A method, device, equipment and medium for mining potential hit songs

Abstract

The present application discloses a method, device, equipment, and medium for mining potential hit songs, which includes: obtaining first comment data of candidate songs and second comment data of reference hit songs; processing the first comment data and the second comment data respectively, screening out first excellent comment data and second excellent comment data; performing isotropic transformation on the full sample space on the first vector corresponding to the first excellent comment data and the second vector corresponding to the second excellent comment data respectively, obtaining the first processed vector and the second processed vector; using the first processed vector and the second processed vector respectively to determine the candidate song vector of the candidate song and the reference vector of the reference hit song; determining the similarity between the candidate song vector and the reference vector, and determining whether the candidate song is a potential hit song based on the similarity. This can improve the recall rate of potential hit songs.

Description

Technical Field

[0001] The present application relates to the field of music processing technology, and in particular to a method, device, equipment, and medium for mining potential hit songs. Background Art

[0002] In music applications, potential hit songs are usually mined in order to attract more users and improve user stickiness. The current methods for mining potential hit songs mainly use song audio information to mine potential hit songs, and use dark horse algorithms to mine using dark horse users. However, both of the above methods have the problem that many related potential hit songs cannot be detected, that is, the recall rate of potential hit songs is low. Summary of the invention

[0003] In view of this, the purpose of this application is to provide a method, device, equipment, and medium for mining potential hit songs, which can improve the recall rate of potential hit songs. The specific scheme is as follows:

[0004] In the first aspect, the present application discloses a method for mining potential hit songs, comprising:

[0005] Obtain the first comment data of the candidate songs and the second comment data of the reference hit songs;

[0006] Processing the first review data and the second review data respectively to select first excellent review data and second excellent review data;

[0007] Performing isotropic transformation on the entire sample space on the first vector corresponding to the first excellent comment data and the second vector corresponding to the second excellent comment data, respectively, to obtain a first processed vector and a second processed vector;

[0008] Determine a candidate song vector of the candidate song and a reference vector of the reference hit song using the first processed vector and the second processed vector respectively;

[0009] Determine the similarity between the candidate song vector and the reference vector, and determine whether the candidate song is a potential hit song based on the similarity.

[0010] Optionally, the processing the first comment data and the second comment data respectively to select the first excellent comment data and the second excellent comment data includes:

[0011] Inputting the first comment data into a pre-trained BERT model to obtain a first classification result output by the BERT model, and using the high-quality comment data and good comment data in the first classification result as the first excellent comment data;

[0012] The second comment data is input into the BERT model to obtain a second classification result output by the BERT model, and the high-quality comment data and good comment data in the second classification result are used as the second excellent comment data.

[0013] Optionally, performing an isotropic transformation on the entire sample space on the first vector corresponding to the first excellent review data to obtain a first processed vector includes:

[0014] The CLS vector of the first excellent comment data after passing through the last layer of the BERT model is used as the first vector, wherein the first vector includes vectors corresponding to different comments;

[0015] Determine an average vector of the vectors corresponding to the comments in the first vector;

[0016] The covariance matrix is ​​determined using the average vector and a first operation formula, wherein the first operation formula is:

[0017]

[0018] V represents the covariance matrix, N represents the total number of comments corresponding to the first vector, x i represents the vector corresponding to the i-th comment in the first vector, x i is a 1×m-dimensional vector, (·) T represents the transpose operation of the matrix, and u represents the average vector;

[0019] The covariance matrix is ​​subjected to singular value decomposition according to the second operation formula to determine the first matrix and the second matrix, wherein the second operation formula is: V = UQU T ,U represents the first matrix, Q represents the second matrix;

[0020] The transformation matrix is ​​determined using the first matrix, the second matrix and a third operation formula, wherein the third operation formula is: W represents the transformation matrix, (·) -1 Represents the inverse operation of a matrix;

[0021] The vectors corresponding to the comments in the first vector are processed using the transformation matrix to obtain a first processed vector, where: represents the processed vector corresponding to the i-th comment in the first processed vector.

[0022] Optionally, determining a candidate song vector of any candidate song using the first processed vector includes:

[0023] The processed vectors corresponding to the comments corresponding to the candidate song in the first processed vectors are used as target processed vectors;

[0024] The sum of the target processed vectors is used as the candidate song vector of the candidate song.

[0025] Optionally, performing an isotropic transformation on the entire sample space on the first vector corresponding to the first excellent comment data and the second vector corresponding to the second excellent comment data respectively includes:

[0026] The pre-trained deep neural network is used to perform isotropic transformation on the entire sample space for the first vector corresponding to the first excellent comment data and the second vector corresponding to the second excellent comment data.

[0027] Optionally, determining whether the candidate song is a potential hit song according to the similarity includes:

[0028] When the candidate songs include multiple songs, the candidate songs whose similarity exceeds a preset similarity threshold are selected as preliminary songs;

[0029] The song with the greatest similarity among the preliminary selected songs is selected as the potential hit song.

