Depression detection model training method and device, equipment, medium and program product

Abstract

The application discloses a kind of training methods, device, equipment, medium and program product of depression detection model, involve depression detection technical field, comprising: obtaining the audio data of sample user and the text data associated with audio data;Through initial model, for each sample user, the acoustic feature of the audio data of the sample user is extracted, and the acoustic feature is extracted, to obtain audio feature;Through initial model, for each sample user, the text feature of the text data of the sample user is extracted;Through initial model, audio feature and text feature are fused to obtain fusion feature;Through initial model, according to fusion feature, determine the depression detection result of the sample user;According to the depression detection result of each sample user and the labeled depression result, the initial model is trained to obtain the depression detection model.The application improves the accuracy and robustness of depression detection.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of depression detection, and in particular to a depression detection model training method, device, equipment, medium and program product. BACKGROUND

[0002] Timely and accurate recognition of depression is crucial for early intervention and treatment. Traditional depression detection mainly relies on clinical scale assessment and manual interview, which has strong subjectivity and low efficiency. With the development of artificial intelligence technology, using machine learning and deep learning models to analyze user language and speech features to detect depression has become a research hotspot.

[0003] Currently, there is also a method in the prior art for detecting depression by constructing a multi-modal depression detection model, extracting text features and speech features for fusion. SUMMARY

[0004] The present application provides a depression detection model training method, device, equipment, medium and program product to improve the accuracy and robustness of depression detection.

[0005] In a first aspect, the present application provides a depression detection model training method, comprising:

[0006] Obtain audio data and text data associated with the audio data of a sample user; the sample user includes at least one depression user and at least one non-depression user;

[0007] Extract the acoustic features of the audio data of each sample user through the initial model, and perform feature extraction on the acoustic features to obtain audio features;

[0008] Extract the text features of the text data of each sample user through the initial model;

[0009] Fuse the audio features and text features through the initial model to obtain fusion features;

[0010] Determine the depression detection result of the sample user according to the fusion features through the initial model;

[0011] Train the initial model according to the depression detection results of each sample user and the labeled depression results to obtain a depression detection model.

[0012] In a second aspect, the present application also provides a depression detection model training device, comprising:

[0013] The acquisition module is configured to acquire audio data of sample users and text data associated with the audio data, wherein the sample users include at least one user with depressive emotion and at least one user without depressive emotion.

[0014] The first extraction module is configured to extract, by using the initial model, acoustic features of the audio data of each sample user, and perform feature extraction on the acoustic features to obtain audio features.

[0015] The second extraction module is configured to extract, by using the initial model, text features of the text data of each sample user.

[0016] The feature fusion module is configured to fuse, by using the initial model, the audio features and the text features to obtain fused features.

[0017] The result determination module is configured to determine, by using the initial model, a depressive detection result of each sample user according to the fused features.

[0018] The training module is configured to train the initial model according to the depressive detection results of the sample users and the labeled depressive results to obtain a depressive detection model.

[0019] In a third aspect, an electronic device is provided, and the electronic device comprises:

[0020] at least one processor; and

[0021] a memory connected with the at least one processor; wherein

[0022] The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the training method of the depressive detection model provided in any of the embodiments of the present application.

[0023] In a fourth aspect, a computer readable storage medium is provided, and the computer readable storage medium stores computer instructions, and the computer instructions are used to enable a processor to implement the training method of the depressive detection model provided in any of the embodiments of the present application when the processor executes the computer instructions.

[0024] In a fifth aspect, a computer program product is provided, and the computer program product comprises a computer program, and the computer program is used to implement the training method of the depressive detection model provided in any of the embodiments of the present application when the computer program is executed by a processor.

[0025] Compared with the prior art of extracting acoustic features as depression features, the application further extracts features from the acoustic features extracted from the audio data, can eliminate redundant information in the acoustic features, reduce the interference of redundant information on the audio features of the user, and further mine high-order features, so that the obtained audio features can more accurately reflect the audio characteristics of depression emotion, thereby improving the accuracy and robustness of the model for depression detection.

[0026] It should be understood that the content described in this part is not intended to identify key or important features of the embodiments of the application, nor is it used to limit the scope of the application. Other features of the application will become apparent from the following description. BRIEF DESCRIPTION OF DRAWINGS

[0027] In order to more clearly illustrate the technical solutions in the embodiments of the application, the drawings needed in the embodiment description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0028] Figure 1 is a flowchart of a depression detection model training method according to the first embodiment of the application;

[0029] Figure 2 is a flowchart of a depression detection model training method according to the second embodiment of the application;

[0030] Figure 3A is a flowchart of a depression detection model training method according to the third embodiment of the application;

[0031] Figure 3B is a flowchart of a depression detection method according to the third embodiment of the application;

[0032] Figure 3C is an architecture schematic diagram of a depression detection model according to the third embodiment of the application;

[0033] Figure 4 is a structural schematic diagram of a depression detection model training device according to the fourth embodiment of the application;

[0034] Figure 5 is a structural diagram of an electronic device for implementing a depression detection model training method according to the embodiment of the application. DETAILED DESCRIPTION

[0035] In order to better understand the technical scheme of the present application, the technical scheme in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative efforts should fall within the scope of the present application.

[0036] It should be noted that the terms "first" and "second" and the like in the specification and claims of the present application and the above-described drawings are used to distinguish similar objects, and do not necessarily indicate a specific order or a chronological sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0037] In the technical scheme of the embodiments of the present application, the acquisition, storage and application of audio data and text data and the like involved are in accordance with the relevant legal regulations and do not violate public order and good customs.

[0038] Embodiment one

[0039] Figure 1 A flowchart of a training method of a depression detection model is provided for the first embodiment of the present application. The present embodiment can be applicable to the training of a depression detection model. The method can be executed by a training device of a depression detection model. The training device of the depression detection model can be realized in the form of hardware and / or software and be specifically configured in an electronic device.

[0040] Referring to Figure 1 A training method of a depression detection model is shown, which comprises:

[0041] S101, acquiring audio data of a sample user and text data associated with the audio data.

