User intention recognition method and system based on outbound robot

Abstract

The invention discloses a user intention recognition method and system based on an outbound robot, and relates to the technical field of intention recognition. Comprising the following steps: acquiring voice data of known task intentions including confirmation, denying, query intentions and themes, extracting text features, and obtaining intention feature comprehensive vectors through a content awareness network model; a neural network model is utilized, the comprehensive vector is used as input, known intentions and themes are used as labels, and training is completed; extracting speech features such as speech speed, volume, fundamental frequency and the like from speech data to be recognized, and calculating a speech judgment coefficient; generating a comprehensive task intention judgment index in combination with the predicted task intention and the voice judgment coefficient, and comparing the comprehensive task intention judgment index with a threshold to confirm the user intention; and finally providing a response through the outbound robot in combination with the task theme. The method not only can optimize the final judgment of the user intention in combination with the dynamic change of the voice features, but also can significantly reduce the occurrence of misjudgment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intent recognition, in particular to a user intent recognition method and system based on an outbound robot. BACKGROUND

[0002] With the rapid development of artificial intelligence and natural language processing technology, intelligent voice interaction has become one of the important ways of human-computer interaction. In particular, outbound robot technology is widely used in marketing, customer service, information collection and other fields. In these application scenarios, outbound robots need to be able to efficiently interact with users through voice, understand the real intentions of users, and provide accurate responses based on user feedback. However, in actual applications, due to the complexity of voice interaction involving voice features, text features and user emotional factors, accurate recognition of user intent and task theme still faces many challenges.

[0003] Currently, outbound robots in user intent recognition usually rely on text analysis methods, converting user voice signals into text data through speech-to-text technology (STT), and further determining the intent based on keyword matching or rule models. However, this text-based analysis method has the following problems: First, only language expression information after speech-to-text is used in the text analysis process, while non-verbal features such as speech rate, volume and tone are ignored. These voice features often provide key information, such as user emotional state and expression, which are important for further refining and accurately determining intent. Second, traditional text analysis methods have limited understanding of complex task intent, especially when distinguishing between confirmation intent, denial intent and query intent, etc. Due to the similar text expression forms of different intents, text semantic-based judgment may be confused.

[0004] Therefore, a user intent recognition method that can comprehensively use voice features and text features is needed to improve the accuracy of outbound robot intent recognition and response efficiency. In particular, by introducing voice feature parameters such as speech rate, volume and voice fundamental frequency, combining content perception network model and neural network learning method, explicit and implicit information in user voice interaction can be more comprehensively mined, thereby achieving accurate analysis of user task intent and task theme.

[0005] In the prior art, a method and system for identifying user intent based on feature fusion are disclosed in CN112102013A, which comprises: obtaining the current call text of a target user and the label features of the target user; obtaining a pre-constructed feature fusion model, wherein the feature fusion model is constructed based on at least the historical call text of the target user and the label features of the target user; inputting the current call text of the target user and the label features of the target user into the feature fusion model for identification to obtain the intent recognition result of the target user. However, this method is only based on text features including the current call text and the historical call text as well as the label features, ignoring non-verbal information in the speech signal such as speech speed, volume, and fundamental frequency, lacking the support of speech features, which makes it difficult for the method to capture the rich emotions and tone information contained in the user intent, especially in the case of unclear intent expression, and only relying on the call text and the historical call text to identify the user intent may result in a high misjudgment rate, and also reduces the accuracy and effectiveness of the identification system.

[0006] The above information disclosed in the background section is only intended to enhance the understanding of the background of the present disclosure, and thus it can include information that does not constitute the prior art known to those of ordinary skill in the art. SUMMARY

[0007] The present application aims to provide a user intent recognition method and system based on an outbound robot to solve the problems raised in the background.

[0008] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0009] A user intent recognition method based on an outbound robot, comprising the following specific steps:

[0010] Collecting voice data of a plurality of known task intents and task subjects, extracting corresponding text feature data from the voice data, and decoding the text feature data to obtain voice text data, wherein the task intent includes confirmation intent, denial intent, and query intent;

[0011] Performing feature extraction on the voice text data based on a content perception network model to obtain an intent feature comprehensive vector, which includes text description data of the task intent and the task subject;

[0012] According to the obtained intent feature comprehensive vector, a neural network learning model is established, the obtained intent feature comprehensive vector is taken as the input of the neural network learning model, and the known corresponding task intent and task subject are taken as the label to train the neural network learning model, to obtain an intent analysis prediction model with the input being the intent feature comprehensive vector and the output being the corresponding task intent and task subject;

[0013] The voice data of the user to be identified is obtained, converted into target voice text data, and the corresponding target intent feature comprehensive vector is obtained, and the corresponding voice feature parameters are extracted from the voice data, the target intent feature comprehensive vector is input into the trained intent analysis prediction model, and the task prediction intent and task prediction theme are obtained, wherein the voice feature parameters include the user's speech speed, volume and voice fundamental frequency value;

[0014] The voice judgment coefficient is calculated based on the extracted voice feature parameters, the comprehensive task intent judgment index is generated based on the obtained voice judgment coefficient and the task prediction intent, the comprehensive task intent judgment index is compared with the intent judgment threshold, the user intent is confirmed based on the comparison result, and the confirmed user intent is combined with the task prediction theme to give a corresponding reply through the outbound robot.

