Song content risk assessment method and song risk comparison model training method

Abstract

The invention relates to a song content risk assessment method and a song risk comparison model training method. The method comprises the following steps: acquiring song features of a plurality of hot songs in a song library, wherein the song features comprise at least one of a song content feature, a song creator feature and a song interaction feature besides a security algorithm feature; according to a preset comparison sorting method, determining two hot songs participating in song risk comparison from the multiple hot songs for multiple times, and inputting song features of the two hot songs determined each time into a pre-trained song risk comparison model to obtain a risk comparison result of the two hot songs determined each time; according to the risk comparison result of the two hot songs determined each time, the multiple hot songs are sorted according to the risk level through a comparison sorting method; and dividing the plurality of ordered hot songs into a plurality of intervals, and setting risk coefficients of the hot songs in each interval. By adopting the method, the song risk assessment effect can be effectively improved.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a song content risk assessment method, a song risk comparison model training method, a computer device, a computer-readable storage medium, and a computer program product. Background Art

[0002] With regard to the music platform, which has accumulated a large number of songs, there are many songs with illegal content due to the complex content and uneven quality. Therefore, it is necessary to review the song content to assess the security risks of the song content.

[0003] In the related technology, the traditional method of song content review is to break down songs into modal content such as audio, text and pictures, and use a separate review model corresponding to each modal content to filter and review. The review standard is single, and for unpopular songs in the music library, due to the lack of modal content data of unpopular songs, it will cause modeling difficulties and it is difficult to accurately assess the song content risks of unpopular songs. Summary of the invention

[0004] Based on this, it is necessary to provide a song content risk assessment method, a song risk comparison model training method, a computer device, a computer-readable storage medium and a computer program product that can improve the effect of song content risk assessment in response to the above-mentioned technical problems.

[0005] In a first aspect, the present application provides a song content risk assessment method. The method comprises:

[0006] Acquire song features of multiple popular songs in the music library, wherein the song features include at least one of song content features, song creator features, and song interaction features in addition to security algorithm features;

[0007] According to a preset comparison and sorting method, two popular songs that participate in the risk comparison of songs are determined from the multiple popular songs for multiple times, and the song features of the two popular songs determined each time are input into the pre-trained song risk comparison model to obtain the risk comparison result of the two popular songs determined each time; the comparison and sorting method sorts the multiple popular songs according to the risk comparison result of the two popular songs determined each time;

[0008] The sorted plurality of popular songs are divided into a plurality of intervals, and risk coefficients of the popular songs in each interval are set, wherein the risk coefficients of the popular songs in the same interval are the same.

[0009] In one embodiment, dividing the sorted plurality of popular songs into a plurality of intervals includes:

[0010] According to the exponential distribution division rule, the sorted plurality of popular songs are divided into the plurality of intervals.

[0011] In one embodiment, after the step of setting the risk factor of the popular songs in each interval, the method further includes:

[0012] Inputting at least one song feature of unpopular songs in the music library into a feature vector learning network to obtain a risk feature vector of the unpopular songs, and inputting at least one song feature of each of the popular songs into the feature vector learning network to obtain a risk feature vector of each of the popular songs; wherein the types of song features of the unpopular songs are the same as the types of song features of the popular songs, and include at least one of a security algorithm feature, a song content feature, a song creator feature, and a song interaction feature;

[0013] Determining at least one target popular song from among the popular songs according to the similarity between the risk feature vectors of the unpopular songs and the risk feature vectors of the popular songs;

[0014] The risk coefficient of the unpopular song is obtained based on the risk coefficient of the at least one target popular song.

[0015] In one embodiment, the method further comprises:

[0016] Obtaining a pre-trained deep neural network for determining the song category of a song; wherein the deep neural network sequentially comprises an input layer, a plurality of feature vector extraction layers, and an output layer, and after at least one song feature of the popular songs in the same interval is input into the input layer, the song category of each of the popular songs output by the output layer is the same;

[0017] Using the input layer and the plurality of feature vector extraction layers as the feature vector learning network;

[0018] The step of inputting at least one song feature of the unpopular songs in the music library into a feature vector learning network to obtain a risk feature vector of the unpopular songs, and inputting at least one song feature of each of the popular songs into the feature vector learning network to obtain a risk feature vector of each of the popular songs comprises:

[0019] Inputting at least one song feature of the unpopular song into the input layer of the feature vector learning network, and using the feature vector output by the last feature vector extraction layer among the multiple feature vector extraction layers as the risk feature vector of the unpopular song;

[0020] At least one song feature of each of the popular songs is input into the input layer of the feature vector learning network, and the feature vector output by the last feature vector extraction layer among the multiple feature vector extraction layers is used as the risk feature vector of each of the popular songs.