[0030] Optionally, determining the similarity between the candidate song vector and the reference vector comprises:

[0031] determining a cosine distance between the candidate song vector and the reference vector;

[0032] determining a Manhattan distance between the candidate song vector and the reference vector;

[0033] The comprehensive distance between the candidate song vector and the reference vector is determined by using the cosine distance and the Manhattan distance to obtain the similarity between the candidate song vector and the reference vector.

[0034] Optionally, the determining the comprehensive distance between the candidate song vector and the reference vector by using the cosine distance and the Manhattan distance comprises:

[0035] The value of the Manhattan distance raised to the power of the preset parameter is used as the processed Manhattan distance;

[0036] The product of the cosine distance and the processed Manhattan distance is taken as the comprehensive distance between the candidate song vector and the reference vector.

[0037] In a second aspect, the present application discloses a device for mining potential hit songs, comprising:

[0038] A data acquisition module, used to acquire first comment data of candidate songs and second comment data of reference hit songs;

[0039] A data cleaning module, used for processing the first review data and the second review data respectively, and screening out first excellent review data and second excellent review data;

[0040] an isotropic transformation module, used to perform isotropic transformation on the entire sample space on the first vector corresponding to the first excellent comment data and the second vector corresponding to the second excellent comment data, respectively, to obtain a first processed vector and a second processed vector;

[0041] A vector association module, used to determine a candidate song vector of the candidate song and a reference vector of the reference hit song using the first processed vector and the second processed vector respectively;

[0042] The similarity determination module is used to determine the similarity between the candidate song vector and the reference vector, and determine whether the candidate song is a potential hit song based on the similarity.

[0043] In a third aspect, the present application discloses an electronic device, comprising:

[0044] Memory and processor;

[0045] Wherein, the memory is used to store computer programs;

[0046] The processor is used to execute the computer program to implement the aforementioned method for mining potential hit songs.

[0047] In a fourth aspect, the present application discloses a computer-readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements the aforementioned method for mining potential hit songs.

[0048] It can be seen that the present application first obtains the first comment data of the candidate song and the second comment data of the reference hit song. Then the first comment data and the second comment data are processed respectively to screen out the first excellent comment data and the second excellent comment data. Then it is also necessary to perform an isotropic transformation on the full sample space for the first vector corresponding to the first excellent comment data and the second vector corresponding to the second excellent comment data, respectively, to obtain the first processed vector and the second processed vector. Then the candidate song vector of the candidate song and the reference vector of the reference hit song are determined using the first processed vector and the second processed vector respectively. Then the similarity between the candidate song vector and the reference vector is determined, and whether the candidate song is a potential hit song is determined based on the similarity. It can be seen that in the present application, the excellent comment data of the song is used to mine the potential hit songs. Since the comment data can better represent the popularity of the song, the use of the excellent comment data of the song to mine the potential hit songs can improve the recall rate of the potential hit songs. In addition, when using comment data, we first clean the comment data, remove poor-quality comment data, and only use good comment data for subsequent potential hit songs. This can effectively solve the impact of poor-quality comments and further improve the recall rate of potential hit songs. On the other hand, before using the vector corresponding to the comment data to represent the song, we first perform an isotropic transformation on the vector, thereby avoiding the impact of the cone distribution of the original comment vector on the subsequent potential hit song mining and ensuring the recall rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0050] Figure 1 A schematic diagram of the system framework applicable to the potential hit song mining solution provided in this application;

[0051] Figure 2 A flowchart of a specific method for mining potential hit songs disclosed in this application;

[0052] Figure 3 This is a partial flow chart of a specific method for mining potential hit songs disclosed in this application;

[0053] Figure 4 This is a partial flow chart of a specific method for mining potential hit songs disclosed in this application;

[0054] Figure 5 A flowchart of a specific method for mining potential hit songs disclosed in this application;

[0055] Figure 6 This is a partial flow chart of a specific method for mining potential hit songs disclosed in this application;

[0056] Figure 7 A flowchart of a specific method for mining potential hit songs disclosed in this application;

[0057] Figure 8 This is a schematic diagram of the structure of a device for mining potential hit songs disclosed in this application;

[0058] Fig. 9 This is a schematic diagram of the structure of an electronic device disclosed in this application. DETAILED DESCRIPTION

[0059] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0060] At present, the main methods for mining potential hit songs are to mine potential hit songs using song audio information, and to use dark horse algorithms to mine using dark horse users. However, both of the above methods have the problem of low recall rates for potential hit songs. In view of this, the present application proposes a method for mining potential hit songs, which can improve the recall rate of potential hit songs.

[0061] The system framework used in the potential hit song mining solution of this application can be found in Figure 1 As shown, it may specifically include: a backend server and a number of user terminals that establish communication connections with the backend server. The user terminals include but are not limited to tablet computers, laptop computers, smart phones, and personal computers (PCs), which are not limited here.