[0015] Further, the method for extracting and decoding the voice data to obtain the voice text data is as follows: inputting the voice data into a Wav2Vec2.0 model, using a CTC decoding method, and converting the voice data into voice text data.

[0016] Further, the specific steps of obtaining the intent feature comprehensive vector include: 1) converting the voice text data to obtain the corresponding input representation; 2) extracting features from the input representation to obtain the intent feature comprehensive vector;

[0017] The process of obtaining the input representation includes: constructing an intent vocabulary library and a theme scene vocabulary library, searching the intent vocabulary library and the theme scene vocabulary library based on the voice text data to obtain a knowledge vector of the voice text data, converting each word in the knowledge vector into a corresponding word vector, and combining the word vectors to obtain the input representation of the voice text data;

[0018] The intent vocabulary library is specifically represented as:

[0019]

[0020] In the formula, represents the intent vocabulary library, represents the pth emotional vocabulary, wherein p is the index of the emotional vocabulary, , is the total number of emotional vocabularies;

[0021] The theme scene vocabulary library is specifically represented as:

[0022]

[0023] In the formula, represents the theme scene vocabulary library, is the pth theme scene vocabulary, wherein p is the index of the theme scene vocabulary, a topic scene vocabulary, wherein an index of the topic scene vocabulary, , a total number of the topic scene vocabulary;

[0024] wherein the knowledge vector is composed of words in the speech text data that are repeated in the intent vocabulary library and the topic scene vocabulary library;

[0025] wherein the word vector is obtained by converting the knowledge vector by a Word2Vec model;

[0026] wherein the input representation is an m x z matrix, m represents the number of words in the knowledge vector, z represents the dimension of the word vector, and .

[0027] Further, the formula for feature extraction of the input representation is:

[0028]

[0029] wherein, is the output of the i-th word vector in the input representation at the j-th hidden layer node, wherein j is the index of the hidden layer node, , represents the total number of hidden layer nodes, is a parameterized activation function, represents the k-th element in the i-th word vector, is the weight between the j-th hidden layer node and the k-th word vector element, and k represents the index of the element in the word vector, wherein ;

[0030] wherein the weight between the j-th hidden layer node and the k-th word vector element is an element in the weight matrix from the input layer to the hidden layer , and the dimension of the weight matrix is z x h.

[0031] Further, the input speech text data is converted and output by the content-aware network model into an intent feature comprehensive vector, and the specific formula for calculating the intent feature comprehensive vector is:

[0032]

[0033]

[0034] wherein, is the output of the i-th word vector after passing through all the hidden layer nodes, is the intent feature comprehensive vector corresponding to the speech text data;

[0035] According to the obtained intention feature comprehensive vector, a neural network learning model is established, wherein a graph analysis prediction model is established based on a convolutional neural network, wherein the convolutional neural network is composed of an input layer, a convolutional layer, a pooling layer, a full connection layer and an output layer, and the activation function in the convolutional layer is function.

[0036] For the full connection layer, the number of full connection layer neurons is set to 32, the initial neural network learning rate is set to 0.001, and the training round number is 100.

[0037] Further, a speech judgment coefficient is calculated based on the extracted speech feature parameters, wherein the formula for calculating the speech judgment coefficient is:

[0038]

[0039] In the formula, is the speech judgment coefficient, is the speech speed of the user voice, is the volume of the user voice, is the fundamental frequency variation of the user voice, , and are weight coefficients of the fundamental frequency variation, the speech speed and the volume of the user voice, respectively, wherein and , and are all greater than 0.

[0040] The fundamental frequency variation of the user voice is calculated by the end fundamental frequency and the start fundamental frequency of the sentence, and the specific calculation formula is:

[0041]

[0042] In the formula, is the end fundamental frequency, is the start fundamental frequency.

[0043] Further, according to the obtained speech judgment coefficient, the task prediction intention is combined to generate a comprehensive task intention judgment index, wherein the logic for calculating the comprehensive task intention judgment index is: numerical coding is performed on the three task intentions, and calculation is performed based on the corresponding numerical values after coding.

[0044] The specific logic for numerical coding of the three task intentions is: in ascending order, the confirm intention, the deny intention and the query intention are sequentially numerically coded, that is, the query intention has the maximum numerical value, and the confirm intention has the minimum numerical value.