[0021] In one embodiment, obtaining the risk factor of the unpopular song based on the risk factor of the at least one target popular song includes:

[0022] The risk coefficient of the at least one target popular song is processed by weighted average to obtain the risk coefficient of the unpopular song.

[0023] In one embodiment, the method further comprises:

[0024] The risk coefficient of each popular song and the risk coefficient of each unpopular song in the music library are weighted averaged to obtain the risk coefficient of the music library.

[0025] In one embodiment, the method further comprises:

[0026] When a need to improve the health of the music library is detected, the risk coefficient threshold of the music library is lowered; when a need to improve the richness of the music library is detected, the risk coefficient threshold of the music library is increased; wherein the risk coefficient threshold of the music library is determined based on the risk coefficient of each song in the music library, each of which includes the popular songs and the unpopular songs;

[0027] Songs in the music library whose risk coefficients do not exceed the adjusted risk coefficient threshold of the music library are determined as songs to be put on the shelves.

[0028] In one embodiment, the method further comprises:

[0029] For songs of a preset type in the music library, determining whether the songs of the preset type are songs on the shelves based on a specific risk factor threshold set for the songs of the preset type;

[0030] For other songs in the music library except the preset type of songs, it is determined whether the other songs are listed songs based on the risk coefficient threshold of the music library.

[0031] In a second aspect, the present application provides a method for training a song risk comparison model. The method comprises:

[0032] Obtain any two popular songs in the music library and construct multiple sample pairs; the two popular songs in the sample pairs are different songs;

[0033] Based on the song risk comparison labels of each sample pair and the song features of the popular songs in each sample pair, a song risk comparison model is trained; the song risk comparison labels are used to characterize the content risks of the two popular songs in the corresponding sample pair; the song features include at least one of the song content features, song creator features, and song interaction features in addition to the security algorithm features;

[0034] The song risk comparison model is used in any of the song content risk assessment methods described.

[0035] In a third aspect, the present application further provides a computer device, wherein the computer device comprises a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the song content risk assessment method as described in the first aspect are implemented.

[0036] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the song content risk assessment method described in the first aspect are implemented.

[0037] In a fifth aspect, the present application further provides a computer program product, which includes a computer program, and when the computer program is executed by a processor, the steps of the song content risk assessment method described in the first aspect are implemented.

[0038] The above-mentioned song content risk assessment method, song risk comparison model training method, computer device, computer-readable storage medium and computer program product obtain song features of multiple popular songs in a music library, wherein the song features include at least one of song content features, song creator features and song interaction features in addition to security algorithm features, and then sort the multiple popular songs according to risk levels according to a preset comparison and sorting method, and then divide the sorted multiple popular songs into multiple intervals, and set risk coefficients for the popular songs in each interval, wherein the risk coefficients of popular songs in the same interval are the same, thereby optimizing the assessment of song content risks, using security algorithm features combined with song content features, song creator features and song interaction features to compare the song content risks between popular songs, and being able to perform comprehensive risk assessment on song content, dividing the intervals according to the sorted popular songs and setting the risk coefficients of the popular songs in each interval, and being able to make overall considerations based on the risk ranking of songs in the music library, and effectively improving the content risk assessment effect of the songs. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] FIG1 is a schematic diagram of a flow chart of a song content risk assessment method in one embodiment;

[0040] FIG2 is a schematic diagram of a sorting process in one embodiment;

[0041] FIG3 is a schematic diagram of searching for songs based on the similarity of risk feature vectors of songs in one embodiment;

[0042] FIG4 is a schematic diagram of learning a song risk feature vector in one embodiment;

[0043] FIG5 is a schematic flow chart of a method for training a song risk comparison model in one embodiment;

[0044] FIG6 is a flow chart of a song content risk assessment method in another embodiment;

[0045] Figure 7 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0046] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0047] In an exemplary embodiment, as shown in FIG1 , a song content risk assessment method is provided. This embodiment uses the method applied to a terminal as an example. It is understandable that the method can also be applied to a server, and can also be applied to a system including a terminal and a server, and is implemented through the interaction between the terminal and the server. The server can be implemented as an independent server or a server cluster composed of multiple servers; the terminal can be, but is not limited to, various personal computers, laptops, smart phones, tablet computers, portable wearable devices, etc.

