the WeChat webpage to obtain a WeChat message.

The first identification text data obtaining module 22 is configured to: if the WeChat message is a voice message, invoke a third-party voice model to identify the voice message, and obtain the identification text data.

The second identification text data obtaining module 23 is configured to directly obtain the identification text data if the WeChat message is a text message.

The response information obtaining and transmitting module 24 is configured to input the recognized text data into the target response model, obtain corresponding response information, and invoke the information sending interface of the WeChat webpage to send the response information.

The target response model is a model obtained by training using the response model training method in Embodiment 1.

Example 5

The embodiment provides one or more non-volatile readable storage media having computer readable instructions that, when executed by one or more processors, cause the one or more processors to execute The response model training method in Embodiment 1 is implemented. To avoid repetition, details are not described herein again.

Alternatively, the computer readable instructions are executed by one or more processors such that when executed by the one or more processors, the functions of the modules/units in the response model training device of Embodiment 2 are implemented, in order to avoid duplication, I won't go into details here.

Alternatively, when the computer readable instructions are executed by one or more processors, such that the one or more processors execute the functions of the steps in the smart chat method in Embodiment 3, in order to avoid duplication, here is not One by one.

Alternatively, the computer readable instructions are executed by one or more processors such that when executed by the one or more processors, the functions of the modules/units in the smart chat device of Embodiment 4 are implemented, to avoid repetition, I will not repeat them one by one.

Example 6

FIG. 9 is a schematic diagram of a computer device according to an embodiment of the present application. As shown in FIG. 9, the computer device 90 of this embodiment includes a processor 91, a memory 92, and computer readable instructions 93 stored in the memory 92 and executable on the processor 91, the computer readable instructions being processed by the processor The response model training method in Embodiment 1 is implemented when executed 91. To avoid repetition, details are not described herein. Alternatively, when the computer readable instructions are executed by the processor 91, the functions of the models/units in the response model training device in Embodiment 2 are implemented. To avoid repetition, details are not described herein. Alternatively, the computer readable instructions are implemented by the processor 91 to implement the functions of the steps in the smart chat method in the third embodiment. To avoid repetition, details are not described herein. Alternatively, the computer readable instructions are executed by the processor 91 to implement the functions of the modules/units in the smart chat device of the fourth embodiment. To avoid repetition, we will not go into details here.

It will be apparent to those skilled in the art that, for convenience and brevity of description, only the division of each functional unit and module described above is exemplified. In practical applications, the above functions may be assigned to different functional units as needed. The module is completed by dividing the internal structure of the device into different functional units or modules to perform all or part of the functions described above.

The above-mentioned embodiments are only used to explain the technical solutions of the present application, and are not limited thereto; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still implement the foregoing embodiments. The technical solutions described in the examples are modified or equivalently replaced with some of the technical features; and the modifications or substitutions do not deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in Within the scope of protection of this application.

Claims

A response model training method, comprising:

Obtain the original training text data;

Performing pre-processing on the original training text data to obtain target training text data;

And dividing the target training text data into a preset ratio to obtain a training set and a test set;

The training set is trained by using an RBM model to obtain an original response model;

The original response model is tested using the test set to obtain a target response model.
The response model training method according to claim 1, wherein the preprocessing the original training text data to obtain the target training text data comprises:

The original training text data is identified in Chinese and English to obtain the recognized text;

Performing word segmentation on the recognized text to obtain at least one word;

Performing vectorization processing on at least one of the words to obtain target training text data.
The response model training method according to claim 2, wherein the performing the vectorization processing on the at least one of the words to obtain the target training text data comprises:

Performing, by using a TF-IDF algorithm, performing operations on at least one of the words, and acquiring a word frequency corresponding to each of the words;

The word frequency corresponding to each word is taken as the dimension of the vector, and the target training text data expressed in the form of a vector is obtained.
The response model training method according to claim 1, wherein the training set is trained by using an RBM model to obtain an original response model, including:

Initialize model parameters;

The model parameters are optimized by using a contrast divergence algorithm to obtain the original response model; wherein the formula of the contrast divergence algorithm is CDK (k, S, W, a, b; ΔW, Δa, Δb); k is Contrast divergence algorithm iteration number; S is the training set; W is the weight matrix; a is the offset vector of the visible layer; b is the offset vector of the hidden layer; ΔW is the rate of change of the weight matrix; Δa is visible The rate of change of the layer offset vector, Δb is the rate of change of the hidden layer offset vector.
The response model training method according to claim 1, wherein the testing the original response model by using the test set to obtain a target response model comprises:

The original response model is tested by using the test set to obtain test accuracy;

If the test accuracy is not less than the preset accuracy, the target response model is acquired.
An intelligent chat method, comprising:

Calling the information acquisition interface of the WeChat web version to obtain the WeChat message;

If the WeChat message is a voice message, the third party voice model is called to identify the voice message, and the recognized text data is obtained;

If the WeChat message is a text message, directly acquiring the recognized text data;

Inputting the identification text data into the target response model, acquiring corresponding response information, and calling the information sending interface of the WeChat webpage to send the response information;

The target response model is a model obtained by training using the response model training method according to any one of claims 1-5.
A response model training device, comprising:

The original training text data acquiring module is configured to obtain original training text data;

a target training text data acquiring module, configured to preprocess the original training text data to obtain target training text data;

a target training text data dividing module, configured to divide the target training text data according to a preset ratio, and acquire a training set and a test set;

An original response model acquisition module, configured to train the training set by using an RBM model, and obtain an original response model;