[0045] The specific formula for calculating the comprehensive task intention judgment index is:

[0046]

[0047] In the formula, is a comprehensive task intention judgment index, is an encoding value corresponding to the task prediction intention, and are weight coefficients of the task prediction intention and the speech judgment coefficient respectively, wherein and are greater than 0;

[0048] The comprehensive task intention judgment index is compared with an intention judgment threshold value, and the user intention is confirmed according to the comparison result, wherein the specific logic for confirming the user intention is as follows:

[0049] When , it is judged that the user intention is a query intention, indicating that the user wants to consult;

[0050] When , it is judged that the user intention is a denial intention, indicating that the user denies or refuses the current dialogue content;

[0051] When , it is judged that the user intention is a confirmation intention, indicating that the user agrees or is interested in the current dialogue content;

[0052] wherein is a pre-set intention judgment threshold value.

[0053] The application also provides a user intention recognition system based on an outbound robot, which is used to execute the above-mentioned user intention recognition method based on an outbound robot and comprises:

[0054] A text data conversion module is configured to collect speech data of a plurality of known task intentions and task subjects, extract corresponding text feature data from the speech data, and decode the text feature data to obtain speech text data, wherein the task intention includes a confirmation intention, a denial intention and a query intention.

[0055] A feature data extraction module is configured to perform feature extraction on the speech text data based on a content perception network model to obtain an intention feature comprehensive vector, which includes text description data of the task intention and the task subject.

[0056] ​The intention analysis prediction module is used for establishing a neural network learning model according to the obtained intention feature comprehensive vector, taking the obtained intention feature comprehensive vector as the input of the neural network learning model, and training the neural network learning model by taking the known corresponding task intention and task theme as labels, so as to obtain an intention analysis prediction model with the input of the intention feature comprehensive vector and the output of the corresponding task intention and task theme.

[0057] The voice feature analysis module is used for obtaining voice data of the user to be recognized, converting the voice data into target voice text data, obtaining the corresponding target intention feature comprehensive vector, extracting the corresponding voice feature parameters from the voice data, inputting the target intention feature comprehensive vector into the intention analysis prediction model which has been trained, obtaining the task predicted intention and the task predicted theme, and the voice feature parameters include the speech speed, the volume and the voice fundamental frequency value of the user.

[0058] The accurate intention recognition module is used for calculating a voice judgment coefficient based on the extracted voice feature parameters, generating a comprehensive task intention judgment index according to the obtained voice judgment coefficient and the task predicted intention, comparing the comprehensive task intention judgment index with an intention judgment threshold, confirming the user intention according to the comparison result, and giving a corresponding reply through the outbound robot according to the confirmed user intention and the task predicted theme.

[0059] Compared with the prior art, the present application has the following beneficial effects:

[0060] Firstly, the scheme realizes the deep understanding of task intention and theme by establishing a content-aware network model. The content-aware network model can extract and enhance the key features of task description from voice text data, generate an intention feature comprehensive vector, and cover the intention information and task theme expressed by the user. It overcomes the limitations of traditional keyword matching or rule-based models, has stronger semantic perception ability, and can more accurately predict the user's interactive intention. At the same time, the intention feature comprehensive vector is trained by using a neural network learning model, which not only improves the model's learning ability for complex intentions, but also expands the applicability of the outbound robot in multi-task processing. Secondly, by introducing voice feature parameters such as speech rate, volume and fundamental frequency, the scheme can distinguish between confirmation intention, denial intention and query intention according to the explicit and implicit signals in the user's voice expression. For example, confirmation intention is usually accompanied by a slow speech rate and a slight increase in fundamental frequency, while denial intention may be accompanied by a faster speech rate, an increase in volume or fluctuations in fundamental frequency. By integrating voice features and text features, the scheme can greatly improve the accuracy of intention analysis and solve the problem of traditional methods that are difficult to distinguish between refined intentions under similar semantic expressions. Finally, by combining the voice judgment coefficient with the intention prediction model, the reliability and robustness of user intention recognition are further improved. In view of the uncertainty and diversity of user expression, this method not only optimizes the final determination of user intention by combining the dynamic changes of voice features, but also significantly reduces the occurrence of misjudgment. BRIEF DESCRIPTION OF DRAWINGS

[0061] Figure 1 The whole method flowchart of the application is shown in the figure.

[0062] Figure 2 The whole system structure diagram of the application is shown in the figure. DETAILED DESCRIPTION

[0063] To make the purpose, technical scheme and advantages of the application clearer, the application will be further described in detail below with specific examples.

[0064] It should be noted that, unless otherwise defined, technical terms or scientific terms used in the present application shall be understood as having the usual meaning to those having ordinary skills in the art to which the present application pertains. The terms "first", "second", and similar words used in the present application do not represent any order, number, or importance, but are only used to distinguish different components. The terms "include", "contain", and similar words mean that the elements or objects before the words cover the elements or objects listed after the words and their equivalents, and do not exclude other elements or objects. The terms "connect" or "connected" and similar words are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. The terms "up", "down", "left", "right", and the like are only used to represent relative positional relationships, which can change accordingly when the absolute positions of the described objects change.