[0048] In this embodiment, the method includes the following steps S101 to S103. Among them:

[0049] In step S101, song features of multiple popular songs in the music library are obtained, and the song features include at least one of song content features, song creator features, and song interaction features in addition to security algorithm features.

[0050] As an example, the hot songs and unpopular songs in the music library can be defined in combination with the completeness of the song features and the user interaction of the songs. Among them, hot songs can be songs with complete song features in the music library, such as complete song content (such as lyrics) and creator information, and relatively more user interactions (such as playing, commenting). Among them, songs whose user interaction times exceed a preset threshold can be determined as songs with relatively more user interactions. For example, a preset number or a preset proportion of songs ranked at the top of the user interaction times in the music library can be considered as songs with relatively more user interactions. Of course, whether the number of user interactions is relatively large can also be determined by other methods, and this application does not make specific limitations. Unpopular songs can be songs that lack song features, such as songs with incomplete song content and creator information, and / or relatively few user interactions.

[0051] As an example, security algorithm features can be used to characterize whether the content of a song is legal and compliant, such as whether it contains lyrics that are illegal or in violation of regulations; song content features can include objective features of the song content (such as language, genre, instrument, release time, etc.) and subjective features of the song content (such as emotion, environment, listening experience, etc.); song creator features can include composer / lyricist portrait features, such as the historical song violation rate corresponding to the composer / lyricist, whether the composer / lyricist has serious violations, and the risk situation of the composer / lyricist; the song interaction features can include user interaction behavior features, such as song play volume, song completion rate, song comment volume, song collection volume, comment violation rate, and whether the comments have serious violations.

[0052] In actual applications, for each popular song in the music library, the security algorithm features of each popular song, as well as at least one of the song content features, song creator features, and song interaction features can be obtained to obtain the song features of each popular song.

[0053] In step S102, according to a preset comparison and sorting method, two popular songs that participate in the risk comparison of songs are determined from multiple popular songs for multiple times, and the song features of the two popular songs determined each time are input into a pre-trained song risk comparison model to obtain the risk comparison result of the two popular songs determined each time; the comparison and sorting method sorts the multiple popular songs according to the risk comparison result of the two popular songs determined each time.

[0054] As an example, the comparison sorting method may include various comparison sorting methods such as bubble sort, quick sort, insertion sort, merge sort, etc.

[0055] Among them, the song risk comparison model can be used to compare the song content risk between two popular songs based on their song features. By selecting security algorithm features and combining them with song content features, song creator features, and song interaction features, the song risk comparison model is modeled, so that the model can be trained based on comprehensive features, making the song risk comparison model have stronger generalization capabilities and stronger recognition capabilities for new types of illegal song content.

[0056] In a specific implementation, for each time two popular songs to be compared are determined, the song features of the two popular songs can be input into a song risk comparison model. The song risk comparison model can be used to compare the song content risks of the two popular songs to obtain a risk comparison result. For example, in the song pair <Popular Song 1, Popular Song 2>, the song content risk of Popular Song 1 is higher than the song content risk of Popular Song 2, or the song content risk of Popular Song 2 is higher than the song content risk of Popular Song 1.

[0057] In one example, a comparative sorting method may be used to sort multiple popular songs in the music library according to the risk level based on the risk comparison results of two popular songs determined each time.

[0058] For example, taking bubble sort as an example, as shown in Figure 2, sorting can be performed based on the idea of ​​bubble sort. For example, if the music library contains 6 popular songs, the first two songs at the end of the sort <Song 5, Song 2> can be selected for judgment. Through the risk comparison result output by the song risk comparison model, it is determined that the risk of Song 5 is higher than that of Song 2 (that is, the popular songs in adjacent sorting positions). The positions of the two songs can be swapped, and then by comparing the risks of <Song 5, Song 3>, the risk of Song 5 is still higher, so the positions of the two songs can be swapped. By comparing the risks of <Song 5, Song 6>, the risk of Song 5 is lower than that of Song 6, and the positions of the two songs remain unchanged.

[0059] Then, for the next song 6 to be judged, we continue to compare the risks of <Song 6, Song 1>. The risk of Song 6 is higher, so the two songs are swapped. We continue to compare the risks of <Song 6, Song 4>. The risk of Song 6 is higher, so the two songs are swapped. Until the first round of bubbling is completed, the result is as shown in Figure 2, where Song 2 is at the bottom of the water and Song 6 is above the water.

