WO2020073530A1

WO2020073530A1 - Customer service robot session text classification method and apparatus, and electronic device and computer-readable storage medium

Info

Publication number: WO2020073530A1
Application number: PCT/CN2018/125249
Authority: WO
Inventors: 许开河; 杨坤; 王少军; 肖京
Original assignee: 平安科技（深圳）有限公司
Priority date: 2018-10-12
Filing date: 2018-12-29
Publication date: 2020-04-16
Also published as: CN109543030A; CN109543030B

Abstract

Disclosed are a customer service robot session text classification method and apparatus. The customer service robot session text classification method comprises: acquiring an input statement in a session carried out by a customer service robot, and converting the input statement into a standard session text, wherein the input statement is a session message waiting for the customer service robot to process and respond to same; performing semantic feature extraction on the standard session text to obtain a semantic vector corresponding to the standard session text; performing category label prediction on the semantic vector of the standard session text to obtain a category label probability vector corresponding to the standard session text; and selecting, from the category label probability vector, a category corresponding to a label with the maximum probability to serve as the category of the standard session text, wherein the category is used for assisting the execution of a response of the customer service robot to the input text. By means of the customer service robot session text classification method and apparatus disclosed in the present application, an input statement acquired by a customer service robot can be accurately classified.

Description

Customer service robot conversation text classification method and device, electronic equipment, and computer-readable storage medium

Technical field

This application requires the priority of the Chinese patent application 201811191509.3 filed on October 12, 2018, and the application name is "Customer Service Robot Session Text Classification Method and Apparatus, Equipment, and Storage Media", and the entire contents are incorporated herein by reference.

The present application relates to the field of data processing technology, and in particular to a customer service robot conversation text classification method and device, electronic equipment, and computer-readable storage medium.

Background technique

In the FAQ (Frequently Asked Questions) scenario of customer service robots, each knowledge point corresponds to a standard question. There are multiple questions for standard questions. These different questions are called extended questions. After the customer service robot obtains the extended question, it needs to use the text classification model to classify the extended question to obtain the category of the standard question corresponding to the extended question, and then extract the answer that matches the standard question category from its own knowledge base according to the resulting standard question category . Therefore, whether the expansion problem is accurately classified is the key to whether the customer service robot can accurately answer customer questions.

The inventor realized that in the process of classifying the expansion problem by the text classification model, the expansion problem is mapped to the vector space, and by dividing the vector space, the standard problem in the vector space to which the vector corresponding to the expansion problem belongs is obtained. Control area, the category corresponding to this control area is the standard problem category corresponding to the extended problem. Some knowledge points correspond to a relatively small number of expansion problems, and the standard problem categories corresponding to these expansion problems also have fewer control areas in the vector space, resulting in easy classification of these expansion problems in text classification, so that these expansions cannot be accurately obtained. The standard question category corresponding to the question.

Therefore, how to accurately classify the extended problems acquired by the customer service robot is a problem to be solved in the prior art.

technical problem

In order to solve the above technical problems, an object of the present application is to provide a customer service robot conversation text classification method and device, electronic equipment, and computer-readable storage medium.

Technical solution

Among them, the technical solutions adopted in this application are:

On the one hand, a conversation text classification method for a customer service robot includes: acquiring input sentences of a customer service robot in a conversation, converting the input sentences into standard conversation text, and the input sentences are conversations waiting for a response from the customer service robot to process a response Message; obtain the semantic vector corresponding to the standard conversation text by performing semantic feature extraction on the standard conversation text; perform category label prediction on the semantic vector of the standard conversation text to obtain the category label probability vector corresponding to the standard conversation text Selecting the category corresponding to the maximum probability label from the category label probability vector as the category of the standard conversation text, the category is used to assist in performing the response of the customer service robot to the input text.

On the other hand, a customer service robot conversation text classification device includes: an input sentence conversion module for acquiring an input sentence of a customer service robot in a conversation, converting the input sentence into standard conversation text, and the input sentence is waiting The customer service robot processes the responded conversation message; the semantic feature extraction module is used to obtain the semantic vector corresponding to the standard conversation text by performing semantic feature extraction on the standard conversation text; the text category prediction module is used to compare the standard The semantic vector of the conversation text is used to predict the category label, and the probability vector of the category label corresponding to the standard conversation text is obtained; the text category acquisition module is used to select the category corresponding to the maximum probability label from the category label probability vector as the criterion A category of conversation text, which is used to assist in performing the response of the customer service robot to the input text.

On the other hand, an electronic device includes a processor and a memory, and a computer-readable instruction is stored on the memory, and when the computer-readable instruction is executed by the processor, the method for classifying a customer service robot conversation text as described above is implemented .

On the other hand, a computer-readable storage medium has stored thereon a computer program, and when the computer program is executed by a processor, the customer service robot conversation text classification method as described above is implemented.