[0065] Embodiments:

[0066] Referring to Figure 1 The present application provides a technical solution:

[0067] A user intent recognition method based on an outbound robot, comprising the following specific steps:

[0068] Step 1: Collecting voice data of a plurality of known task intents and task topics, extracting corresponding text feature data from the voice data, and decoding the text feature data to obtain voice text data, wherein the task intents include confirmation intent, denial intent, and query intent.

[0069] The method for decoding the voice data to obtain the voice text data is as follows: inputting the voice data into a Wav2Vec2.0 model, using a CTC decoding method, and converting the voice data of the individual user into voice text data.

[0070] Wav2Vec2.0 is a self-supervised learning voice model developed by Facebook AI (now Meta), which is suitable for low-resource voice recognition tasks. CTC decoding is a decoding method widely used in sequence-to-sequence tasks such as voice recognition and handwriting recognition, which calculates the probability distribution of output characters and minimizes the difference between the target text sequence and the real label to output the correct character sequence. The Wav2Vec2.0 model and the CTC decoding method are both mature prior art, and their specific operation processes are not described here.

[0071] The task topics include appointment services, event invitations, after-sales services, product promotions, and the like.

[0072] Step 2: Feature extraction of the voice text data based on the content-aware network model to obtain an intent feature comprehensive vector, which includes task intent and text description data of task subject.

[0073] The specific steps of obtaining the intent feature comprehensive vector include: 1) converting the voice text data to obtain a corresponding input representation; 2) feature extraction of the input representation to obtain the intent feature comprehensive vector;

[0074] Wherein, the process of obtaining the input representation includes: constructing an intent vocabulary and a theme scene vocabulary, retrieving the intent vocabulary and the theme scene vocabulary based on the voice text data to obtain a knowledge vector of the voice text data, and converting each word in the knowledge vector into a corresponding word vector, and combining the word vectors to obtain the input representation of the voice text data;

[0075] Wherein, the intent vocabulary is specifically represented as:

[0076]

[0077] In the formula, represents the intent vocabulary, represents the pth emotional vocabulary, wherein p is the index of the emotional vocabulary, , is the total number of emotional vocabularies;

[0078] Wherein, the theme scene vocabulary is specifically represented as:

[0079]

[0080] In the formula, represents the theme scene vocabulary, is the pth theme scene vocabulary, wherein is the index of the theme scene vocabulary, , is the total number of theme scene vocabularies;

[0081] Wherein, the knowledge vector is composed of words in the voice text data that are repeated in the intent vocabulary and the theme scene vocabulary;

[0082] Wherein, the word vector is obtained by converting the knowledge vector by the Word2Vec model;

[0083] Wherein, the input representation is an m×z matrix, m represents the number of words in the knowledge vector, z represents the dimension of the word vector, and .

[0084] ​The Word2Vec is a model for embedding words into a continuous vector space, which is a classical word embedding method proposed by Google. The goal of Word2Vec is to learn the relationship between words and their context, and capture the semantic information of words.

[0085] The words in these vocabulary libraries are selected based on in-depth analysis of a large amount of voice data from actual scene user daily conversations. The intent vocabulary library covers the words that users express task intent may use when they feel positive, negative, questioning, and other emotions, such as "OK" "No need" "How much is it?" and so on; the theme scene vocabulary library contains theme description words that often appear in events, such as "party", "event", "appointment", "product", etc. Each word is assigned a vector representation generated based on the Kolmogorov-Arnold representation theorem, which can accurately capture the intent and theme characteristics of the word in the dialogue context.

[0086] Each word in the knowledge vector is converted into a word vector, and the dimension of the word vector model is set to z. The input representation of the entire voice text is constructed based on the word vector, which is specifically represented as an m×z matrix, where m represents the number of words in the knowledge vector;

[0087] The formula for feature extraction of the input representation is:

[0088]

[0089] In the formula, is the output of the i-th word vector in the input representation at the j-th hidden layer node, where j is the index of the hidden layer node, , represents the total number of hidden layer nodes, is a parameterized activation function, represents the k-th element in the i-th word vector, is the weight between the j-th hidden layer node and the k-th word vector element, and k represents the index of the element in the word vector, .

[0090] where the weight between the j-th hidden layer node and the k-th word vector element is an element in the weight matrix from the input layer to the hidden layer , and the dimension of the weight matrix is z×h.

[0091] The input voice text data is converted into an intent feature comprehensive vector by the content-aware network model, and the specific formula for calculating the intent feature comprehensive vector is:

[0092]

[0093]

[0094] wherein, is the output of the i-th word vector after passing through all hidden layer nodes, is the intent feature integrated vector corresponding to the speech text data;

[0095] This step converts the speech text into a computer-processed mathematical form, preparing for subsequent operations such as feature extraction.

[0096] where the parameterized activation function refers to those activation functions whose behavior is adjusted by parameters (such as weights and biases), wherein, is a parameterized activation function, and a specific function expression is set based on the Kolmogorov-Arnold representation theorem.