[0042] In the present embodiment, the sample user can be a user providing sample data for training the model. The sample user can include, but is not limited to, at least one user in a depressive mood and at least one user not in a depressive mood. The user in a depressive mood can be a user in a depressive mood or a patient with depression, etc. The user not in a depressive mood can be a user not in a depressive mood or a patient not with depression, etc.

[0043] The sample data used for training the model is audio data of the sample user and text data associated with the audio data. The text data associated with the audio data can be text data in the same communication process as the audio data. For example, the audio data can be subjected to speech-to-text processing to obtain the text data associated with the audio data, or text data in the same communication process as the audio data can be obtained from a communication database of the sample user.

[0044] Specifically, under the condition of obtaining the authorization of the user, the audio data of at least one user in a depressed mood and the text data associated with the audio data are obtained, and the audio data of at least one user in a depressed mood and the text data associated with the audio data are used as the audio data of the sample user and the text data associated with the audio data.

[0045] In an optional embodiment, the audio data and the associated text data can be data in any language, for example, Chinese text and Chinese speech. The audio data and the associated text data can be obtained in various ways, for example, written records and their recordings obtained through volunteer investigation.

[0046] In an optional embodiment, the obtained audio data and the associated text data can be preprocessed. For example, for the text data, repeated characters and special symbols in the text data are deleted; words such as "hmm" and "then" that do not contribute much to the sentiment tendency are removed to obtain text data in the form of a text sequence; for the audio data, filtering or speech enhancement is performed on the audio data to remove environmental noise and device noise and improve the signal-to-noise ratio; silent segments in the audio data are detected and removed, and only audio segments with human voice are retained; if the audio data includes audio segments of multiple people, only the audio segments of the sample user in the audio data are retained, and the audio segments of other people are removed.

[0047] In S102, acoustic features of the audio data of each sample user are extracted by using the initial model, and the acoustic features are subjected to feature extraction to obtain audio features.

[0048] In this embodiment, the initial model can be a machine learning model with parameter initialization. The acoustic features can include, but are not limited to, at least one of the fluctuation amplitude of the fundamental frequency, the trend of the energy change, the speech rate, the pause frequency, and the mel-frequency cepstral coefficient. The acoustic features and the audio features can be represented in the form of a matrix or a vector.

[0049] Specifically, by using the initial model and a certain algorithm, acoustic features of the audio data of each sample user are extracted, and the acoustic features are subjected to feature extraction to obtain audio features.

[0050] S103, extracting, by the initial model, text features of the text data of each sample user.

[0051] In this embodiment, the text features can be represented in the form of a matrix or a vector. Specifically, the initial model can be used to extract, by using a certain algorithm, the text features of the text data of each sample user.

[0052] S104, fusing, by the initial model, the audio features and the text features to obtain fused features.

[0053] In this embodiment, the fused features can be the result of fusing the audio features and the text features. The fused features can be represented in the form of a matrix or a vector.

[0054] Specifically, the initial model can be used to fuse, by using a certain algorithm, the audio features and the text features to obtain the fused features.

[0055] S105, determining, by the initial model, a depression detection result of the sample user according to the fused features.

[0056] Specifically, the initial model can be used to determine, by using a certain algorithm, the depression detection result of the sample user according to the fused features.

[0057] Optionally, determining the depression detection result of the sample user according to the fused features includes: performing average pooling processing on the fused features to obtain a processing result; determining a depression probability of the sample user according to the processing result; comparing the depression probability with a set probability threshold to obtain a comparison result; and determining the depression detection result of the sample user according to the comparison result.

[0058] It should be noted that the set probability threshold can be set by the technician according to actual needs or practical experience, and the present application does not limit this.

[0059] Specifically, the initial model can be used to determine, by using a certain algorithm, the depression detection result of the sample user according to the fused features.

[0060] ;

[0061] ;

[0062] wherein, represents a nonlinear activation function; represents a weight matrix for linear transformation; denotes a weight matrix for linear transformation; denotes a processing result; and denotes a bias matrix; denotes a two-dimensional probability vector; denotes a normalized exponential function;

[0063] comparing the depression probability with a set probability threshold to obtain a comparison result; if the depression probability is greater than or equal to the set probability threshold, determining that the depression detection result is depression; otherwise, determining that the depression detection result is non-depression.

[0064] It can be understood that, by using the technical solution, the average pooling processing is performed on the fusion features to obtain a processing result; according to the processing result, the depression probability of the sample user is determined; and the comparison result between the depression probability and the set probability threshold is used to determine the depression detection result of the sample user. The extreme value in the fusion features can be processed by the average pooling processing, the local noise in the fusion features can be reduced to interfere with the overall features, and the overall feature trend in the fusion features can be comprehensively fused to improve the accuracy of the binary classification detection of depression and non-depression.

[0065] S106, training the initial model according to the depression detection result of each sample user and the labeled depression result to obtain a depression detection model.

[0066] In this embodiment, the labeled depression result can be the real depression result of the sample user manually labeled. Specifically, a cross-entropy loss function can be constructed according to the depression detection result and the labeled depression result; and an Adam optimizer can be used to update all parameters in the initial model by a back propagation algorithm.

[0067] In an optional embodiment, to prevent overfitting, regularization techniques such as random inactivation (dropout), weight decay, and early stopping can be added.

[0068] It should be noted that the training method of the depression detection model provided by the present application is not only suitable for training a binary classification (depression / non-depression) depression detection model, but also can be extended to a multi-classification task to identify different degrees of depression symptoms or distinguish depression from other emotional disorders.

[0069] Compared with the prior art of extracting acoustic features as depression features, the application further extracts features from the acoustic features extracted from the audio data, can eliminate redundant information in the acoustic features, reduce the interference of the redundant information on the audio features of the user, and further mine high-order features, so that the obtained audio features can more accurately reflect the audio characteristics of depression emotion, thereby improving the accuracy and robustness of the model in depression detection.

[0070] Embodiment two

[0071] Figure 2 A flowchart of a depression detection model training method provided by the second embodiment of the application, the second embodiment of the application optimizes and improves the extraction operation of the audio features on the basis of the technical solutions of the above-mentioned embodiments.