[0062] In the present application, the backend server executes the steps of the method for mining potential hit songs, including obtaining first comment data of a candidate song and second comment data of a reference hit song; processing the first comment data and the second comment data respectively to screen out first excellent comment data and second excellent comment data; performing isotropic transformation on the entire sample space on the first vector corresponding to the first excellent comment data and the second vector corresponding to the second excellent comment data respectively to obtain a first processed vector and a second processed vector; using the first processed vector and the second processed vector respectively to determine the candidate song vector of the candidate song and the reference vector of the reference hit song; determining the similarity between the candidate song vector and the reference vector, and determining whether the candidate song is a potential hit song based on the similarity.

[0063] After the background server determines the potential hit songs, it can deliver the determined potential hit songs to each user terminal that has established a link with it.

[0064] See also Figure 2 As shown, the embodiment of the present application discloses a method for mining potential hit songs, the method comprising:

[0065] Step S11: Obtain the first comment data of the candidate songs and the second comment data of the reference hit songs.

[0066] In the specific implementation process, it is necessary to first obtain the first comment data of the candidate songs and the second comment data of the reference hit songs, wherein the candidate songs may be new songs recently released by pre-selected singers, and songs manually selected from recently released songs, and the reference hit songs may be hit songs manually selected by editors, or songs on the hot search list, hot song list, etc. The first comment data and the second comment data may be comment data of recent period, for example, comment data of recent month, or comment data of recent year, etc.

[0067] Step S12: Process the first review data and the second review data respectively to select first excellent review data and second excellent review data.

[0068] After obtaining the first comment data and the second comment data, since both the first comment data and the second comment data may contain high-quality, good, ordinary and poor-quality comments, the first comment data and the second comment data need to be cleaned in order to retain the good data in the first comment data and the second comment data and filter out the ordinary and poor-quality comments.

[0069] That is, the first comment data and the second comment data need to be processed respectively, and the first excellent comments are screened out from the first comment data and the second excellent comments are screened out from the second comment data.

[0070] Specifically, the first comment data can be input into a pre-trained BERT (Bidirectional Encoder Representations from Transformer) model to obtain a first classification result output by the BERT model, and the high-quality comment data and good comment data in the first classification result are used as the first excellent comment data; the second comment data can be input into the BERT model to obtain a second classification result output by the BERT model, and the high-quality comment data and good comment data in the second classification result are used as the second excellent comment data.

[0071] That is, the BERT model can be fine-tuned using the labeled comment data to obtain a trained BERT model, and then the first comment data and the second comment data can be processed using the trained BERT model to screen out the first excellent comment data and the second excellent comment data. Specifically, the first comment data is input into the BERT model to obtain a first classification result, so as to classify the first comment data into high-quality, good, ordinary, and low-quality comment data, and then the high-quality comment data and the good comment data in the first classification result are used as the first excellent comment data. The second comment data is input into the BERT model to obtain a second classification result, so as to classify the second comment data into high-quality, good, ordinary, and low-quality comment data, and then the high-quality comment data and the good comment data in the second classification result are used as the second excellent comment data.

[0072] Step S13: performing isotropic transformation on the entire sample space for the first vector corresponding to the first excellent comment data and the second vector corresponding to the second excellent comment data, respectively, to obtain a first processed vector and a second processed vector.

[0073] After filtering out the first excellent comment data and the second excellent comment data, it is also necessary to perform isotropic transformation on the first vector corresponding to the first excellent comment data and the second vector corresponding to the second excellent comment data in the entire sample space to obtain the first processed vector and the second processed vector.

[0074] In the specific implementation process, the pre-trained deep neural network can be used to perform isotropic transformation on the first vector corresponding to the first excellent review data and the second vector corresponding to the second excellent review data in the whole sample space. That is, the deep neural network is first trained with sample data so that the trained deep neural network can perform isotropic transformation on the input vector in the whole sample space and output the transformed vector, wherein the deep neural network can be a VAE (Variational Autoencoder) model or other deep neural networks.

[0075] Here, an isotropic transformation is performed on the first vector and the second vector in the entire sample space to avoid the influence of the conical distribution of the original comment vector on the subsequent calculation using the relevant vector to determine whether the candidate song is a potential hit song.

[0076] Step S14: Determine the candidate song vector of the candidate song and the reference vector of the reference hit song using the first processed vector and the second processed vector respectively.

[0077] After obtaining the first processed vector and the second processed vector, it is also necessary to use the first processed vector and the second processed vector respectively to determine the candidate song vector of the candidate song and the reference vector of the reference hit song.

[0078] In the actual implementation process, the first processed vector includes vectors corresponding to excellent comment data of different candidate songs, so the candidate song vector of any candidate song is determined by using the first processed vector, including: taking the processed vectors corresponding to each comment corresponding to the candidate song in the first processed vector as the target processed vector; and taking the sum of each of the target processed vectors as the candidate song vector of the candidate song.