[0097] Step 3: Based on the obtained intent feature integrated vector, a neural network learning model is established, taking the obtained intent feature integrated vector as the input of the neural network learning model, and taking the known corresponding task intent and task theme as the label to train the neural network learning model, to obtain an intent analysis prediction model with the input of the intent feature integrated vector and the output of the corresponding task intent and task theme.

[0098] Based on the obtained intent feature integrated vector, a neural network learning model is established, wherein a convolutional neural network is used to establish a graph analysis prediction model, wherein the convolutional neural network consists of an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer, and the activation function in the convolutional layer is ;

[0099] For the fully connected layer, the number of fully connected layer neurons is set to 32, the initial neural network learning rate is set to 0.001, and the training round number is set to 100.

[0100] Convolutional Neural Networks (CNN) are known for their ability to capture local features of data, and are particularly good at handling data with spatial structure. In the intent recognition task, the feature integrated vector can be regarded as a kind of high-dimensional data representation, and CNN can extract local feature patterns and cross-dimensional correlation characteristics of task intent from it.

[0101] The parameter sharing mechanism of convolutional neural network allows the same convolution kernel to be reused when processing features, which greatly reduces the number of model parameters and reduces the computational complexity. In the intent recognition task, the feature integrated vector is usually high-dimensional. CNN extracts local features through convolution operation and shares weights in different positions, which can effectively reduce resource consumption while maintaining high computational efficiency.

[0102] Step 4: Obtain the voice data of the user to be identified, convert it into target voice text data, and obtain the corresponding target intent feature comprehensive vector. At the same time, extract the corresponding voice feature parameters from the voice data, input the target intent feature comprehensive vector into the trained intent analysis prediction model, and obtain the task prediction intent and task prediction theme. The voice feature parameters include the user's speech speed, volume, and voice fundamental frequency value.

[0103] The specific method for extracting the corresponding voice feature parameters from the voice data is: the method used to extract the voice feature parameters of the corresponding user is MFCC processing, and the specific steps include: dividing the continuous voice signal into short time frames. Generally, the length of these frames is between 20 ms and 40 ms, and the overlapping part is generally taken as 50% of the frame length; window function processing, applying a window function (such as a Hamming window or a Hanning window) to each frame to reduce boundary effects. The window function can help smooth the signal and reduce frequency domain leakage; FFT conversion, performing fast Fourier transform (FFT) on the frame processed by the window function to convert the time domain signal to the frequency domain signal, which can obtain the frequency spectrum information of the signal; Mel filter, applying a Mel frequency filter bank to the FFT result. The Mel scale is a way to simulate human ear perception of different frequencies. The filter divides the frequency spectrum into multiple bandpass filters, allowing for higher resolution in the low frequency part; logarithmic operation, taking the logarithm of the output of the Mel filter. This step is to simulate the human auditory system's perception of sound amplitude. Finally, the voice feature parameters include the user's speech speed and volume.

[0104] The commonly used method for extracting the fundamental frequency includes autocorrelation method, zero-crossing method, and FFT harmonic analysis method. The fundamental frequency extraction method is a common prior art and will not be described here.

[0105] Step 5: Calculate the voice judgment coefficient based on the extracted voice feature parameters. According to the obtained voice judgment coefficient, combine the task prediction intent to generate a comprehensive task intent judgment index. Compare the comprehensive task intent judgment index with the intent judgment threshold value, confirm the user's intent according to the comparison result, and give a reply through the outbound robot according to the confirmed user's intent and the task prediction theme.

[0106] The voice judgment coefficient is calculated based on the extracted voice feature parameters, and the formula for calculating the voice judgment coefficient is:

[0107]

[0108] In the formula, is the voice judgment coefficient, is the user's speech speed, is the user's voice volume, a base frequency variation of the user voice, , and are weight coefficients of the base frequency variation, the speech rate and the volume of the user voice, respectively, wherein and , and are all greater than 0.

[0109] It should be noted that the speech judgment coefficient analyzes the user task intention by comprehensively considering the speech rate, the volume and the base frequency variation of the user voice, wherein the greater the value of the speech judgment coefficient , the closer the user's intention is to the query.

[0110] Among them, the base frequency is an important parameter in the speech, reflecting the high and low changes of the intonation, reflecting the fluctuation degree of the intonation in the user voice, and is usually closely related to the tone, emotion and expression method, such as the confirmation intention usually has smaller base frequency variation and more stable intonation, the denial intention may be accompanied by intonation rising or small fluctuation, and the query intention usually has larger base frequency variation, especially at the end of the sentence, which is manifested as intonation rising, so the base frequency variation of the user voice is directly proportional to the speech judgment coefficient , and the direct proportional relationship is reflected by the exponential function in the denominator, which embodies the importance of the base frequency feature to the intention judgment.