[0072] Further, the "extracting acoustic features to obtain audio features" is refined into "performing convolution processing on the acoustic features to obtain convolution features; extracting time-dependent features of the audio data according to the convolution features; extracting periodic features of the audio data according to the convolution features; performing time attention weighting processing on the sub-convolution features of each frame in the convolution features to obtain weighted convolution features; extracting linear-dependent features of the audio data; and fusing the time-dependent features, the periodic features, the weighted convolution features and the linear-dependent features to obtain the audio features", so as to improve the extraction operation of the audio features.

[0073] It should be noted that the parts not described in detail in the embodiments of the application can refer to the descriptions of the foregoing embodiments.

[0074] Referring to Figure 2 The depression detection model training method comprises the following steps:

[0075] S201, acquiring audio data of a sample user and text data associated with the audio data; the sample user includes at least one user with depression emotion and at least one user without depression emotion.

[0076] S202, extracting acoustic features of the audio data of each sample user by an initial model.

[0077] Optionally, the acoustic feature of the sample user is extracted by: dividing the audio data of the sample user into at least one frame of sub-audio data; determining, for each frame of sub-audio data, a mel-frequency cepstral coefficient and a pitch fundamental frequency of the frame of sub-audio data; determining a mel-spectrum diagram of the audio data; and fusing the mel-frequency cepstral coefficient and the pitch fundamental frequency of each frame of sub-audio data and the mel-spectrum diagram to obtain the acoustic feature of the sample user.

[0078] The sub-audio data can be obtained by dividing the audio data; the mel-spectrum diagram can be represented by a matrix; the acoustic feature can be represented by a matrix, and the matrix dimension can be T x n; T represents the number of frames, and n represents the dimension of each frame of feature.

[0079] Specifically, the audio data is divided into at least one frame of sub-audio data by framing and windowing; for each frame of sub-audio data, a short-time Fourier transform is performed on the frame of audio data to obtain a power spectrum of the frame of audio data; the power spectrum is weighted and summed by a mel filter bank to obtain output energy; the mel filter output energy is logarithmically transformed and discrete cosine transformed to obtain a mel-frequency cepstral coefficient; a pitch period of the frame of sub-audio data is calculated to determine a pitch fundamental frequency of the frame of sub-audio data; the mel filter output energy of each frame of sub-audio data is arranged in time sequence of time frames to form a mel-spectrum diagram; and the mel-frequency cepstral coefficient and the pitch fundamental frequency of each frame of sub-audio data and the mel-spectrum diagram are fused to obtain the acoustic feature of the sample user.

[0080] It can be understood that, by using the above technical solution, the mel-frequency cepstral coefficient reflecting the spectral shape of the audio, the pitch fundamental frequency reflecting the prosodic feature, and the mel-spectrum diagram reflecting the dynamic change details of the spectrum over time can be extracted from the audio data; and by fusing the mel-frequency cepstral coefficient and the pitch fundamental frequency of each frame of sub-audio data and the mel-spectrum diagram into the acoustic feature, the different features can complement each other's defects, so that the acoustic feature accurately reflecting the emotional feature is obtained.

[0081] S203, performing convolution processing on the acoustic feature by using an initial model to obtain a convolution feature.

[0082] In this embodiment, the convolution feature can be a result of the convolution processing on the acoustic feature; and the convolution feature can be represented by a matrix or a vector. The initial model can include a convolution module; the convolution module can be used to perform the convolution processing on the acoustic feature; and the convolution module includes at least one convolution kernel.

[0083] Specifically, the convolution feature is obtained by performing the convolution processing on the acoustic feature by using the convolution module in the initial model; and the convolution feature can be determined by the following formula:

[0084] ;

[0085] wherein, denotes a convolution feature; denotes a sub convolution feature of the i-th frame output by the j-th convolution kernel in the convolution feature; denotes a sigmoid activation function; denotes a weight matrix of the j-th convolution kernel; denotes an acoustic feature; denotes a number of frames covered by the convolution kernel; denotes a bias matrix of the j-th convolution kernel; denotes a convolution operation.

[0086] S204, extracting a time-dependent feature of the audio data according to the convolution feature through an initial model.

[0087] In the embodiment, the time-dependent feature can represent information correlation between sub audio data of adjacent frames in the audio data. The initial model can include a Gated Recurrent Unit (GRU). The Gated Recurrent Unit can be used to extract the time-dependent feature of the audio data.

[0088] Specifically, the convolution feature can be processed through the Gated Recurrent Unit to obtain a final hidden state of the Gated Recurrent Unit. The final hidden state of the Gated Recurrent Unit is determined as the time-dependent feature. For example, the final hidden state can be determined by the following formula:

[0089] ;

[0090] ;

[0091] ;

[0092] ;

[0093] wherein, denotes the final hidden state of the t-th frame; denotes an update gate of the t-th frame; denotes an element-wise multiplication; denotes the final hidden state of the t-1-th frame; denotes a candidate hidden state of the t-th frame; denotes a hyperbolic tangent activation function; denotes a reset gate of the t-th frame; denotes a weight matrix of the candidate hidden state; denotes a bias matrix of the candidate hidden state; denotes a sub convolution feature of the t-th frame in the convolution feature; This represents the weight matrix of the updated gate; This represents the bias matrix of the update gate; This represents the weight matrix of the reset gate; This represents the bias matrix of the reset gate.

[0094] S205. Using the initial model, extract the periodic features of the audio data based on the convolutional features.

[0095] In this embodiment, periodicity features can characterize the periodic repetition of audio data. The initial model may include a skip loop module; the skip loop module can extract periodicity features from the audio data through skip connections of fixed length.

[0096] Specifically, the convolutional features can be processed through a jump loop module to obtain the final hidden state of the jump loop module; the final hidden state of the jump loop module is then determined as a periodic feature; for example, the final hidden state can be determined by the following formula:

[0097] ;

[0098] ;

[0099] ;

[0100] ;

[0101] in, This represents the final hidden state of the jump in frame t; This represents the jump update gate for frame t; This represents the final hidden state of frame tp; Indicates the preset period length; This represents the candidate hidden state in frame t; This represents the jump reset gate in frame t; The weight matrix representing the candidate hidden states for skipping; The bias matrix represents the candidate hidden state; This represents the weight matrix of the jump update gate; The bias matrix represents the jump update gate; The weight matrix represents the jump reset gate; This represents the bias matrix of the reset gate.