[0079] Step S15: Determine the similarity between the candidate song vector and the reference vector, and determine whether the candidate song is a potential hit song based on the similarity.

[0080] After determining the candidate song vector of the candidate song and the reference vector of the reference hit song, the similarity between the candidate song vector and the reference vector can be determined, and based on the similarity, it can be determined whether the candidate song is a potential hit song.

[0081] In a specific implementation process, the similarity between the candidate song vector and the reference vector is determined. The similarity may be cosine similarity, inner product, etc., which may be determined based on actual conditions and is not specifically limited here.

[0082] Determining whether the candidate song is a potential hit song based on the similarity includes: when the candidate songs include multiple songs, taking the candidate songs whose similarity exceeds a preset similarity threshold as preliminary selected songs; and taking the song with the greatest similarity among the preliminary selected songs as the potential hit song.

[0083] That is, when the candidate songs include multiple songs, the candidate songs whose similarity exceeds a preset similarity threshold are first selected as preliminary songs, and then the songs with the greatest similarity among the preliminary songs are selected as potential hit songs.

[0084] Of course, when the candidate songs include multiple songs, the candidate songs whose similarity with the reference hit songs exceeds the preset similarity threshold can be directly selected as potential hit songs. Alternatively, the candidate songs whose similarity exceeds the preset similarity threshold are first selected as preliminary songs, and then the first few songs with the greatest similarity among the preliminary songs are selected as potential hit songs. The specific first few songs can be determined according to actual conditions and are not specifically limited here.

[0085] It can be seen that the present application first obtains the first comment data of the candidate song and the second comment data of the reference hit song. Then the first comment data and the second comment data are processed respectively to screen out the first excellent comment data and the second excellent comment data. Then it is also necessary to perform an isotropic transformation on the full sample space for the first vector corresponding to the first excellent comment data and the second vector corresponding to the second excellent comment data, respectively, to obtain the first processed vector and the second processed vector. Then the candidate song vector of the candidate song and the reference vector of the reference hit song are determined using the first processed vector and the second processed vector respectively. Then the similarity between the candidate song vector and the reference vector is determined, and whether the candidate song is a potential hit song is determined based on the similarity. It can be seen that in the present application, the excellent comment data of the song is used to mine the potential hit songs. Since the comment data can better represent the popularity of the song, the use of the excellent comment data of the song to mine the potential hit songs can improve the recall rate of the potential hit songs. In addition, when using comment data, we first clean the comment data, remove poor-quality comment data, and only use good comment data for subsequent potential hit songs. This can effectively solve the impact of poor-quality comments and further improve the recall rate of potential hit songs. On the other hand, before using the vector corresponding to the comment data to represent the song, we first perform an isotropic transformation on the vector, thereby avoiding the impact of the cone distribution of the original comment vector on the subsequent potential hit song mining and ensuring the recall rate.

[0086] See also Figure 3As shown, the first vector corresponding to the first excellent review data is isotropically transformed in the entire sample space to obtain a first processed vector, including:

[0087] Step S131: The CLS vector of the first excellent comment data when it passes through the last layer of the BERT model is used as the first vector, wherein the first vector includes vectors corresponding to different comments.

[0088] First, it is necessary to obtain the first vector corresponding to the first excellent comment data. Specifically, the CLS vector of the first excellent comment when passing through the last layer of the BERT model can be used as the first vector, wherein the first vector includes vectors corresponding to different comments.

[0089] That is, the first vector is a vector set composed of vectors of different comments, including vectors corresponding to excellent comments of each of the candidate songs.

[0090] Step S132: Determine the average vector of the vectors corresponding to the comments in the first vector.

[0091] After obtaining the first vector, it is also necessary to determine the average vector of the vectors corresponding to each comment in the first vector. Specifically, the values ​​of the corresponding dimensions of each comment data are added together, and then divided by the total number of excellent comments in the first excellent comment data.

[0092] Step S133: Determine the covariance matrix using the average vector and a first operation formula, wherein the first operation formula is:

[0093]

[0094] V represents the covariance matrix, N represents the total number of comments corresponding to the first vector, x i represents the vector corresponding to the i-th comment in the first vector, x i is a 1×m-dimensional vector, (·) T represents the transpose operation of the matrix, and u represents the average vector.

[0095] After obtaining the average vector of the first vector, it is also necessary to use the average vector of the first vector and the first calculation formula to determine the covariance matrix. Specifically, Wherein, V represents the covariance matrix, N represents the total number of comments corresponding to the first vector, and x i represents the vector corresponding to the i-th comment in the first vector, x i is a 1×m-dimensional vector, (·) T represents the transpose operation of the matrix, and u represents the average vector.

[0096] x i is a 1×m-dimensional vector, and u is also a 1×m-dimensional vector, so (x i -u) is also a 1×m-dimensional vector, (x i -u) T is an m×1 dimensional vector, so the above covariance matrix V is an m×m matrix.