[0060] When the music library contains multiple popular songs, a second round of bubbling processing can be performed after the first round of bubbling is completed, and multiple processing rounds are performed until all songs are sorted according to the risk level of the song content.

[0061] In step S103, the sorted multiple popular songs are divided into multiple intervals, and the risk coefficients of the popular songs in each interval are set, wherein the risk coefficients of the popular songs in the same interval are the same.

[0062] As an example, the risk coefficients set for different intervals may be in the range of 0-1, such as 0.1, 0.15, 0.16, 0.5, . . .

[0063] After sorting the multiple popular songs in the music library according to the risk level, the sorted multiple popular songs can be divided into multiple intervals according to the exponential distribution division rule (or the uniform interval division method), and then the risk coefficient of the popular songs in each interval can be set, so that the risk coefficients of the popular songs in the same interval are the same. Therefore, based on quantifying the risk level of the popular songs in the music library, risk intervals with different risk coefficients can be divided.

[0064] Compared with the traditional method, the song risk assessment rules are single and rigid and cannot adapt to various changes; the technical solution of this embodiment, for popular songs in the music library, because their content information and feature data are complete, can use security algorithm features combined with song content features, song creator features, and song interaction features to quantitatively assess the risk level of song content, which can improve the risk assessment effect of songs and effectively identify new types of illegal content in songs.

[0065] In the above-mentioned song content risk assessment method, by obtaining song features of multiple popular songs in the music library, and then according to a preset comparison and sorting method, two popular songs that participate in the comparison of song risks are determined from multiple popular songs for multiple times, and the song features of the two popular songs determined each time are input into the pre-trained song risk comparison model to obtain the risk comparison result of the two popular songs determined each time. The comparison and sorting method sorts the multiple popular songs according to the risk comparison result of the two popular songs determined each time, and then divides the sorted multiple popular songs into multiple intervals, and sets the risk coefficient of the popular songs in each interval, thereby realizing the optimization of the assessment of song content risk. The security algorithm features are combined with the song content features, the song creator features, and the song interaction features to compare the song content risks between popular songs, so as to make a comprehensive risk judgment on the song content, divide the intervals according to the sorted popular songs and set the risk coefficient of the popular songs in each interval, and can be considered as a whole based on the risk ranking of the songs in the music library, thereby effectively improving the content risk assessment effect of the song.

[0066] In an exemplary embodiment, dividing the sorted multiple popular songs into multiple intervals includes the following steps: dividing the sorted multiple popular songs into multiple intervals according to an exponential distribution partitioning rule.

[0067] In one example, as a probability distribution, the probability density function of the exponential distribution decays exponentially. The weighted sum of the sorted popular songs according to the risk level can be performed to obtain the risk index of each popular song. The risk index can then be normalized so that its value range is between 0 and 1. The number of intervals to be divided can then be determined according to actual needs. Multiple intervals can be obtained by combining the number of intervals and the characteristics of the exponential distribution. For example, based on the characteristics of the exponential distribution, a series of decreasing interval thresholds can be set. The threshold can be determined based on the normalization of the risk index to ensure that the risk level of songs in different intervals decays exponentially. This can avoid songs with similar risks being evaluated differently. By dividing the risk intervals based on the exponential distribution, it is beneficial to more finely control the overall risk health of the music library.

[0068] In an exemplary embodiment, after the step of setting the risk coefficient of the popular songs in each interval, the following steps may also be included: inputting at least one song feature of the unpopular songs in the music library into the feature vector learning network to obtain the risk feature vector of the unpopular songs, and inputting at least one song feature of each popular song into the feature vector learning network to obtain the risk feature vector of each popular song; wherein the types of song features of the unpopular songs are the same as the types of song features of the popular songs, and include at least one of security algorithm features, song content features, song creator features, and song interaction features; according to the similarity between the risk feature vector of the unpopular songs and the risk feature vector of each popular song, determining at least one target popular song in each popular song; and obtaining the risk coefficient of the unpopular songs based on the risk coefficient of at least one target popular song.

[0069] As an example, unpopular songs may be songs that lack song features, such as incomplete song content and creator information, and / or songs with relatively few user interactions. The determination method for relatively few user interactions can refer to the determination method for relatively many user interactions described above, and will not be repeated here.