The target response model acquisition module is configured to test the original response model by using the test set to obtain a target response model.
An intelligent chat device, comprising:

a WeChat message obtaining module, configured to invoke an information obtaining interface of a WeChat webpage to obtain a WeChat message;

a first identification text data obtaining module, configured to: if the WeChat message is a voice message, invoke a third-party voice model to identify the voice message, and obtain the identification text data;

a second identification text data obtaining module, configured to directly obtain the identification text data if the WeChat message is a text message;

a response information obtaining and sending module, configured to input the recognized text data into the target response model, obtain corresponding response information, and invoke an information sending interface of a WeChat webpage to send the response information;

The target response model is a model obtained by training using the response model training method according to any one of claims 1-5.
A computer device comprising a memory, a processor, and computer readable instructions stored in the memory and operative on the processor, wherein the processor executes the computer readable instructions as follows step:

Obtain the original training text data;

Performing pre-processing on the original training text data to obtain target training text data;

And dividing the target training text data into a preset ratio to obtain a training set and a test set;

The training set is trained by using an RBM model to obtain an original response model;

The original response model is tested using the test set to obtain a target response model.
The computer device according to claim 9, wherein the pre-processing the original training text data to obtain the target training text data comprises:

The original training text data is identified in Chinese and English to obtain the recognized text;

Performing word segmentation on the recognized text to obtain at least one word;

Performing vectorization processing on at least one of the words to obtain target training text data.
The computer device according to claim 10, wherein the performing the vectorization processing on the at least one of the words to obtain the target training text data comprises:

Performing, by using a TF-IDF algorithm, performing operations on at least one of the words, and acquiring a word frequency corresponding to each of the words;

The word frequency corresponding to each word is taken as the dimension of the vector, and the target training text data expressed in the form of a vector is obtained.
The computer device according to claim 9, wherein the training of the training set by using an RBM model to obtain an original response model comprises:

Initialize model parameters;

The model parameters are optimized by using a contrast divergence algorithm to obtain the original response model; wherein the formula of the contrast divergence algorithm is CDK (k, S, W, a, b; ΔW, Δa, Δb); k is Contrast divergence algorithm iteration number; S is the training set; W is the weight matrix; a is the offset vector of the visible layer; b is the offset vector of the hidden layer; ΔW is the rate of change of the weight matrix; Δa is visible The rate of change of the layer offset vector, Δb is the rate of change of the hidden layer offset vector.
The computer device according to claim 9, wherein the testing the original response model by using the test set to obtain a target response model comprises:

The original response model is tested by using the test set to obtain test accuracy;

If the test accuracy is not less than the preset accuracy, the target response model is acquired.
A computer device comprising a memory, a processor, and computer readable instructions stored in the memory and operative on the processor, wherein the processor executes the computer readable instructions as follows step:

Calling the information acquisition interface of the WeChat web version to obtain the WeChat message;

If the WeChat message is a voice message, the third party voice model is called to identify the voice message, and the recognized text data is obtained;

If the WeChat message is a text message, directly acquiring the recognized text data;

Inputting the identification text data into the target response model, acquiring corresponding response information, and calling the information sending interface of the WeChat webpage to send the response information;

The target response model is a model obtained by training using the response model training method according to any one of claims 1-5.
One or more non-transitory readable storage mediums storing computer readable instructions, wherein when the computer readable instructions are executed by one or more processors, cause the one or more processors to execute The following steps:

Obtain the original training text data;

Performing pre-processing on the original training text data to obtain target training text data;

And dividing the target training text data into a preset ratio to obtain a training set and a test set;

The training set is trained by using an RBM model to obtain an original response model;

The original response model is tested using the test set to obtain a target response model.
The non-volatile readable storage medium according to claim 15, wherein the pre-processing the original training text data to obtain the target training text data comprises:

The original training text data is identified in Chinese and English to obtain the recognized text;

Performing word segmentation on the recognized text to obtain at least one word;

Performing vectorization processing on at least one of the words to obtain target training text data.
The non-volatile readable storage medium according to claim 16, wherein the performing the vectorization processing on the at least one of the words to obtain the target training text data comprises:

Performing, by using a TF-IDF algorithm, performing operations on at least one of the words, and acquiring a word frequency corresponding to each of the words;

The word frequency corresponding to each word is taken as the dimension of the vector, and the target training text data expressed in the form of a vector is obtained.
The non-volatile readable storage medium according to claim 15, wherein the training of the training set by using an RBM model to obtain an original response model comprises:

Initialize model parameters;

The model parameters are optimized by using a contrast divergence algorithm to obtain the original response model; wherein the formula of the contrast divergence algorithm is CDK (k, S, W, a, b; ΔW, Δa, Δb); k is Contrast divergence algorithm iteration number; S is the training set; W is the weight matrix; a is the offset vector of the visible layer; b is the offset vector of the hidden layer; ΔW is the rate of change of the weight matrix; Δa is visible The rate of change of the layer offset vector, Δb is the rate of change of the hidden layer offset vector.
The non-volatile readable storage medium according to claim 15, wherein the testing the original response model by using the test set to obtain a target response model comprises:

The original response model is tested by using the test set to obtain test accuracy;

If the test accuracy is not less than the preset accuracy, the target response model is acquired.
One or more non-transitory readable storage mediums storing computer readable instructions, wherein when the computer readable instructions are executed by one or more processors, cause the one or more processors to execute The following steps:

Calling the information acquisition interface of the WeChat web version to obtain the WeChat message;

If the WeChat message is a voice message, the third party voice model is called to identify the voice message, and the recognized text data is obtained;

If the WeChat message is a text message, directly acquiring the recognized text data;

Inputting the identification text data into the target response model, acquiring corresponding response information, and calling the information sending interface of the WeChat webpage to send the response information;

The target response model is a model obtained by training using the response model training method according to any one of claims 1-5.