[0097] Step S134: Perform singular value decomposition on the covariance matrix according to the second operation formula to determine the first matrix and the second matrix, wherein the second operation formula is: V = UQU T ,U represents the first matrix, and Q represents the second matrix.

[0098] After obtaining the covariance matrix, it is also necessary to perform singular value decomposition (SVD) on the covariance matrix according to the second operation formula to determine the first matrix and the second matrix, wherein the second operation formula is: V = UQU T ,U represents the first matrix, Q represents the second matrix, and both U and Q are m×m matrices.

[0099] Step S135: Determine a transformation matrix using the first matrix, the second matrix and a third operation formula, wherein the third operation formula is: W represents the transformation matrix, (·) -1 Represents the matrix inversion operation.

[0100] After obtaining the first matrix and the second matrix, the transformation matrix can be determined using the first matrix, the second matrix and a third operation formula, wherein the third operation formula is: W represents the transformation matrix, (·) -1 Represents the matrix inversion operation. Since U and Q are both m×m matrices, W is also an m×m matrix.

[0101] Step S136: Use the transformation matrix to process the vectors corresponding to each comment in the first vector to obtain a first processed vector, where: represents the processed vector corresponding to the i-th comment in the first processed vector.

[0102] After obtaining the transformation matrix, the vector corresponding to each comment in the first vector can be processed using the transformation matrix to obtain a first processed vector, where: represents the processed vector corresponding to the i-th comment in the first processed vector. i-u) is a 1×m dimensional vector, and W is an m×m matrix, so we get It is also a 1×m-dimensional vector.

[0103] After obtaining the first processed vector, the candidate song vector of the candidate song can be determined using the first processed vector, wherein the method of determining the candidate song vector of any candidate song using the first processed vector includes: taking the processed vector corresponding to each comment corresponding to the candidate song in the first processed vector as the target processed vector; and taking the sum of each of the target processed vectors as the candidate song vector of the candidate song.

[0104] That is, the candidate song vector of any candidate song Where a represents the number of good reviews for the song.

[0105] The specific process of performing isotropic transformation on the second vector in the entire sample space is the same as the above process and will not be described again here.

[0106] See also Figure 4 The figure shows a partial flow chart of the method for mining potential hit songs. After obtaining the original comment set, that is, after obtaining the first excellent comment data and the second excellent comment data, the CLS vector of the first excellent comment data in the BERT model is obtained to obtain the first vector, the CLS vector of the second excellent comment data in the BERT model is obtained to obtain the second vector, and then the first vector and the second vector are isotropically transformed in the entire sample space. The transformed vector representation is output, that is, the first processed vector and the second processed vector, and then the comment vectors associated with the song in the first processed vector and the second processed vector are summed as the song vector.

[0107] See also Figure 5 As shown, the embodiment of the present application discloses a specific method for mining potential hit songs, which includes:

[0108] Step S21: Obtain the first comment data of the candidate songs and the second comment data of the reference hit songs.

[0109] Step S22: Process the first review data and the second review data respectively to select first excellent review data and second excellent review data.

[0110] Step S23: performing isotropic transformation on the entire sample space for the first vector corresponding to the first excellent comment data and the second vector corresponding to the second excellent comment data, respectively, to obtain a first processed vector and a second processed vector.

[0111] Step S24: Determine the candidate song vector of the candidate song and the reference vector of the reference hit song using the first processed vector and the second processed vector respectively.

[0112] The specific implementation process of step S21 and step S24 can refer to the specific content disclosed in the above embodiments, which will not be repeated here.

[0113] Step S25: Determine the cosine distance between the candidate song vector and the reference vector.

[0114] After obtaining the candidate song vector and the reference vector, it is necessary to determine the similarity between the candidate song vector and the reference vector. Specifically, the cosine distance between the candidate song vector and the reference vector may be determined first.

[0115] Step S26: Determine the Manhattan distance between the candidate song vector and the reference vector.

[0116] In order to preserve the differences in the comment data, it is also necessary to determine the Manhattan distance between the candidate song vector and the reference vector. The cosine distance describes the directional consistency, and the Manhattan distance describes the absolute distance in space, corresponding to the difference in the number of comments.

[0117] Step S27: Determine the comprehensive distance between the candidate song vector and the reference vector using the cosine distance and the Manhattan distance to obtain the similarity between the candidate song vector and the reference vector.

[0118] After obtaining the cosine distance and the Manhattan distance, the comprehensive distance can be determined using the cosine distance and the Manhattan distance to obtain the similarity between the candidate song vector and the reference vector.

[0119] Specifically, the value of the preset parameter power of the Manhattan distance is used as the processed Manhattan distance; the product of the cosine distance and the processed Manhattan distance is used as the comprehensive distance between the candidate song vector and the reference vector. That is, s=s 1 *s 2 α , where s represents the comprehensive distance, s 1 Represents the cosine distance, s 2 represents the Manhattan distance, and α represents the preset parameter.