In the above technical solution, the input sentence of the customer service robot in the ongoing conversation is the extended question acquired by the customer service robot, and the standard conversation text is the standard question corresponding to this extended question. After obtaining the input sentence in the conversation conducted by the customer service robot, the application first converts the input sentence into standard conversation text, and then classifies the resulting standard conversation text.

Beneficial effect

Since the number of standard conversation texts with different categories is often only one, the size of the corresponding control area of the different standard conversation text categories in the text category marking space is the same, so that when text classification is performed on the standard conversation texts, there is no reason for the The inconsistent size of the control area leads to misclassification, so that the input sentences obtained by the customer service robot can be accurately classified.

It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and do not limit the present application.

BRIEF DESCRIPTION

The drawings here are incorporated into the specification and constitute a part of the specification, show embodiments consistent with the present application, and are used together with the specification to explain the principles of the present application.

Fig. 1 is a hardware block diagram of a customer service robot according to an exemplary embodiment.

Fig. 2 is a flowchart illustrating a method for classifying customer service robot conversation text according to an exemplary embodiment.

Fig. 3 is a schematic diagram illustrating a process of encoding and decoding an input sentence according to an exemplary embodiment.

Fig. 4 is a flow chart showing a method for classifying customer service robot conversation text according to another exemplary embodiment.

Fig. 5 is a block diagram showing a customer service robot conversation text classification device according to an exemplary embodiment.

Through the above drawings, clear embodiments of the present application have been shown, which will be described in more detail later. These drawings and text descriptions are not intended to limit the scope of the present application in any way, but by referring to specific embodiments The concept of the present application will be explained to those skilled in the art.

Embodiments of the invention

The exemplary embodiments will be explained in detail here, examples of which are shown in the drawings. When the following description refers to the accompanying drawings, unless otherwise indicated, the same numerals in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of devices and methods consistent with some aspects of the application as detailed in the appended claims.

Fig. 1 is a hardware block diagram of a customer service robot according to an exemplary embodiment. It should be noted that the customer service robot is only an example adapted to the present disclosure, and cannot be considered as providing any limitation on the scope of use of the present disclosure.

As shown in FIG. 1, the customer service robot may include one or more of the following components: a processing component 101, a memory 102, a power component 103, a multimedia component 104, an audio component 105, a sensor component 107, and a communication component 108. Among them, the above components are not all necessary. The customer service robot may add other components or reduce some components according to its own functional requirements, which is not limited in this embodiment.

The processing component 101 generally controls the overall operations of the customer service robot, such as operations associated with display, data communication, camera operations, and log data processing. The processing component 101 may include one or more processors 109 to execute instructions to complete all or part of the steps of the above operations. In addition, the processing component 101 may include one or more modules to facilitate interaction between the processing component 101 and other components. For example, the processing component 101 may include a multimedia module to facilitate interaction between the multimedia component 104 and the processing component 101.

The memory 102 is configured to store various types of data to support the operation of the customer service robot. Examples of these data include instructions for any application or method to operate on the customer service robot. The memory 102 may be implemented by any type of volatile or nonvolatile storage device or a combination thereof, such as SRAM (Static Random Access Memory), EEPROM (Electrically Erasable Programmable Read Only Memory), ROM (Read Only Memory), disk or CD. One or more modules are also stored in the memory 102, and the one or more modules are configured to be executed by the one or more processors 109 to complete all or part of any of the following customer service robot conversation text classification methods step.

The power supply component 103 provides power for various components of the customer service robot. The power supply component 103 may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power for the customer service robot.

The multimedia component 104 includes a screen that provides an output interface between the customer service robot and the user. In some embodiments, the screen may include an LCD (liquid crystal display) and a TP (touch panel). If the screen includes a touch panel, the screen may be implemented as a touch screen to receive input signals from the user. The touch panel includes one or more touch sensors to sense touch, swipe, and gestures on the touch panel. The touch sensor may not only sense the boundary of the touch or sliding action, but also detect the duration and pressure related to the touch or sliding operation.

The audio component 105 is configured to output and / or input audio signals. For example, the audio component 105 includes a microphone. When the customer service robot is in an operation mode, such as a recording mode and a voice recognition mode, the microphone is configured to receive an external audio signal. The received audio signal may be further stored in the memory 102 or transmitted via the communication component 108. The audio component 105 further includes a speaker for outputting audio signals to implement conversation operations between the customer service robot and the customer.

The sensor assembly 107 includes one or more sensors for providing computer equipment with various aspects of status assessment. For example, the sensor component 107 can also detect changes in the coordinates of the customer service robot or a component of the customer service robot and temperature changes in the customer service robot. In some embodiments, the sensor assembly 107 may also include a magnetic sensor, a pressure sensor, or a temperature sensor.

The communication component 108 is configured to facilitate wired or wireless communication between the customer service robot and other devices. Customer service robots can access wireless networks based on communication standards, such as WiFi, 2G, or 3G, or a combination thereof.