[0102] S206. Using the initial model, perform temporal attention weighting on the sub-convolutional features of each frame in the convolutional features to obtain weighted convolutional features.

[0103] In this embodiment, the weighted convolution feature can be a result of time attention weighting processing on the sub convolution feature. The initial model can include a time attention module for calculating the weighted convolution feature.

[0104] Specifically, the weighted convolution feature can be calculated by the time attention module; for example, the weighted convolution feature can be determined by the following formula:

[0105] ;

[0106] ;

[0107] ;

[0108] wherein, denotes the weighted convolution feature; denotes the sub weighted convolution feature of the t-th frame in the weighted convolution feature; denotes the time attention weight; denotes the convolution feature of the t-n-th frame; denotes the time attention score; denotes the attention score vector; denotes the weight matrix in the process of calculating the time attention score; denotes the bias matrix in the process of calculating the time attention score.

[0109] S207, extracting a linear dependence feature of the audio data by the initial model.

[0110] In this embodiment, the linear dependence feature can be used to represent the association relationship between the audio data of a certain frame and the audio data of a set number of frames before the time point; wherein the set number can be set by the technical personnel according to actual demand or practical experience. The initial model can include an autoregressive module for extracting the linear dependence feature of the audio data.

[0111] Specifically, the linear dependence feature is extracted by processing the audio data by the autoregressive module; for example, the linear dependence feature can be determined by the following formula:

[0112] ;

[0113] wherein, denotes the linear dependence feature; denotes the sub linear dependence feature of the t-th frame in the linear dependence feature; denotes the preset frame number; denotes the sub audio data of the t-m-th frame; the m-th linear transformation matrix.

[0114] S208, fusing the time-dependent feature, the periodic feature, the weighted convolution feature and the linear-dependent feature through the initial model to obtain an audio feature.

[0115] Specifically, the audio feature can be determined by fusing the time-dependent feature, the periodic feature, the weighted convolution feature and the linear-dependent feature according to the first weight matrix, the second weight matrix, the third weight matrix, the fourth weight matrix and a fusion bias matrix; for example, the audio feature can be determined by the following formula:

[0116] ;

[0117] wherein, represents the audio feature; represents a sub-audio feature at the t-th moment in the audio feature; represents the first weight matrix; represents the second weight matrix; represents the third weight matrix; represents the fourth weight matrix; represents the fusion bias matrix.

[0118] S209, extracting a text feature of text data of each sample user through the initial model.

[0119] S210, fusing the audio feature and the text feature through the initial model to obtain a fusion feature.

[0120] S211, determining a depression detection result of the sample user according to the fusion feature through the initial model.

[0121] S212, training the initial model according to the depression detection result of each sample user and the labeled depression result to obtain a depression detection model.

[0122] The technical scheme of the embodiment of the application extracts the time-dependent feature of the audio data according to the convolution feature, extracts the periodic feature of the audio data according to the convolution feature, performs time attention weighting processing on the sub-convolution feature of each frame in the convolution feature to obtain the weighted convolution feature, extracts the linear-dependent feature of the audio data, fuses the time-dependent feature, the periodic feature, the weighted convolution feature and the linear-dependent feature to obtain the audio feature, which can capture the features of the audio data at different time scales, especially can capture the information correlation between the adjacent frame sub-audio data in the audio data and the periodic repetition characteristics, so that the audio feature can more accurately reflect the subtle change characteristics and the periodic repetition characteristics in the audio data of the user, and the model can more accurately distinguish the depressed users and the non-depressed users according to the audio feature.

[0123] Embodiment three

[0124] Figure 3A A flowchart of a training method of a depression detection model is provided for Embodiment Three of the present application. The present embodiment optimizes and improves the determination operation of the fusion features on the basis of the technical solutions of the above-mentioned embodiments.

[0125] Further, the operation of fusing the audio features and the text features to obtain the fusion features is refined as the operation of splicing the audio features and the text features to obtain spliced features, and the operation of modeling the spliced features by a Transformer decoder to obtain the fusion features, so as to perfect the training operation of the target depression detection model.

[0126] It should be noted that the parts not described in detail in the present embodiment can be referred to the descriptions of the foregoing embodiments.

[0127] Referring to Figure 3A The training method of a depression detection model includes the following steps.

[0128] S301, audio data of a sample user and text data associated with the audio data are obtained; the sample user includes at least one user with depressive emotion and at least one user without depressive emotion.

[0129] S302, for each sample user, acoustic features of the audio data of the sample user are extracted by an initial model, and the acoustic features are subjected to feature extraction to obtain audio features.

[0130] S303, for each sample user, text features of the text data of the sample user are extracted by the initial model.

[0131] Optionally, for each sample user, the text features of the text data of the sample user include the following steps: the text data is subjected to word segmentation to obtain segmentation results; each segmentation result is mapped to a vector to obtain a word embedding vector of the text data; for each segmentation result, a position encoding of the segmentation result is generated according to position information of the segmentation result; the word embedding vector and the position encoding are fused to obtain a fusion vector; the fusion vector is subjected to feature extraction to obtain the text features.

[0132] The fusion vector is a fusion result between the word embedding vector and the position code; the fusion can be represented in the form of a matrix. The segmentation result can be a sequence composed of segmented words; the order of the segmented words in the sequence corresponds to the position information of the segmented words in the text data; for example, the first segmented word in the text data has an order of 1 in the segmentation result; the segmented word adjacent to the first segmented word and arranged after the first segmented word in the text data has an order of 2 in the segmentation result. The position information can be used to represent the position of the segmented word in the text data. The initial model can include an embedding layer, and the embedding layer can be provided with a pre-trained word embedding sub-model, which can be used to convert text into a vector.

[0133] Specifically, the text data is segmented to obtain a segmentation result; for example, the segmentation result can be represented by the following formula:

[0134] ;

[0135] wherein, represents the segmentation result; represents the segmented word with an order of 1 in the segmentation result; represents the segmented word with an order of 2 in the segmentation result; represents the segmented word with an order of n in the segmentation result;

[0136] Each segmentation result is mapped to a vector by a pre-trained word embedding sub-model in the initial model to obtain a word embedding vector of the text data; for each segmentation result, a position code of the segmentation result is generated according to the position information of each segmentation result; the dimension of the position code is the same as the dimension of the word embedding vector.