[0120] Step S28: Determine the similarity between the candidate song vector and the reference vector, and determine whether the candidate song is a potential hit song based on the similarity.

[0121] See also Figure 6The figure shows a partial flow chart of the potential hit song mining method. The cosine distance and Manhattan distance between the candidate song vector and the source song vector (reference vector) are calculated, and then the comprehensive distance is calculated based on the cosine distance and Manhattan distance, and the comprehensive distance result is output.

[0122] See also Figure 7 The figure shows a flowchart of the potential hit song mining method. First, the source song pool and related comments are obtained from the database, that is, the hit songs and related comments are referred to, the candidate song pool and related comments are obtained, and high-quality and good comments are screened out respectively. Then, the high-quality and good comments are used to vectorize the songs, and then the similarity between the candidate song vector and the reference vector is calculated, and the candidate song with the highest similarity to the source song is output, and multiple rounds of delivery are carried out on and off the site.

[0123] The aforementioned potential hit song mining method can be integrated into a music APP. The music APP can use the recently added songs in the station as candidate songs at a certain time interval, use the songs in the hot song list as reference hit songs, obtain the comment data of the candidate songs in the station for nearly one month as the first comment data, obtain the comment data of the reference hit songs in the station for nearly one month as the second comment data, and then process the first comment data and the second comment data respectively to screen out the first excellent comment data and the second excellent comment data. Then, the first vector corresponding to the first excellent comment data and the second vector corresponding to the second excellent comment data are respectively subjected to isotropic transformation on the full sample space to obtain the first processed vector and the second processed vector. Then, the first processed vector and the second processed vector are respectively used to determine the candidate song vector of the candidate song and the reference vector of the reference hit song. Then, the similarity between the candidate song vector and the reference vector can be determined, and whether the candidate song is a potential hit song can be determined based on the similarity. After determining the potential hit song, the determined potential hit song is placed on the music APP to attract more users and improve the stickiness of existing users.

[0124] See also Figure 8 As shown, the embodiment of the present application discloses a device for mining potential hit songs, comprising:

[0125] The data acquisition module 11 is used to acquire the first comment data of the candidate songs and the second comment data of the reference hit songs;

[0126] A data cleaning module 12, used for processing the first review data and the second review data respectively, and screening out first excellent review data and second excellent review data;

[0127] An isotropic transformation module 13 is used to perform isotropic transformation on the first vector corresponding to the first excellent comment data and the second vector corresponding to the second excellent comment data in the whole sample space to obtain a first processed vector and a second processed vector;

[0128] A vector association module 14, for determining a candidate song vector of the candidate song and a reference vector of the reference hit song by using the first processed vector and the second processed vector respectively;

[0129] The similarity determination module 15 is used to determine the similarity between the candidate song vector and the reference vector, and determine whether the candidate song is a potential hit song based on the similarity.

[0130] It can be seen that the present application first obtains the first comment data of the candidate song and the second comment data of the reference hit song. Then the first comment data and the second comment data are processed respectively to screen out the first excellent comment data and the second excellent comment data. Then it is also necessary to perform an isotropic transformation on the full sample space for the first vector corresponding to the first excellent comment data and the second vector corresponding to the second excellent comment data, respectively, to obtain the first processed vector and the second processed vector. Then the candidate song vector of the candidate song and the reference vector of the reference hit song are determined using the first processed vector and the second processed vector respectively. Then the similarity between the candidate song vector and the reference vector is determined, and whether the candidate song is a potential hit song is determined based on the similarity. It can be seen that in the present application, the excellent comment data of the song is used to mine the potential hit songs. Since the comment data can better represent the popularity of the song, the use of the excellent comment data of the song to mine the potential hit songs can improve the recall rate of the potential hit songs. In addition, when using comment data, we first clean the comment data, remove poor-quality comment data, and only use good comment data for subsequent potential hit songs. This can effectively solve the impact of poor-quality comments and further improve the recall rate of potential hit songs. On the other hand, before using the vector corresponding to the comment data to represent the song, we first perform an isotropic transformation on the vector, thereby avoiding the impact of the cone distribution of the original comment vector on the subsequent potential hit song mining and ensuring the recall rate.

[0131] In a specific implementation process, the data cleaning module 12 is used to:

[0132] Inputting the first comment data into a pre-trained BERT model to obtain a first classification result output by the BERT model, and using the high-quality comment data and good comment data in the first classification result as the first excellent comment data;

[0133] The second comment data is input into the BERT model to obtain a second classification result output by the BERT model, and the high-quality comment data and good comment data in the second classification result are used as the second excellent comment data.