[0111] The speech rate is an important feature to judge the user emotion and intention, and the speech rate of the confirmation intention is generally moderate, appearing natural and calm, the speech rate of the denial intention may be accelerated, indicating the urgency of expressing one's own negative opinion, and the speech rate of the query intention is usually fast, indicating the urgency of obtaining uncertain information, so the speech judgment coefficient is directly proportional to the speech rate of the user voice, and the logarithmic function is used to represent the influence of compressing high speech rate values, avoiding excessive interference on the determination result when the speech rate is too high.

[0112] The volume reflects the loudness of the user's speech, which is usually related to the emotion intensity and the tone, and the volume of the confirmation intention is generally moderate, with a calm tone, the denial intention may be accompanied by a larger volume, because the user may be emotionally excited or try to emphasize his denial, and the query intention is accompanied by a larger volume, expressing the urgency of obtaining uncertain information, so the speech judgment coefficient is directly proportional to the volume of the user voice, and the square root is used to represent the influence of moderating the high volume value on the result, enhancing the robustness of the model to different volume levels.

[0113] Among them, the base frequency variation Directly related to the tone features of the query intention, directly reflects, shows the change of user tone emotion, is highly related to the judgment of intention (confirmation, denial, query), the speech speed is the secondary feature of the user when expressing the intention, is closely related to the anxious degree or emotion of the user, and the volume is related to the emotional intensity and expression method of the user, but the correlation with the specific intention is weak, so the setting And , And are all greater than 0.

[0114] The pitch variation of the user voice The greater, the greater the pitch variation of the user voice

[0115] The pitch variation of the user voice Is calculated by the end pitch and the start pitch of the sentence, and the specific calculation formula is:

[0116]

[0117] In the formula, The end pitch is The start pitch is.

[0118] The end pitch Specifically, the average pitch within 1 second before the end of the voice, and the start pitch Specifically, the average pitch within 1 second after the start of the voice.

[0119] According to the obtained voice judgment coefficient, the task predicted intention is combined to generate a comprehensive task intention judgment index, and the logic for calculating the comprehensive task intention judgment index is: numerical coding is performed on the three task intentions, and calculation is performed based on the corresponding numerical values after coding;

[0120] The specific logic of numerical coding for the three task intentions is: in ascending order, the confirmation intention, the denial intention and the query intention are sequentially numerically coded, that is, the query intention value is the largest, and the confirmation intention value is the smallest.

[0121] The specific formula on which the comprehensive task intention judgment index is calculated is:

[0122]

[0123] In the formula, The comprehensive task intention judgment index is The coding value corresponding to the task predicted intention is And The weight coefficients of the task predicted intention and the voice judgment coefficient respectively, wherein And And are all greater than 0;

[0124] wherein the comprehensive task intention judgment index is greater, indicating that the user's intention is closer to the query intention, and the above-mentioned encoding value corresponding to the task prediction intention and the speech judgment coefficient are directly proportional to the query intention, which will not be repeated here.

[0125] It should be noted that, since the task prediction intention is analyzed at the semantic level through text, the reflection on the intention is more direct and accurate, so the exponential function is expressed by an exponential function which indicates that the encoding value corresponding to the task prediction intention has a significant impact on the comprehensive task intention judgment index;

[0126] Due to the natural differences in voice and speaking habits of different users, the logarithmic function reduces the impact of the speech judgment coefficient on the comprehensive task intention judgment index.

[0127] Since the speech feature parameters differ for different users, and the encoding value corresponding to the task prediction intention is determined through related text, it has high accuracy, so and are both greater than 0.

[0128] Compare the comprehensive task intention judgment index with the intention judgment threshold, and according to the comparison result, confirm the user's intention, wherein the specific logic for confirming the user's intention is:

[0129] When , it is judged that the user's intention is the query intention, indicating that the user wants to consult;

[0130] When , it is judged that the user's intention is the denial intention, indicating that the user denies or refuses the current dialogue content;

[0131] When , it is judged that the user's intention is the confirmation intention, indicating that the user agrees or is interested in the current dialogue content;

[0132] wherein is a pre-set intention judgment threshold.

[0133] Please refer to Figure 2 , the present application also provides a user intention recognition system based on an outbound robot, which is used to execute the above-mentioned user intention recognition method based on an outbound robot, comprising:

[0134] ​The text data conversion module is configured to collect voice data of known task intentions and task subjects, extract corresponding text feature data from the voice data, and decode the text feature data to obtain voice text data, wherein the task intentions include confirmation intentions, denial intentions, and query intentions.

[0135] The feature data extraction module is configured to extract features from the voice text data based on a content-aware network model to obtain an intention feature comprehensive vector including text description data of the task intentions and the task subjects.

[0136] The intention analysis prediction module is configured to establish a neural network learning model based on the obtained intention feature comprehensive vector, train the neural network learning model by taking the obtained intention feature comprehensive vector as an input and the known corresponding task intentions and task subjects as labels, and obtain an intention analysis prediction model with the input being the intention feature comprehensive vector and the output being the corresponding task intentions and task subjects.