[0137] For example, the position code can be determined by the following formula:

[0138] ;

[0139] ;

[0140] ;

[0141] wherein, represents the position of the segmentation result in the text data; represents the position code of the segmented word at the position; represents the position code of the segmented word at the position; represents the position code at an even dimension; represents the position code at an odd dimension; represents the number of dimensions of the word embedding vector.

[0142] The word embedding vector corresponding to each segmented word is added to the position encoding corresponding thereto to obtain a fusion vector; for example, the fusion vector can be determined by the following formula:

[0143]

[0144] wherein, represents the fusion vector; represents the word embedding vector corresponding to the segmented word ; represents the word embedding vector corresponding to the segmented word ; represents the word embedding vector corresponding to the segmented word ; represents the position encoding corresponding to the segmented word ; represents the position encoding corresponding to the segmented word ; represents the position encoding corresponding to the segmented word .

[0145] The initial model can also include a Transformer encoder; the Transformer encoder is stacked by a plurality of same Encoders, each Encoder containing a Multi-Head Attention mechanism and a Feed-Forward Neural Network, and each mechanism is connected by a Residual Connection and a Layer Normalization; the fusion vector is subjected to feature extraction by the Transformer encoder to obtain a text feature;

[0146] For example, the text feature can be determined by the following formula:

[0147]

[0148]

[0149]

[0150]

[0151]

[0152]

[0153]

[0154] ​​​​​​​​;

[0155] wherein, denotes the text feature, and also denotes the output result of the jth layer encoder of the transformer encoder; denotes a layer normalization function; denotes the output result of the multi-head attention sub-layer in the jth layer encoder of the transformer encoder; denotes a feed-forward neural network; denotes the output result of the j-1th layer encoder of the transformer encoder, and also denotes the input of the jth layer encoder; denotes a multi-head attention mechanism; denotes the query matrix obtained by the j-1th layer encoder performing linear transformation on ; denotes the key matrix obtained by the j-1th layer encoder performing linear transformation on ; denotes the value matrix obtained by the j-1th layer encoder performing linear transformation on ; denotes a feature concatenation function; denotes the scaled dot-product attention calculated by the 1st attention sub-head of the j-1th layer encoder; denotes the scaled dot-product attention calculated by the 2nd attention sub-head of the j-1th layer encoder; denotes the scaled dot-product attention calculated by the hth attention sub-head of the j-1th layer encoder; denotes a scaled dot-product attention function; denotes or a dimension value; denotes the query matrix obtained by the 1st layer encoder performing linear transformation on the fusion vector; denotes the query matrix obtained by the 1st layer encoder performing linear transformation on the fusion vector; denotes the value matrix obtained by the 1st layer encoder performing linear transformation on the fusion vector; denotes the linear transformation matrix for the hth attention sub-head of the j-1th layer encoder to transform ; denotes the linear transformation matrix for the hth attention sub-head of the j-1th layer encoder to transform ; denotes the linear transformation matrix for the hth attention sub-head of the j-1th layer encoder to transform ; denotes the linear transformation matrix for converting the fusion vector to the query matrix in the 1st layer encoder; represents a linear transformation matrix in the first layer encoder for converting the fusion vector into a key matrix; represents a linear transformation matrix in the first layer encoder for converting the fusion vector into a value matrix; represents a linear transformation matrix.

[0156] It can be understood that, by using the technical solutions described above, the text data is converted into a word embedding vector, position encoding is fused, and then the fusion vector is further feature extracted to obtain text features, which can deeply analyze the text data, extract semantic representation and emotional features about depression in the text data, and thus improve the accuracy of text features in representing depression characteristics.

[0157] S304, splicing the audio features and the text features through the initial model to obtain spliced features.

[0158] In this embodiment, the spliced features can be the result of splicing the audio features and the text features. The spliced features can be represented in the form of a matrix or a vector.

[0159] Specifically, the audio features can be linearly transformed to the same dimension as the text features; the audio features after dimension transformation are spliced with the linear features to obtain spliced features. Exemplarily, the spliced features can be represented by the following formula:

[0160] ;

[0161] ;

[0162] wherein, represents the spliced features; represents the audio features after dimension transformation; represents a weight matrix in the transformation process; represents a bias matrix in the transformation process.

[0163] S305, sequence modeling of the spliced features is performed through a Transformer decoder to obtain fusion features.

[0164] In this embodiment, the initial model can include a Transformer decoder; the Transformer decoder can be used for sequence modeling of the spliced features to obtain fusion features; the fusion features can be the result output by the Transformer decoder, i.e., the result of sequence modeling of the spliced features.

[0165] The Transformer decoder consists of multiple identical decoders stacked together. Each decoder contains three parts: a masked multi-head attention mechanism, an encoder-decoder multi-head attention mechanism, and a feed-forward neural network.

[0166] For example, the fusion features can be determined by the following formula:

[0167] ;

[0168] ;

[0169] ;

[0170] in, This represents the fused features, and is also the output of the j-th (last) layer decoder in the Transformer decoder; This represents the encoder-decoder multi-head attention operation result in the j-th layer decoder; This represents the result of the mask multi-head attention operation in the j-th layer decoder; Indicates splicing characteristics; This represents the output of the (j-1)th layer decoder in the Transformer decoder; This indicates a multi-head attention mechanism in the mask; This indicates the decoder pair for the (j-1)th layer. The query matrix is ​​obtained by performing a linear transformation, where the first-layer decoder can be obtained by performing a linear transformation on the concatenated features; This represents the key matrix obtained by performing a linear transformation on the concatenated features in the (j-1)th layer decoder; This represents the value matrix obtained by the (j-1)th layer decoder performing a linear transformation on the spliced ​​features.