[0134] In a specific implementation process, the isotropic transformation module 13 is used to:

[0135] The CLS vector of the first excellent comment data after passing through the last layer of the BERT model is used as the first vector, wherein the first vector includes vectors corresponding to different comments;

[0136] Determine an average vector of the vectors corresponding to the comments in the first vector;

[0137] The covariance matrix is ​​determined using the average vector and a first operation formula, wherein the first operation formula is:

[0138]

[0139] V represents the covariance matrix, N represents the total number of comments corresponding to the first vector, x i represents the vector corresponding to the i-th comment in the first vector, x i is a 1×m-dimensional vector, (·) T represents the transpose operation of the matrix, and u represents the average vector;

[0140] The covariance matrix is ​​subjected to singular value decomposition according to the second operation formula to determine the first matrix and the second matrix, wherein the second operation formula is: V = UQU T ,U represents the first matrix, Q represents the second matrix;

[0141] The transformation matrix is ​​determined using the first matrix, the second matrix and a third operation formula, wherein the third operation formula is: W represents the transformation matrix, (·) -1 Represents the inverse operation of a matrix;

[0142] The vectors corresponding to the comments in the first vector are processed using the transformation matrix to obtain a first processed vector, where: represents the processed vector corresponding to the i-th comment in the first processed vector.

[0143] In a specific implementation process, the vector association module 14 is used to:

[0144] The processed vectors corresponding to the comments corresponding to the candidate song in the first processed vectors are used as target processed vectors;

[0145] The sum of the target processed vectors is used as the candidate song vector of the candidate song.

[0146] In a specific implementation process, the isotropic transformation module 13 is used to:

[0147] The pre-trained deep neural network is used to perform isotropic transformation on the entire sample space for the first vector corresponding to the first excellent comment data and the second vector corresponding to the second excellent comment data.

[0148] In a specific implementation process, the similarity determination module 15 is used to:

[0149] When the candidate songs include multiple songs, the candidate songs whose similarity exceeds a preset similarity threshold are selected as preliminary songs;

[0150] The song with the greatest similarity among the preliminary selected songs is selected as the potential hit song.

[0151] In a specific implementation process, the similarity determination module 15 is used to:

[0152] determining a cosine distance between the candidate song vector and the reference vector;

[0153] determining a Manhattan distance between the candidate song vector and the reference vector;

[0154] The comprehensive distance between the candidate song vector and the reference vector is determined by using the cosine distance and the Manhattan distance to obtain the similarity between the candidate song vector and the reference vector.

[0155] In a specific implementation process, the similarity determination module 15 is used to:

[0156] The value of the Manhattan distance raised to the power of the preset parameter is used as the processed Manhattan distance;

[0157] The product of the cosine distance and the processed Manhattan distance is taken as the comprehensive distance between the candidate song vector and the reference vector.

[0158] Fig. 9 This is a schematic diagram of the structure of an electronic device 20 provided in an embodiment of the present application. The electronic device 20 may specifically include but is not limited to a tablet computer, a laptop computer, or a desktop computer.

[0159] Generally, the electronic device 20 in this embodiment includes: a processor 21 and a memory 22 .

[0160] Among them, the processor 21 may include one or more processing cores, such as a quad-core processor, an octa-core processor, etc. The processor 21 may be implemented using at least one of DSP (digital signal processing), FPGA (field-programmable gate array), and PLA (programmable logic array). The processor 21 may also include a main processor and a coprocessor. The main processor is a processor for processing data in an awake state, also known as a CPU (central processing unit); the coprocessor is a low-power processor for processing data in a standby state. In some embodiments, the processor 21 may be integrated with a GPU (graphics processing unit), which is responsible for rendering and drawing images that need to be displayed on the display screen. In some embodiments, the processor 21 may include an AI (artificial intelligence) processor, which is used to process computing operations related to machine learning.

[0161] The memory 22 may include one or more computer-readable storage media, and the computer-readable storage media may be non-transitory. The memory 22 may also include a high-speed random access memory, and a non-volatile memory, such as one or more disk storage devices, flash memory storage devices. In this embodiment, the memory 22 is at least used to store the following computer program 221, wherein after the computer program is loaded and executed by the processor 21, the steps of the potential hit song mining method disclosed in any of the aforementioned embodiments can be implemented.

[0162] In some embodiments, the electronic device 20 may further include a display screen 23 , an input / output interface 24 , a communication interface 25 , a sensor 26 , a power source 27 , and a communication bus 28 .

[0163] Those skilled in the art will understand that Fig. 9 The structure shown in the figure does not constitute a limitation on the electronic device 20, and may include more or fewer components than shown in the figure.

[0164] Furthermore, an embodiment of the present application also discloses a computer-readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements the method for mining potential hit songs disclosed in any of the aforementioned embodiments.

[0165] Among them, the specific process of the above-mentioned potential hit song mining method can refer to the corresponding content disclosed in the aforementioned embodiment, which will not be repeated here.

[0166] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0167] The steps of the method or algorithm described in conjunction with the embodiments disclosed herein may be implemented directly using hardware, a software module executed by a processor, or a combination of the two. The software module may be placed in a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

[0168] Finally, it should be noted that, in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a series of processes, methods, articles or equipment containing other elements include not only those elements, but also other elements not explicitly listed, or also include elements inherent to such processes, methods, articles or equipment. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or equipment that includes the elements.