In an exemplary embodiment, the customer service robot may be controlled by one or more ASICs (application specific integrated circuits), DSPs (digital signal processors), PLDs (programmable logic devices), FPGAs (field programmable gate arrays), and controllers , Microcontroller, microprocessor or other electronic components to implement the text classification method of customer service robot conversation shown below.

In this embodiment, the customer service robot is a machine device for automatically performing dialogue work, and may specifically be a terminal device such as a smart phone, tablet computer, notebook computer, or other machine equipment with a specific shape and function. Not going.

Fig. 2 is a flow chart of a method for classifying conversation text of a customer service robot according to an exemplary embodiment. The method is applicable to the customer service robot shown in Fig. 1. As shown in Figure 2, the method may include the following steps:

In step 210, the input sentence of the customer service robot in the ongoing conversation is obtained, and the input sentence is converted into standard conversation text.

Among them, the input sentence of the customer service robot in the conversation is a conversation message waiting for the response of the customer service robot to process the response, which is easy to understand. The input sentence is an expansion problem that the customer inputs to the customer service robot during the conversation between the customer service robot and the customer. For example, the customer enters "Hello, I would like to ask what is the annual fee of the owner card", "Will I apply for an annual fee for the owner card", "I ask, my owner card "What is the annual fee-free requirement" and other input sentences, and these input sentences are all extended questions corresponding to the standard question "Car Owner's Card Annual Fee".

The input sentence may be obtained by the customer service robot by recognizing the voice signal input by the customer. For example, the customer service robot obtains the question voice input by the customer through the microphone configured by itself, and performs speech recognition on the obtained question voice to obtain the input sentence.

The input sentence can also be obtained through the touch screen configured by the customer service robot. For example, the customer enters the question he wants to ask on the touch screen configured by the customer service robot. At this time, the customer service robot directly obtains the text information entered on the touch screen as an input sentence.

After the input sentence is obtained, the input sentence is converted into standard conversation text corresponding to the input sentence. Among them, the standard conversation text is the standard question corresponding to the extension question, such as the above-mentioned "Car Owner's Card Annual Fee".

In an embodiment, the input sentence can be converted into standard conversation text corresponding to the input sentence through text translation, which may include the following steps:

By encoding the input sentence, the key semantic features of the input sentence are extracted;

Decode the key semantic features to obtain the standard conversation text corresponding to the input sentence.

Among them, the coding of the input sentence is carried out using a neural network model to automatically analyze the key semantic features of the input sentence. The key semantic feature is an important feature used to express the semantics of the input sentence. It is highly related to the semantics of the input sentence and can include the structural features and keywords of the input sentence.

In this embodiment, an LSTM (Long Short-Term Memory, Long Short-Term Neural Network) model can be used to encode input sentences. The specific process is as follows: input each word vector of the input sentence into the LSTM model in sequence, and input The word vector is traversed to obtain a hidden state vector obtained through traversal, and the hidden state vector is a semantic vector corresponding to the input sentence.

The word vector of the input sentence is obtained by vectorizing the words in the input sentence. First, perform word segmentation on the input sentence to divide the input text into several word sequences arranged in sequence. For example, if the input text is "Do I have to pay an annual fee for the owner card I apply for", the word segmentation process will result in the phrase "Please ^ Me ^ Apply ^ 's ^ Owner Card ^ Want ^ Receive ^ Annual fee ^" The word segmentation processing of the input sentence may be performed by using a word segmentation algorithm, such as a word segmentation method based on string matching, a word segmentation method based on understanding, or a word segmentation method based on statistics.

Then, each word in the word sequence is mapped to a low-dimensional vector to obtain a word vector corresponding to each word. Specifically, one-hot (one-hot code) vector coding or word2vec (word embeddings (word vector) vector coding method, or other methods can also be used, not limited here.

It should be noted that, since the vector obtained by the one-hot vector coding method does not store the correlation between the words in the input sentence, it is necessary to add weight information to the one-hot vector obtained by each word. The weight of each word is related to the degree of semantic relevance of the word to the input sentence. For example, in the above input sentence "Do I need to charge an annual fee for the owner card I apply for", the "owner card" and "annual fee" The two words have a greater semantic relevance to the input sentence, and the corresponding weights should be higher, while the words "I ask" and "I" are obviously not highly semantically related to the input sentence, and the corresponding weights are also higher. low.

Each word vector obtained through the word2vec vector coding method is also associated with the semantics of the input sentence, and each word vector obtained through the word2vec method can still reflect the degree of relevance of each word to the input sentence semantics.

Each word vector of the input sentence is input into the LSTM model in sequence, and the specific process of traversing the input word vector in chronological order is shown in FIG. 3. The word vectors X1, X2, X3 are sequentially input into the LSTM model in chronological order, and the state of the hidden layer at different times is updated. The update of the state of the hidden layer at each time depends on the state of the hidden layer updated at the previous time, and will be updated to EOS end of sentence, the first hidden state vector L as the semantic vector of the input sentence.