[0171] Furthermore, in the masked multi-head attention operation, the masked multi-head attention result obtained from the operation of each sub-head can be expressed by the following formula:

[0172] ;

[0173] in, This represents the mask attention operation function; express or Dimensional values; M represents a mask matrix; in the mask matrix, if j>i, the element at position (i, j) is -∞, otherwise, the element at position (i, j) is 0. This ensures that position i can only focus on information from positions 0 to i.

[0174] S306, determining a depression detection result of the sample user according to the fusion feature through the initial model.

[0175] S307, training the initial model according to the depression detection results of the sample users and the labeled depression results to obtain a depression detection model.

[0176] Optionally, Figure 3B is a flowchart of a depression detection method. Figure 3B The depression detection method shown can be a method flowchart of a depression detection method performed by a depression detection model trained by a training method of a depression detection model provided by an embodiment of the present application. As shown in Figure 3B As shown, the text data of the user is preprocessed and text features are extracted, and the audio data of the user is preprocessed and audio features are extracted; the text features and the audio features are spliced to obtain spliced features; the spliced features are sequentially modeled to obtain fusion features; and the depression detection result of the user is output according to the fusion features.

[0177] Optionally, Figure 3C is an architecture schematic diagram of a depression detection model. As shown in Figure 3C As shown, the depression detection model includes an embedding layer, a position encoding, a Transformer encoder; an acoustic feature extraction layer, a convolution layer, a recurrent layer, a recurrent skip layer, a time attention layer, an autoregressive layer, a recurrent skip layer, a time attention layer, an autoregressive layer, an acoustic feature fusion layer, a feature splicing layer, a Transformer decoder, a pooling layer, and a classifier.

[0178] The Transformer encoder includes N identical encoders; each encoder includes a multi-head attention mechanism and a feedforward neural network; residual connections and layer normalization are used to connect between each mechanism; the Transformer decoder includes N identical decoders, and each decoder includes three parts: a masked multi-head attention mechanism, an encoder-decoder multi-head attention mechanism, and a feedforward neural network.

[0179] The text preprocessing layer can be used for preprocessing the text data input into the depression detection model, such as deleting repeated characters and repeated words; the embedding layer can be used for converting the output of the text preprocessing layer, that is, the preprocessed text data, into a word embedding vector; the position encoding layer can be used for generating position encoding of each token and fusing the position encoding with the word embedding vector; and the Transformer encoder can be used for processing the output of the position encoding layer through multi-head attention mechanism of a multi-layer encoder, residual connection & layer normalization, feedforward neural network and residual connection & layer normalization to extract text features.

[0180] The audio preprocessing layer can be used for preprocessing the audio data input into the depression detection model, such as filtering or speech enhancement; the acoustic feature extraction layer can be used for extracting acoustic features from the preprocessed audio data; the convolution layer can be used for performing convolution operation on the acoustic features to extract convolution features; the recurrent layer can be used for extracting time-dependent features; the recurrent skip layer can be used for extracting periodic features; the temporal attention layer can be used for extracting weighted convolution features; the autoregressive layer can be used for extracting linearly dependent features; and the acoustic feature fusion layer can be used for fusing the acoustic features to obtain audio features.

[0181] The feature splicing layer can be used for splicing the audio features and the text features to obtain spliced features; the Transformer decoder can be used for performing sequence modeling on the spliced features through masked multi-head attention mechanism of a multi-layer decoder, residual connection & layer normalization, multi-head attention mechanism, residual connection & layer normalization, feedforward neural network and residual connection & layer normalization to obtain fusion features; the pooling layer can be used for performing average pooling processing on the fusion features; and the classifier can be used for outputting a depression detection result according to the output of the pooling layer.

[0182] The technical scheme of the embodiment of the present application splices the audio features and the text features to obtain spliced features, performs sequence modeling on the spliced features through the Transformer decoder to obtain fusion features, can perform cross-modal time alignment on the audio feature part and the text feature part in the spliced features, avoids the defect of information discontinuity in the fusion features compared with the technical scheme of directly splicing in the prior art, and improves the accuracy of depression detection by the model according to the fusion features.

[0183] Embodiment four

[0184] Figure 4 A structural schematic diagram of a depression detection model training device provided by the embodiment four of the present application. The embodiment of the present application can be applicable to the case of training a depression detection model, the device can execute a depression detection model training method, the depression detection model training device can be realized in the form of hardware and / or software, and the device can be configured in an electronic device.

[0185] Referring to Figure 4 The training device of the depression detection model shown in the embodiment of the present application comprises a first extraction module 401, a first fusion module 402, a second extraction module 403, a second fusion module 404, a result determination module 405 and a training module 406, wherein,

[0186] The first extraction module 401 is configured to extract acoustic features of audio data of each sample user by using an initial model, and perform feature extraction on the acoustic features to obtain audio features.

[0187] The first fusion module 402 is configured to fuse the time-dependent features, the periodic features, the weighted convolution features and the linear-dependent features to obtain audio features.

[0188] The second extraction module 403 is configured to extract text features of text data of each sample user by using the initial model.

[0189] The second fusion module 404 is configured to fuse the audio features and the text features by using the initial model to obtain fused features.

[0190] The result determination module 405 is configured to determine a depression detection result of the sample user according to the fused features by using the initial model.

[0191] The training module 406 is configured to train the initial model according to the depression detection results of the sample users and the labeled depression results to obtain a depression detection model.

[0192] In the embodiment of the present application, the acoustic features of the audio data of the sample user are extracted by using the initial model, and the feature extraction is performed on the acoustic features to obtain the audio features. The audio features and the text features are fused to obtain the fused features. Compared with the prior art in which the acoustic features are extracted as the depression features, the acoustic features extracted from the audio data are further subjected to the feature extraction in the present application, the redundant information in the acoustic features can be eliminated, the interference of the redundant information on the audio features of the user is reduced, and the high-order features are further mined, so that the obtained audio features can more accurately reflect the audio characteristics of the depression emotion, thereby improving the accuracy and robustness of the model in the depression detection.

[0193] Optionally, the first extraction module 401 comprises:

[0194] The extraction unit is configured to extract acoustic features of audio data of the sample user.

[0195] The first processing unit is configured to perform convolution processing on the acoustic features to obtain convolution features.