[0137] The voice feature analysis module is configured to obtain voice data of a user to be recognized, convert the voice data into target voice text data, obtain corresponding target intention feature comprehensive vectors, extract corresponding voice feature parameters from the voice data, input the target intention feature comprehensive vectors into the trained intention analysis prediction model, obtain predicted task intentions and predicted task subjects, and the voice feature parameters include the user's speech speed, volume, and voice fundamental frequency value.

[0138] The accurate intention recognition module is configured to calculate a voice judgment coefficient based on the extracted voice feature parameters, generate a comprehensive task intention judgment index based on the obtained voice judgment coefficient and the predicted task intentions, compare the comprehensive task intention judgment index with an intention judgment threshold, confirm the user's intention according to the comparison result, and give a corresponding reply through an outbound robot based on the confirmed user's intention and the predicted task subject.

[0139] The above formulas are all dimensionless values calculated, the formulas are obtained by software simulation of a large amount of data to obtain a formula of the nearest real situation, and the preset parameters in the formulas are set by a person skilled in the art according to actual conditions.

[0140] The above embodiments can be realized by software, hardware, firmware, or any combination thereof, in whole or in part. When realized by software, the above embodiments can be realized in the form of a computer program product in whole or in part. A person skilled in the art can realize that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized by hardware or software methods depends on the specific application and design constraints of the technical solutions.

[0141] The units described as separate components may or may not be physically separate, and the components displayed as units may or may not be physical units, which may be located in one place, or distributed on multiple network units. Part or all of the units can be selected to achieve the purpose of the embodiment scheme according to actual needs.