By traversing each word vector in the input sentence in the LSTM model, the output first hidden state vector L can establish the global semantic expression of each word combined with the input sentence, so that the obtained semantic vector fully correlates with the key semantics of the input sentence feature. In addition, Bi-LSTM (Bi-Long Short-Term Memory, bidirectional long-term neural network) model can also be used to traverse each word vector in the input sentence, which is not limited here.

Correspondingly, the key features of the input sentence are decoded using another LSTM model or Bi-LSTM model. The LSTM model is used as an example for description below.

The specific decoding process is still shown in Figure 3. The semantic vector L of the encoded input sentence is used as the initial value in the LSTM model, the probability distribution of the output words at this time is calculated, the probability of the possible output words is obtained, and then the probability of the output word Sampling is performed to obtain the final word O output at this moment, and the state of the hidden layer is updated. Next, the word vector O finally output at this time is used as the input at the next time, and the updated hidden layer state is passed to the next time to calculate the word P output at the next time. This cycle until the end of the output sentence indicates that the decoding is complete.

The word sequence obtained by arranging the words output by decoding in chronological order is the standard conversation text obtained by text translation of the input sentence.

It should be noted that the text translation of the input sentence may be performed by the processor configured by the customer service robot, or may be performed by a server that has established a wired or wireless network connection with the customer service robot in advance. limited.

In step 230, a semantic vector corresponding to the standard conversation text is obtained by performing semantic feature extraction on the standard conversation text.

Among them, in order to realize the conversation between the customer service robot and the customer, after the input sentence is converted into standard conversation text, it is also necessary to classify the standard conversation text so that the customer service robot executes the input sentence according to the category corresponding to the standard conversation text response.

In order to obtain the category of the standard conversation text, in one embodiment, a convolutional neural network (CNN) model is used to extract the semantic features of the standard conversation text to obtain the semantic vector corresponding to the standard conversation text.

Obtain the second hidden state vector obtained by decoding the key semantic features, and form the hidden state vector matrix from the second hidden state vector;

Semantic feature extraction of standard conversation text based on hidden state vector matrix;

The semantic vector corresponding to the standard conversation text is obtained by pooling the extracted semantic features.

The second hidden state vector obtained by decoding the key semantic features is the hidden layer state vector corresponding to each output word. Several second hidden state vectors obtained by decoding are arranged in sequence to form a vector matrix with dimension dimension length_state (length of state sequence) × hidden_size (number of hidden state vectors) to obtain a hidden state vector matrix. The state sequence length is the number of elements contained in the second hidden state vector, and the resulting hidden state vector matrix is used as the input layer of the convolutional neural network.

After obtaining the hidden state vector matrix, the convolutional layer of the convolutional neural network convolves the hidden state vector matrix to convolve the input layer to obtain several features Map (feature label). The size of the convolution window is the length of the state sequence in the hidden state vector matrix × the number of hidden state vectors.

After the convolutional layer of the convolutional neural network is used to convolve the hidden state vector matrix, a number of feature labels with a column number of 1 are obtained. These feature labels are used to represent the semantic features of standard conversational text.

The pooling of the extracted semantic features is performed by the pooling layer of the convolutional neural network model. The pooling layer extracts the feature vector corresponding to the maximum value from each feature label obtained by the convolutional layer, and obtains the semantic vector corresponding to the standard conversation text by combining these extracted feature vectors.

In step 250, class label prediction is performed on the semantic vector of the standard conversation text to obtain the class label probability vector corresponding to the standard conversation text.

Among them, the category label prediction of the semantic vector of the standard conversation text is to predict the probability that the standard conversation text satisfies each control region in the text category label space according to the semantic vector of the standard conversation text. The text category mark space is a preset sample space, and the control areas in the text category mark space correspond to preset standard question categories. Each element in the predicted category label probability vector is the probability that the standard conversation text satisfies each control region in the text category label space.

In an embodiment, the category label prediction on the semantic vector of the standard conversation text may specifically include the following steps:

Through the target parameter matrix, the semantic vector of the standard conversation text is nonlinearly mapped to obtain the state vector of the standard conversation text in the text category label space;

Probability normalization is performed on the state vector of the standard conversation text in the text category label space to obtain the category label probability vector corresponding to the standard conversation text.

Among them, the non-linear mapping of the semantic vector of the standard conversation text through the target parameter matrix is used to map the semantic vector of the standard conversation text to the text category label space, so as to associate the standard conversation text with the sample space. The target parameter matrix is a non-linear mapping condition that maps the semantic vector of the standard conversation text to the text category label space. Therefore, the parameter values in the target parameter matrix will directly affect the accuracy of class label prediction of the semantic vector of the standard conversation text. degree.