[0196] a second processing unit configured to extract time-dependent features of the audio data according to the convolution features;

[0197] a third processing unit configured to extract periodic features of the audio data according to the convolution features;

[0198] a fourth processing unit configured to perform time attention weighting processing on the sub-convolution features of each frame in the convolution features to obtain weighted convolution features.

[0199] Optionally, the extraction unit comprises:

[0200] a division sub-unit configured to divide the audio data of the sample user into at least one frame of sub-audio data;

[0201] a first determination sub-unit configured to determine, for each frame of sub-audio data, a mel-frequency cepstral coefficient and a pitch fundamental frequency of the frame of sub-audio data;

[0202] a second determination sub-unit configured to determine a mel-spectrum graph of the audio data;

[0203] a first fusion sub-unit configured to fuse the mel-frequency cepstral coefficient and the pitch fundamental frequency of each frame of sub-audio data, and the mel-spectrum graph to obtain the acoustic features of the sample user.

[0204] Optionally, the second fusion module 404 comprises:

[0205] a concatenation unit configured to concatenate the audio features and the text features to obtain concatenated features;

[0206] a sequence modeling unit configured to perform sequence modeling on the concatenated features by a Transformer decoder to obtain fused features.

[0207] Optionally, the second extraction module 403 comprises:

[0208] a word segmentation unit configured to perform word segmentation on the text data to obtain word segmentation results;

[0209] an embedding unit configured to map each word segmentation result into a vector to obtain word embedding vectors of the text data;

[0210] a generation unit configured to generate, for each word segmentation result, a position encoding of the word segmentation result according to position information of the word segmentation result;

[0211] a third fusion unit configured to fuse the word embedding vectors and the position encodings to obtain fused vectors;

[0212] an extraction unit configured to perform feature extraction on the fused vectors to obtain text features.

[0213] Optionally, the result determining module 405 comprises:

[0214] a pooling unit configured to perform average pooling processing on the fused features to obtain a processing result;

[0215] a probability determining unit configured to determine a depression probability of the sample user according to the processing result;

[0216] a comparison unit configured to compare the depression probability with a set probability threshold to obtain a comparison result;

[0217] a result determining unit configured to determine a depression detection result of the sample user according to the comparison result.

[0218] The training device of the depression detection model provided in the embodiment of the present application can execute the training method of the depression detection model provided in any embodiment of the present application, and has the function modules and beneficial effects corresponding to the training method of the depression detection model.

[0219] Embodiment five

[0220] Figure 5 A structural schematic diagram of a depression detection model training device 510 that can be used to implement the embodiments of the present application is shown. The depression detection model training device is intended to represent various forms of digital computers, such as laptops, desktops, workstations, personal digital assistants, servers, blade servers, mainframes, and other appropriate computers. The depression detection model training device can also represent various forms of mobile devices, such as personal digital processors, cellular telephones, smart phones, wearable devices (such as headgear, eyewear, watches, etc.), and other similar computing devices. The components shown here, their connections, and their functions, as well as the software implemented by the depression detection model training device, are meant to be examples only, and are not intended to limit the implementations of the present application described and / or claimed in this document.

[0221] As Figure 5As shown, the training device 510 of the depression detection model comprises at least one processor 511, and a memory, such as a read-only memory (ROM) 512, a random access memory (RAM) 513, etc., which is communicatively connected to the at least one processor 511, wherein the memory stores a computer program executable by the at least one processor. The processor 511 can perform various appropriate actions and processes according to the computer program stored in the read-only memory (ROM) 512 or loaded from the storage unit 518 to the random access memory (RAM) 513. In the RAM 513, various programs and data required for the operation of the training device 510 of the depression detection model can also be stored. The processor 511, the ROM 512, and the RAM 513 are connected to each other through a bus 514. An input / output (I / O) interface 515 is also connected to the bus 514.

[0222] Various components in the training device 510 of the depression detection model are connected to the I / O interface 515, including: an input unit 516, such as a keyboard, a mouse, etc.; an output unit 517, such as various types of displays, speakers, etc.; a storage unit 518, such as a magnetic disk, an optical disk, etc.; and a communication unit 519, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 519 allows the training device 510 of the depression detection model to exchange information / data with other devices through a computer network, such as the Internet, and / or various telecommunication networks.

[0223] The processor 511 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the processor 511 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any appropriate processor, controller, microcontroller, etc. The processor 511 performs various methods and processes described above, such as the training method of the depression detection model.

[0224] In some embodiments, the training method of the depression detection model can be implemented as a computer program tangibly embodied in a computer readable storage medium, such as the storage unit 518. In some embodiments, part or all of the computer program can be loaded and / or installed on the training device 510 of the depression detection model via the ROM 512 and / or the communication unit 519. When the computer program is loaded into the RAM 513 and executed by the processor 511, one or more steps of the training method of the depression detection model described above can be performed. Alternatively, in other embodiments, the processor 511 can be configured to perform the training method of the depression detection model by any other appropriate means (e.g., by means of firmware).

[0225] Various implementations of the systems and techniques described above can be realized in digital electronic circuitry, integrated circuitry, a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), a system on a chip (SOC), a complex programmable logic device (CPLD), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include implementation in one or more computer programs that are executable and / or interpretable on a programmable system including at least one programmable processor, which can be special or general purpose, coupled to receive data and instructions from, and to transmit data and instructions to, a storage system, at least one input device, and at least one output device.

[0226] Computer programs used to implement the methods of the application can be written in any combination of one or more programming languages. These computer programs can be provided to a processor of a general purpose computer, special purpose computer, or other programmable processing apparatus to produce a machine, such that the computer program, when executed by the processor of the computer or other programmable processing apparatus, enables the implementation of the functions / acts specified in the flow diagrams and / or block diagrams. The computer program can be executed entirely on a machine, partially on a machine, partially on a machine as part of a standalone software package, or entirely on a remote machine or server.