[0142] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A method for identifying user intention based on an outbound call robot, characterized in that: The specific steps include: Collecting voice data of known task intention and task subject, extracting corresponding text feature data from the voice data, and decoding the text feature data to obtain voice text data, the task intention including confirmation intention, denial intention and query intention; Based on the content-aware network model, the voice text data is extracted to obtain an intention feature comprehensive vector, which includes the text description data of the task intention and the task subject; According to the obtained intention feature comprehensive vector, a neural network learning model is established, the obtained intention feature comprehensive vector is taken as the input of the neural network learning model, and the known corresponding task intention and task subject are taken as the label to train the neural network learning model, so as to obtain an intention analysis prediction model with the input of the intention feature comprehensive vector and the output of the corresponding task intention and task subject; Obtain the voice data of the user to be identified, convert it into target voice text data, and obtain the corresponding target intention feature comprehensive vector, while extracting the corresponding voice feature parameters from the voice data, input the target intention feature comprehensive vector into the trained intention analysis prediction model, obtain the task prediction intention and task prediction subject, and the voice feature parameters include the user's speech speed, volume and voice fundamental frequency value; Based on the extracted voice feature parameters, a voice judgment coefficient is calculated, and according to the obtained voice judgment coefficient, the comprehensive task intention judgment index is generated combined with the task prediction intention, the comprehensive task intention judgment index is compared with the intention judgment threshold, and according to the comparison result, the user intention is confirmed, and according to the confirmed user intention, the task prediction subject is combined to give a corresponding reply through the outbound robot. 2.The user intent recognition method based on an outbound robot according to claim 1, wherein: The method for extracting and decoding the voice data to obtain the voice text data is as follows: input the voice data into the Wav2Vec2.0 model, use the CTC decoding method, and convert the voice data into voice text data. 3.The user intent recognition method based on an outbound robot according to claim 2, wherein: The specific steps for obtaining the intention feature comprehensive vector include: 1) converting the voice text data to obtain the corresponding input representation; 2) extracting features from the input representation to obtain the intention feature comprehensive vector; The process of obtaining the input representation includes: constructing an intention vocabulary and a subject scene vocabulary, searching the intention vocabulary and the subject scene vocabulary based on the voice text data to obtain a knowledge vector of the voice text data, and converting each word in the knowledge vector into a corresponding word vector to obtain the input representation of the voice text data; The intention vocabulary is specifically represented as: In the formula, represents an intended vocabulary of words, represents the pth emotional word, wherein p is an index of the emotional word, , is the total number of emotional words; The subject scene vocabulary is specifically represented as: wherein, denotes a subject scene vocabulary, is the first subject scene vocabulary, wherein is an index of the subject scene vocabulary, , is a total number of subject scene vocabularies; The knowledge vector is composed of words repeated in the intention vocabulary and the subject scene vocabulary in the voice text data; The word vector is obtained by converting the knowledge vector by the Word2Vec model; wherein the input is represented as an mxz matrix, m represents the number of words in the knowledge vector, z represents the dimension of the word vector, and . 4.The method of identifying user intent based on an outbound robot according to claim 3, wherein: The formula for extracting features from the input representation is: wherein, is the output of the i-th word vector in the input representation at the j-th hidden layer node, where j is the index of the hidden layer node, , denotes the total number of hidden layer nodes, is a parameterized activation function, denotes the k-th element in the i-th word vector, is the weight between the j-th hidden layer node and the k-th word vector element, k denotes the index of the element in the word vector, wherein ; where wjkis the weight between the jth hidden layer node and the kth word vector element is the weight matrix from the input layer to the hidden layer is the weight matrix from the input layer to the hidden layer has dimension z x h. 5.The user intent recognition method based on an outbound robot according to claim 4, wherein: The input voice text data is converted into an intention feature comprehensive vector by the content-aware network model, and the specific formula for calculating the intention feature comprehensive vector is: In the formula, is the output of the i-th word vector after passing through all the hidden layer nodes, is the intent feature integrated vector corresponding to the speech text data; According to the obtained intention feature comprehensive vector, a neural network learning model is established, wherein a graph analysis prediction model is established based on a convolutional neural network, wherein the convolutional neural network is composed of an input layer, a convolutional layer, a pooling layer, a full connection layer and an output layer, wherein the activation function in the convolutional layer is a sigmoid function. For the fully connected layer, the number of fully connected layer neurons is set to 32, the initial neural network learning rate is set to 0.001, and the number of training rounds is 100. 6.The user intent recognition method based on an outbound robot according to claim 5, wherein: The speech judgment coefficient is calculated based on the extracted speech feature parameters, and the formula for calculating the speech judgment coefficient is: In the formula, is a voice judgment coefficient, is a speech rate of the user voice, is a volume of the user voice, is a fundamental frequency variation of the user voice, , and are weight coefficients of the fundamental frequency variation, the speech rate, and the volume of the user voice, respectively, wherein and , and are all greater than 0. The amount of change in the fundamental frequency of the user voice The amount of change in the fundamental frequency of the user voice is calculated by the end fundamental frequency and the start fundamental frequency of the sentence, and the specific calculation formula is as follows: Where, is the sentence-end fundamental frequency, is the fundamental frequency of the sentence start. 7.The user intent recognition method based on an outbound robot according to claim 6, wherein: According to the obtained speech judgment coefficient, the comprehensive task intention judgment index is generated by combining the task predicted intention, and the logic for calculating the comprehensive task intention judgment index is: the three task intentions are numerically coded, and the corresponding numerical values after coding are calculated; The specific logic for numerically coding the three task intentions is: in ascending order, the confirm intention, the deny intention and the query intention are sequentially numerically coded, that is, the query intention value is the largest, and the confirm intention value is the smallest. The specific formula for calculating the comprehensive task intention judgment index is: In the formula, is a comprehensive task intention judgment index, is an encoding value corresponding to the task prediction intention, and are weight coefficients of the task prediction intention and the speech judgment coefficient respectively, wherein and and are greater than 0. The comprehensive task intention judgment index is compared with the intention judgment threshold, and the user intention is confirmed according to the comparison result, and the specific logic for confirming the user intention is: When the user intention is determined as a query intention, it indicates that the user wants to make an inquiry. When the user intention is determined as a denial intention, indicating that the user denies or refuses the current dialogue content; When the user intention is a confirmation intention, indicating that the user agrees or is interested in the current dialogue content; wherein is a pre-set intention determination threshold value. 8.A user intent recognition system based on an outbound robot, characterized by: The user intention recognition system based on the outbound robot is used to execute the user intention recognition method based on the outbound robot in any one of claims 1-7, comprising: The text data conversion module is used to collect a plurality of speech data of known task intentions and task topics, extract corresponding text feature data from the speech data, and decode the text feature data to obtain speech text data, wherein the task intention includes confirm intention, deny intention and query intention; The feature data extraction module is used to extract features from the speech text data based on a content perception network model to obtain an intention feature comprehensive vector, which includes text description data of task intention and task subject; The intention analysis prediction module is used to establish a neural network learning model according to the obtained intention feature comprehensive vector, take the obtained intention feature comprehensive vector as the input of the neural network learning model, and train the neural network learning model with known corresponding task intention and task topic as labels to obtain an intention analysis prediction model with the input of the intention feature comprehensive vector and the output of the corresponding task intention and task topic. The speech feature analysis module is used to obtain the speech data of the user to be identified, convert it into target speech text data, and obtain the corresponding target intention feature comprehensive vector, while extracting the corresponding speech feature parameters from the speech data, inputting the target intention feature comprehensive vector into the trained intention analysis prediction model to obtain the task predicted intention and task predicted topic, wherein the speech feature parameters include the user's speech speed, volume and speech fundamental frequency value. The accurate intention recognition module is used to calculate the speech judgment coefficient based on the extracted speech feature parameters, generate the comprehensive task intention judgment index according to the obtained speech judgment coefficient, compare the comprehensive task intention judgment index with the intention judgment threshold, confirm the user intention according to the comparison result, and give a corresponding reply through the outbound robot according to the confirmed user intention and the task predicted topic.

Images