The mapping of the semantic vectors of standard conversational texts to the sample mark space is specifically to perform a weighted sum operation on the semantic vectors and the target parameter matrix, which can be expressed as: z = Wx, where "W" indicates the target parameter matrix and "x" indicates The semantic vector of standard conversational text, "z" correspondingly represents the state vector of the standard conversational text in the space of text category mark.

In an embodiment, the probability normalization of the state vector of the standard conversation text in the text category label space is performed by the Softmax multi-class prediction function. The definition of Softmax multi-class prediction function is as follows:

Among them, "k" represents the number of divided text categories in the text category label space. The softmax multi-class prediction function maps the state vector of standard conversation text in the text category label space to a probability vector formed by the combination of probability value sequences between (0,1) to obtain the category label probability corresponding to the standard conversation text vector.

In step 270, the category corresponding to the maximum probability label is selected from the category label probability vector as the category of standard conversation text.

Among them, as mentioned above, each element in the category label probability vector is the probability that the standard conversation text satisfies each control region in the text category label space, and the category marked by the most probable control region is closest to the true category of the standard conversation text .

Therefore, selecting the category corresponding to the maximum probability label from the category label probability vector as the category of the standard conversation text can be as close as possible to the true category of the standard conversation text, thereby accurately predicting the type of the standard conversation text.

It should be noted that the process of classifying the standard conversation text in this embodiment may be performed by the processor configured by the customer service robot, or may be pre-established with the customer service robot Executed by a server connected to a wired or wireless network.

In this embodiment, the input sentence is first converted into standard conversational text, and then the standard conversational text is classified. Since each control area in the text category mark space corresponds to each preset standard question, and the standard conversation text is the standard question corresponding to the input sentence, when classifying the standard conversation text, each text area in the text category mark space The size of a control area is the same, so that in this embodiment, when classifying standard conversational text, it will not be erroneously divided because of the size of the control area in the text category mark space. Therefore, the method provided by the present application can accurately predict the category corresponding to the input sentence.

In an application scenario, after the customer service robot obtains the category of the input sentence, it selects the answer sentence matching this category from its own knowledge base, and outputs the voice of the answer sentence through the speaker configured by the customer service robot, or through the configured The LCD screen displays text of the answer sentence, so as to have a conversation with the user.

The above-mentioned method for classifying customer service robot conversation text can be used as an offline training stage and an online prediction stage, respectively. The purpose of offline training is to optimize the target parameter matrix described in step 250 to obtain the optimal target parameter matrix. In the online prediction stage, the optimal target parameter matrix obtained in the offline training stage is directly used to classify the input sentences, and the optimal category of the input sentences is directly output.

Fig. 4 is a method for classifying conversation text of a customer service robot shown in another exemplary embodiment, which is applicable to the offline training stage. As shown in FIG. 4, after obtaining the category of the standard conversation text, the method may further include the following steps:

In step 310, the translation deviation of the input sentence into the standard conversational text and the classification deviation of the text classification of the standard conversational text are summed to obtain the input sentence classification deviation.

Among them, the translation deviation is the error value between the standard conversation text converted from the input sentence and the real standard conversation text of the input sentence, and the classification deviation is between the category obtained by text classification of the standard conversation text and the real category of the standard conversation text Error value.

In one embodiment, the input sentence classification deviation is calculated according to the cross-entropy loss function. In the process of text classification of input sentences, since the accuracy of text translation of input sentences and the accuracy of text classification of standard conversation texts can both affect the accuracy of text classification of input sentences, the definition of the present invention is crossover The entropy loss function includes the sum of these two loss functions.

The cross-entropy loss function defined in the present invention is:

Among them, p (x) represents the probability of translating the input sentence into the standard conversation sample x in text translation. Only when the input sentence is translated into the real standard conversation text, the value of p (x) is 1, in other cases p (x) The value of x) is 0. q (x) represents the probability of text translation of the input sentence. p (i) represents the probability of labeling the standard conversation text as category i in text classification. Only when the standard conversation text is marked as the real text category, the value of p (i) is 1, in other cases p (i) The value is 0. q (i) represents the class probability obtained by text classification of standard conversation text.

According to the above cross-entropy loss function, the value of the input sentence classification deviation H (p, q) is calculated. If the calculated input sentence classification deviation is less than the preset threshold, it means that the target parameter matrix currently used for classifying and predicting the input sentence is not optimal.

In step 330, the target parameter matrix is updated by minimizing the input sentence classification deviation.

Among them, if the resulting input sentence classification deviation is less than a preset threshold, the input sentence classification deviation needs to be minimized.

In one embodiment, a gradient descent algorithm is used to minimize input sentence classification deviation. The specific processing procedure is: performing a derivative operation on the above cross-entropy loss function to obtain the partial derivative of the cross-entropy loss function with respect to the current target parameter matrix. The obtained partial derivatives are also called gradient values. Then, the target parameter matrix currently used and the obtained partial derivative are subtracted to obtain a new parameter matrix, and the target parameter matrix is updated according to the new parameter matrix.