[0169] The above is a detailed introduction to a method, device, equipment, and medium for mining potential hit songs provided by the present application. This article uses specific examples to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea; at the same time, for general technical personnel in this field, according to the ideas of the present application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.

Claims

1. A method for mining potential hit songs, characterized in that: include: Obtain the first comment data of the candidate songs and the second comment data of the reference hit songs; Processing the first review data and the second review data respectively to select first excellent review data and second excellent review data; Performing isotropic transformation on the entire sample space on the first vector corresponding to the first excellent comment data and the second vector corresponding to the second excellent comment data, respectively, to obtain a first processed vector and a second processed vector; Determine a candidate song vector of the candidate song and a reference vector of the reference hit song using the first processed vector and the second processed vector respectively; Determine the similarity between the candidate song vector and the reference vector, and determine whether the candidate song is a potential hit song based on the similarity; The first vector corresponding to the first excellent review data is subjected to an isotropic transformation in the entire sample space to obtain a first processed vector, including: The CLS vector of the last layer of the pre-trained BERT model of the first excellent comment data is used as the first vector; the first vector includes vectors corresponding to different comments; Determine an average vector of the vectors corresponding to the comments in the first vector; Determine a covariance matrix using the mean vector and the first operation formula, perform singular value decomposition on the covariance matrix according to the second operation formula to determine a first matrix and a second matrix, and determine a transformation matrix using the first matrix, the second matrix and a third operation formula; The vector corresponding to each comment in the first vector is processed using the transformation matrix to obtain a first processed vector.

2. The method for mining potential hit songs according to claim 1, characterized in that: The processing of the first review data and the second review data respectively to select the first excellent review data and the second excellent review data comprises: Inputting the first comment data into a pre-trained BERT model to obtain a first classification result output by the BERT model, and using the high-quality comment data and good comment data in the first classification result as the first excellent comment data; The second comment data is input into the BERT model to obtain a second classification result output by the BERT model, and the high-quality comment data and good comment data in the second classification result are used as the second excellent comment data.

3. The method for mining potential hit songs according to claim 2, characterized in that: The first operation formula is: V represents the covariance matrix, N represents the total number of comments corresponding to the first vector, x i represents the vector corresponding to the i-th comment in the first vector, x i is a 1×m-dimensional vector, () T represents the transpose operation of the matrix, and u represents the average vector; The second operation formula is: V = UQU T ,U represents the first matrix, Q represents the second matrix; The third operation formula is: W represents the transformation matrix, ( · ) - 1 Represents the inverse operation of a matrix; The calculation formula of the first processed vector is: represents the processed vector corresponding to the i-th comment in the first processed vector.

4. The method for mining potential hit songs according to claim 3, characterized in that: Determining a candidate song vector of any candidate song using the first processed vector includes: The processed vectors corresponding to the comments corresponding to the candidate song in the first processed vectors are used as target processed vectors; The sum of the target processed vectors is used as the candidate song vector of the candidate song.

5. The method for mining potential hit songs according to claim 1, characterized in that: The performing isotropic transformation on the entire sample space of the first vector corresponding to the first excellent comment data and the second vector corresponding to the second excellent comment data respectively includes: The pre-trained deep neural network is used to perform isotropic transformation on the entire sample space for the first vector corresponding to the first excellent comment data and the second vector corresponding to the second excellent comment data.

6. The method for mining potential hit songs according to claim 1, characterized in that: Determining whether the candidate song is a potential hit song according to the similarity includes: When the candidate songs include multiple songs, the candidate songs whose similarity exceeds a preset similarity threshold are selected as preliminary songs; The song with the greatest similarity among the preliminary selected songs is selected as the potential hit song.

7. The method for mining potential hit songs according to any one of claims 1 to 6, characterized in that: The determining the similarity between the candidate song vector and the reference vector comprises: determining a cosine distance between the candidate song vector and the reference vector; determining a Manhattan distance between the candidate song vector and the reference vector; The comprehensive distance between the candidate song vector and the reference vector is determined by using the cosine distance and the Manhattan distance to obtain the similarity between the candidate song vector and the reference vector.

8. The method for mining potential hit songs according to claim 7, characterized in that: The method of determining the comprehensive distance between the candidate song vector and the reference vector by using the cosine distance and the Manhattan distance comprises: The value of the Manhattan distance raised to the power of the preset parameter is used as the processed Manhattan distance; The product of the cosine distance and the processed Manhattan distance is taken as the comprehensive distance between the candidate song vector and the reference vector.

9. An electronic device, characterized in that: include: Memory and processor; Wherein, the memory is used to store computer programs; The processor is used to execute the computer program to implement the method for mining potential hit songs as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that: Used to store a computer program, wherein when the computer program is executed by a processor, the method for mining potential hit songs as described in any one of claims 1 to 8 is implemented.

Images