[0227] In the context of the present application, a computer-readable storage medium can be a tangible medium that can contain or store computer programs for use by or in connection with an instruction execution system, apparatus, or device. The computer-readable storage medium can include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. Alternatively, the computer-readable storage medium can be a machine-readable signal medium. More specific examples of the machine-readable storage medium will include one or more lines of electrical connections, portable computer disks, hard disk drives, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or Flash memory), optical fibers, portable compact disc read-only memories (CD-ROMs), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0228] To provide for interaction with a user, the systems and techniques described here can be implemented on a depression detection model training device having a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user and a keyboard and a pointing device (e.g., a mouse or a trackball) by which the user can provide input to the depression detection model training device. Other kinds of devices can be used to provide for interaction with a user as well; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form, including acoustic, speech, or tactile input.

[0229] The systems and techniques described here can be implemented in a computing system that includes a back end component (e.g., as a data server), or that includes a middleware component (e.g., an application server), or that includes a front end component (e.g., a user computer having a graphical user interface or a Web browser through which a user can interact with an implementation of the systems and techniques described here), or any combination of such back end, middleware, or front end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network (LAN), a wide area network (WAN), blockchain network, and the Internet.

[0230] The computing system can include clients and servers. A client and server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other. A server can be a cloud server, also known as cloud computing server or cloud host, which is a host product in the cloud computing service system, to solve the defects of large management difficulty and weak business scalability in traditional physical host and VPS (Virtual Private Server).

[0231] It should be understood that the various forms of flow shown above can be re-ordered, added to, or deleted from without departing from the scope of the present disclosure. For example, the steps recited in the present disclosure can be executed in parallel, executed in sequence, or executed in different orders, as long as the desired results of the technical solutions of the present disclosure are achieved, and the present disclosure is not limited herein.

[0232] The above detailed description does not limit the scope of the application. Various modifications, combinations, sub-combinations and alternatives can be made to the detailed description. Any modification, equivalent replacement and improvement made within the spirit and principle of the application shall fall within the scope of the application.

Claims

1. A training method for a depression detection model, characterized in that, The method comprises: extracting linear dependence features of each sample user through an initial model; dividing audio data of the sample user into at least one frame of sub-audio data; the sample user comprises a depressed mood user and a non-depressed mood user; determining a mel-frequency cepstrum coefficient and a pitch fundamental frequency of each frame of sub-audio data; determining a mel-spectrum graph of the audio data; fusing the mel-frequency cepstrum coefficient and the pitch fundamental frequency of each frame of sub-audio data and the mel-spectrum graph to obtain acoustic features of the sample user; performing convolution processing on the acoustic features to obtain convolution features; extracting time dependence features of the audio data according to the convolution features; extracting periodicity features of the audio data according to the convolution features; performing time attention weighting processing on each frame of sub-convolution features in the convolution features to obtain weighted convolution features; fusing the time dependence features, the periodicity features, the weighted convolution features and the linear dependence features to obtain audio features; extracting text features of text data of each sample user through the initial model; fusing the audio features and the text features through the initial model to obtain fused features; determining a depression detection result of the sample user according to the fused features through the initial model; training the initial model according to depression detection results and labeled depression results of each sample user to obtain a depression detection model.

2. The method of claim 1, wherein, The fusing the audio features and the text features to obtain fused features comprises: splicing the audio features and the text features to obtain spliced features; performing sequence modeling on the spliced features through a Transformer decoder to obtain fused features.

3. The method of claim 1, wherein, The extracting text features of text data of each sample user comprises: performing word segmentation on the text data to obtain word segmentation results; mapping each word segmentation result into a vector to obtain word embedding vectors of the text data; generating position encodings of each word segmentation result according to position information of the word segmentation results; fusing the word embedding vectors and the position encodings to obtain fused vectors; extracting features of the fused vectors to obtain text features.

4. The method of claim 1, wherein, The determining a depression detection result of the sample user according to the fused features comprises: performing average pooling processing on the fused features to obtain a processing result; determining a depression probability of the sample user according to the processing result; comparing the depression probability with a set probability threshold to obtain a comparison result; determining the depression detection result of the sample user according to the comparison result. 5.A device for training a depression detection model, comprising: The device comprises: a first extraction module configured to extract time dependence features, periodicity features, weighted convolution features and linear dependence features of audio data of each sample user through an initial model; the sample user comprises a depressed mood user and a non-depressed mood user. The first fusion module is configured to fuse the time-dependent feature, the periodic feature, the weighted convolution feature, and the linear-dependent feature to obtain an audio feature. The second extraction module is configured to extract, by the initial model, a text feature of text data of each sample user. The second fusion module is configured to fuse, by the initial model, the audio feature and the text feature to obtain a fusion feature. The result determination module is configured to determine, by the initial model, a depression detection result of the sample user according to the fusion feature. The training module is configured to train the initial model according to the depression detection result and a labeled depression result of each sample user to obtain a depression detection model. The first extraction module includes: The extraction unit is configured to extract an acoustic feature of audio data of the sample user. The first processing unit is configured to perform convolution processing on the acoustic feature to obtain a convolution feature. The second processing unit is configured to extract a time-dependent feature of the audio data according to the convolution feature. The third processing unit is configured to extract a periodic feature of the audio data according to the convolution feature. The fourth processing unit is configured to perform time attention weighting processing on a sub-convolution feature of each frame in the convolution feature to obtain a weighted convolution feature. The extraction unit includes: The division sub-unit is configured to divide the audio data of the sample user into at least one frame of sub-audio data. The first determination sub-unit is configured to determine, for each frame of sub-audio data, a mel-frequency cepstral coefficient and a pitch fundamental frequency of the frame of sub-audio data. The second determination sub-unit is configured to determine a mel-spectrum graph of the audio data. The first fusion sub-unit is configured to fuse the mel-frequency cepstral coefficient and the pitch fundamental frequency of each frame of sub-audio data, and the mel-spectrum graph to obtain the acoustic feature of the sample user.

6. An electronic device, comprising: The electronic device includes: at least one processor; and a memory connected with the at least one processor in communication; wherein The memory stores a computer program that can be executed by the at least one processor, and the computer program is executed by the at least one processor to enable the at least one processor to execute the training method of the depression detection model according to any one of claims 1-4.

7. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer instructions for enabling the processor to execute the training method of the depression detection model according to any one of claims 1-4 when executed.

8. A computer program product, characterised in that, The computer program product includes a computer program that, when executed by a processor, implements the training method of the depression detection model according to any one of claims 1-4.

Images