After obtaining the updated parameter matrix, the input sentence is still trained for the next text classification according to the methods described in steps 210 to 270, and the target parameter matrix adopted at this time is the updated parameter matrix. After the category corresponding to the input sentence is obtained, the input sentence classification deviation is calculated according to the cross entropy loss function.

If the obtained input sentence classification deviation is still less than the preset threshold, then repeat the method described in steps 310 and 330 to update the target parameter matrix, and perform the next text classification on the input sentence according to the method described in steps 210 to 270 Training, until the obtained input sentence classification deviation is greater than a preset threshold, it means that the target parameter matrix used for text classification training of the input sentence this time is optimal, and the offline training phase is completed at this time. In the online prediction phase, the target parameters used in this training are directly used for online prediction of input sentences.

As shown in FIG. 5, in an exemplary embodiment, the present application further provides a customer service robot conversation text classification device, which includes:

The input sentence conversion module 410 is used to obtain the input sentence of the customer service robot in the ongoing conversation, convert the input sentence into standard conversation text, and the input sentence is a conversation message waiting for the response of the customer service robot to process the response;

The semantic feature extraction module 430 is used to obtain the semantic vector corresponding to the standard conversation text by performing semantic feature extraction on the standard conversation text;

The text category prediction module 450 is used to predict category labels on the semantic vector of the standard conversation text to obtain the category label probability vector corresponding to the standard conversation text;

The text category acquisition module 470 is used to select the category corresponding to the maximum probability label from the category label probability vector as the category of standard conversation text, which is used to assist in executing the response of the customer service robot to the input text.

In another exemplary embodiment, the input sentence conversion module 410 may include:

The sentence coding unit is used to extract the key semantic features of the input sentence by coding the input sentence;

The sentence decoding unit is used to decode key semantic features to obtain the standard conversation text corresponding to the input sentence.

In another exemplary embodiment, the sentence encoding unit may include:

The word vector acquisition subunit is used to obtain the word vector corresponding to the word in the input sentence by vectorizing the word in the input sentence;

The semantic vector acquisition subunit is used to traverse the word vector corresponding to the words in the input sentence in chronological order, and extract the first hidden state vector obtained by the traversal as the semantic vector of the input sentence.

In another exemplary embodiment, the semantic feature extraction module 430 includes:

The feature acquisition unit is used to acquire a second hidden state vector obtained by decoding key semantic features, and the second hidden state vector forms a hidden state vector matrix;

Feature extraction unit, used to extract semantic features of standard conversation text according to the hidden state vector matrix;

The feature pooling unit is used to obtain the semantic vector corresponding to the standard conversation text by pooling the extracted semantic features.

In another exemplary embodiment, the text category prediction module 450 may further include:

The state vector acquisition unit is used to nonlinearly map the semantic vector of the standard conversation text through the target parameter matrix to obtain the state vector of the standard conversation text in the text category label space;

The category label probability vector acquisition unit is used to normalize the probability of the state vector of the standard conversation text in the text category label space to obtain the category label probability vector corresponding to the standard conversation text.

In another exemplary embodiment, the robot conversation text classification apparatus further includes:

The input sentence classification deviation acquisition module is used to sum the translation deviation of the input sentence into the standard conversation text and the classification deviation of the text classification of the standard conversation text to obtain the input sentence classification deviation;

The parameter update module is used to update the target parameter matrix by minimizing the input sentence classification deviation.

It should be noted that the device provided in the above embodiment and the method provided in the above embodiment belong to the same concept, and the specific manner in which each module performs operations has been described in detail in the method embodiment, and will not be repeated here.

In an exemplary embodiment, the present application further provides an electronic device, the electronic device includes:

Processor; memory, computer-readable instructions are stored on the memory, and when the computer-readable instructions are executed by the processor, the customer service robot conversation text classification method shown above is realized.

In an exemplary embodiment, the present application also provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, implements the customer service robot conversation text classification method as described above.

The above content is only a preferred exemplary embodiment of the present application and is not intended to limit the implementation of the present application. Those of ordinary skill in the art can easily make corresponding changes or modifications according to the main idea and spirit of the present application. Therefore, the scope of protection of this application shall be subject to the scope of protection required by the claims.

Claims

A customer service robot conversation text classification method, including:

Obtain the input sentence of the customer service robot in the ongoing conversation, convert the input sentence into standard conversation text, and the input sentence is a conversation message waiting for the customer service robot to process a response;

Obtaining a semantic vector corresponding to the standard conversation text by performing semantic feature extraction on the standard conversation text;

Performing category label prediction on the semantic vector of the standard conversation text to obtain a category label probability vector corresponding to the standard conversation text;

The category corresponding to the maximum probability label is selected from the category label probability vector as the category of the standard conversation text, and the category is used to assist in performing the response of the customer service robot to the input text.
The method according to claim 1, wherein the acquiring input sentences of the customer service robot in the conducted conversation, and converting the input sentences into standard conversation text includes:

Encoding the input sentence to extract key semantic features of the input sentence;

Decode the key semantic features to obtain standard conversation text corresponding to the input sentence.
The method of claim 2, wherein the extracting key semantic features of the input sentence by encoding the input sentence comprises:

Obtaining the word vector corresponding to the word in the input sentence by vectorizing the word in the input sentence;

The word vectors corresponding to the words in the input sentence are traversed in chronological order, and the first hidden state vector obtained by the traversal is extracted as the semantic vector of the input sentence.
The method according to any one of claims 1 to 3, wherein the obtaining the semantic vector corresponding to the standard conversation text by performing semantic feature extraction on the standard conversation text includes:

Acquiring a second hidden state vector obtained by decoding the key semantic features, and forming a hidden state vector matrix from the second hidden state vector;

Performing semantic feature extraction on the standard conversation text according to the hidden state vector matrix;

The semantic vector corresponding to the standard conversation text is obtained by pooling the extracted semantic features.
The method according to any one of claims 1 to 4, wherein the class label prediction on the semantic vector of the standard conversation text to obtain the class label probability vector corresponding to the standard conversation text includes:

Performing a non-linear mapping on the semantic vector of the standard conversation text through the target parameter matrix to obtain the state vector of the standard conversation text in the text category label space;

Probability normalization is performed on the state vector of the standard conversation text in the text category label space to obtain a category label probability vector corresponding to the standard conversation text.
The method of claim 5, further comprising:

Summing the translation deviations of converting the input sentence into the standard conversational text and the classification deviation of the text classification of the standard conversational text to obtain the input sentence classification deviation;

The target parameter matrix is updated by minimizing the input sentence classification deviation.
A customer service robot conversation text classification device, including:

The input sentence conversion module is configured to obtain the input sentence of the customer service robot in the ongoing conversation, convert the input sentence into standard conversation text, and the input sentence is a conversation message waiting for the customer service robot to process a response;

A semantic feature extraction module configured to obtain a semantic vector corresponding to the standard conversation text by performing semantic feature extraction on the standard conversation text;

A text category prediction module configured to predict category labels on the semantic vector of the standard conversation text to obtain a category label probability vector corresponding to the standard conversation text;

The text category acquisition module is configured to select the category corresponding to the maximum probability label from the category label probability vector as the category of the standard conversation text, the category is used to assist in performing the response of the customer service robot to the input text .
The apparatus of claim 7, wherein the sentence conversion module comprises:

The sentence coding unit is configured to extract key semantic features of the input sentence by coding the input sentence;

The sentence decoding unit is configured to decode key semantic features and obtain standard conversation text corresponding to the input sentence.
The apparatus of claim 8, wherein the sentence encoding unit comprises:

The word vector acquisition subunit is configured to obtain the word vector corresponding to the word in the input sentence by vectorizing the word in the input sentence;

The semantic vector acquisition subunit is configured to traverse the word vector corresponding to the words in the input sentence in chronological order, and extract the first hidden state vector obtained by the traversal as the semantic vector of the input sentence.
The apparatus according to any one of claims 7 to 9, wherein the semantic feature extraction module includes:

The feature acquisition unit is configured to acquire a second hidden state vector obtained by decoding key semantic features, and the second hidden state vector constitutes a hidden state vector matrix;

The feature extraction unit is configured to extract the semantic features of the standard conversation text according to the hidden state vector matrix;

The feature pooling unit is configured to obtain the semantic vector corresponding to the standard conversation text by pooling the extracted semantic features.
The apparatus according to any one of claims 7 or 10, wherein the text category prediction module includes:

The state vector acquisition unit is configured to nonlinearly map the semantic vector of the standard conversation text through the target parameter matrix to obtain the state vector of the standard conversation text in the text category label space;

The category label probability vector acquisition unit is configured to normalize the probability of the state vector of the standard conversation text in the text category label space to obtain the category label probability vector corresponding to the standard conversation text.
The apparatus of claim 11, further comprising:

The input sentence classification deviation acquisition module is configured to sum the translation deviation of the input sentence into the standard conversation text and the classification deviation of the text classification of the standard conversation text to obtain the input sentence classification deviation;

The parameter updating module is configured to update the target parameter matrix by minimizing the input sentence classification deviation.
An electronic device, including:

processor;

And a memory, where computer readable instructions are stored on the memory, and when the computer readable instructions are executed by the processor, the processor is used to implement the following steps:

Obtain the input sentence of the customer service robot in the ongoing conversation, convert the input sentence into standard conversation text, and the input sentence is a conversation message waiting for the customer service robot to process a response;

Obtaining a semantic vector corresponding to the standard conversation text by performing semantic feature extraction on the standard conversation text;

Performing category label prediction on the semantic vector of the standard conversation text to obtain a category label probability vector corresponding to the standard conversation text;

The category corresponding to the maximum probability label is selected from the category label probability vector as the category of the standard conversation text, and the category is used to assist in performing the response of the customer service robot to the input text.
The electronic device of claim 13, wherein the input sentence of the customer service robot in the ongoing conversation is converted into standard conversation text, and the processor is used to implement the following steps:

Encoding the input sentence to extract key semantic features of the input sentence;

Decode the key semantic features to obtain standard conversation text corresponding to the input sentence.
The electronic device according to claim 14, wherein the key semantic feature of the input sentence is extracted by encoding the input sentence, and the processor is configured to implement the following steps:

Obtaining the word vector corresponding to the word in the input sentence by vectorizing the word in the input sentence;

The word vectors corresponding to the words in the input sentence are traversed in chronological order, and the first hidden state vector obtained by the traversal is extracted as the semantic vector of the input sentence.
The electronic device according to any one of claims 13 to 15, wherein the semantic vector corresponding to the standard conversation text is obtained by performing semantic feature extraction on the standard conversation text, and the processor is configured to implement the following steps:

Acquiring a second hidden state vector obtained by decoding the key semantic features, and forming a hidden state vector matrix from the second hidden state vector;

Performing semantic feature extraction on the standard conversation text according to the hidden state vector matrix;

The semantic vector corresponding to the standard conversation text is obtained by pooling the extracted semantic features.
The electronic device according to any one of claims 13 or 16, wherein the class label prediction is performed on the semantic vector of the standard conversation text to obtain a category label probability vector corresponding to the standard conversation text, and the processor uses To achieve the following steps:

Performing a non-linear mapping on the semantic vector of the standard conversation text through the target parameter matrix to obtain the state vector of the standard conversation text in the text category label space;

Probability normalization is performed on the state vector of the standard conversation text in the text category label space to obtain a category label probability vector corresponding to the standard conversation text.
The electronic device of claim 17, the processor is further configured to implement the following steps:

Summing the translation deviations of converting the input sentence into the standard conversational text and the classification deviation of the text classification of the standard conversational text to obtain the input sentence classification deviation;

The target parameter matrix is updated by minimizing the input sentence classification deviation.
A computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the processor is used to implement the following steps:

Obtain the input sentence of the customer service robot in the ongoing conversation, convert the input sentence into standard conversation text, and the input sentence is a conversation message waiting for the customer service robot to process a response;

Obtaining a semantic vector corresponding to the standard conversation text by performing semantic feature extraction on the standard conversation text;

Performing category label prediction on the semantic vector of the standard conversation text to obtain a category label probability vector corresponding to the standard conversation text;

The category corresponding to the maximum probability label is selected from the category label probability vector as the category of the standard conversation text, and the category is used to assist in performing the response of the customer service robot to the input text.
The computer-readable storage medium according to claim 19, wherein the input sentence of the customer service robot in the ongoing conversation is converted into standard conversation text, and the processor is used to implement the following steps:

Encoding the input sentence to extract key semantic features of the input sentence;

Decode the key semantic features to obtain standard conversation text corresponding to the input sentence.
The computer-readable storage medium of claim 20, wherein the key semantic feature of the input sentence is extracted by encoding the input sentence, and the processor is configured to implement the following steps:

Obtaining the word vector corresponding to the word in the input sentence by vectorizing the word in the input sentence;

The word vectors corresponding to the words in the input sentence are traversed in chronological order, and the first hidden state vector obtained by the traversal is extracted as the semantic vector of the input sentence.
The computer-readable storage medium according to any one of claims 19 to 21, wherein the semantic vector corresponding to the standard conversation text is obtained by performing semantic feature extraction on the standard conversation text, and the processor is used to implement The following steps:

Acquiring a second hidden state vector obtained by decoding the key semantic features, and forming a hidden state vector matrix from the second hidden state vector;

Performing semantic feature extraction on the standard conversation text according to the hidden state vector matrix;

The semantic vector corresponding to the standard conversation text is obtained by pooling the extracted semantic features.
The computer-readable storage medium according to any one of claims 19 or 22, wherein the class label prediction is performed on the semantic vector of the standard conversation text to obtain a class label probability vector corresponding to the standard conversation text. The processor is used to implement the following steps:

Performing a non-linear mapping on the semantic vector of the standard conversation text through the target parameter matrix to obtain the state vector of the standard conversation text in the text category label space;

Probability normalization is performed on the state vector of the standard conversation text in the text category label space to obtain a category label probability vector corresponding to the standard conversation text.
The computer-readable storage medium of claim 23, the processor is further configured to implement the following steps:

Summing the translation deviations of converting the input sentence into the standard conversational text and the classification deviation of the text classification of the standard conversational text to obtain the input sentence classification deviation;

The target parameter matrix is updated by minimizing the input sentence classification deviation.