[0070] In one example, for unpopular songs in a music library, risk assessment of unpopular songs is more difficult due to missing information and features. Based on a feature vector learning network (such as a deep neural network), at least one song feature of unpopular songs in the music library can be input into the feature vector learning network to obtain a risk feature vector of the unpopular songs, and at least one song feature of each popular song can be input into the feature vector learning network to obtain a risk feature vector of each popular song, wherein the types of song features of unpopular songs input into the feature vector learning network are the same as those of popular songs, and include at least one of security algorithm features, song content features, song creator features, and song interaction features; based on the similarity of the risk feature vectors, at least one target popular song can be determined from the popular songs in the music library, and then the risk coefficient of the unpopular song can be obtained based on the risk coefficient of at least one target popular song.

[0071] Optionally, a song heterogeneous graph can be established based on information such as song content labels and knowledge graphs, and then the risk feature vector of the song can be learned through graph representation learning methods. For example, by collecting song content labels including song genres, styles, emotions, themes, etc., and using entities such as songs, singers, albums, genres, and their relationships to build a knowledge graph, a song heterogeneous graph can be established based on song content labels and knowledge graph information. The song heterogeneous graph not only contains direct relationships between songs, but also indirect relationships based on content labels. Then, a graph representation learning model suitable for processing heterogeneous graphs can be used for model training, so that after the training is completed, the risk feature vector of the song can be extracted from the nodes (representing the song) of the song heterogeneous graph.

[0072] In another example, the risk coefficient of the unpopular song to be identified can be determined based on the risk coefficient of the similar popular songs (i.e., the target popular songs) found out; illustratively, when there is one target popular song, the risk coefficient of the target popular song can be used as the risk coefficient of the unpopular song; when there are multiple target popular songs, the risk coefficients of the multiple target popular songs can be comprehensively processed (such as weighted average) to obtain the risk coefficient of the unpopular song.

[0073] Specifically, since it is difficult to obtain complete feature information for unpopular songs, the effect is not good if the song risk comparison model is used. By learning the risk feature vector, we can model the song similarity based on the song similarity, find songs similar to unpopular songs among the popular songs with determined risk coefficients, and then predict the risk of unpopular songs based on the risk of similar popular songs.

[0074] For example, as shown in Figure 3, the large dots in Figure 3 represent popular songs, and the risk coefficient of each popular song has been determined; the small dots in Figure 3 represent unpopular songs. Through the similarity of the song risk feature vectors, the K popular songs (i.e., at least one target popular song) with the most similar risks to the unpopular songs can be found. In this way, the risk of unpopular songs can be evaluated based on the risk level of popular songs with determined risk coefficients.

[0075] Compared with the traditional method, it is impossible to accurately evaluate the risk coefficient of unpopular songs; the technical solution of this embodiment can find popular songs with similar risks based on the risk feature vector for unpopular songs, and use the risk coefficients of popular songs with similar risks to evaluate the risk level of unpopular songs, thereby accurately evaluating the song content risk of unpopular songs.

[0076] In an exemplary embodiment, the following steps may also be included: obtaining a pre-trained deep neural network for determining the song category of a song; wherein the deep neural network sequentially includes an input layer, multiple feature vector extraction layers, and an output layer, and after at least one song feature of popular songs in the same interval is input into the input layer, the song category of each popular song output by the output layer is the same; the input layer and the multiple feature vector extraction layers are used as a feature vector learning network.

[0077] Inputting at least one song feature of unpopular songs in a music library into a feature vector learning network to obtain a risk feature vector for the unpopular songs, and inputting at least one song feature of each popular song into the feature vector learning network to obtain a risk feature vector for each popular song can include the following steps: inputting at least one song feature of unpopular songs into an input layer of the feature vector learning network, and using a feature vector output by a feature vector extraction layer of the last layer among multiple feature vector extraction layers as a risk feature vector for the unpopular songs; inputting at least one song feature of each popular song into an input layer of the feature vector learning network, and using a feature vector output by a feature vector extraction layer of the last layer among multiple feature vector extraction layers as a risk feature vector for each popular song.

[0078] In practical applications, as shown in FIG4, in order to find similar popular songs to measure the risk of unpopular songs, a pre-trained deep neural network for determining the song category of a song can be used to learn a risk feature vector based on the song features contained in the popular songs. By screening various song features that can also be obtained from unpopular songs from the song features of popular songs, after at least one song feature of popular songs in the same interval is input into the input layer of the deep neural network, the output layer of the deep neural network can output the same song category of each popular song, and then use the deep neural network for feature embedding learning. For example, at least one song feature of the selected popular songs can be input into the neural network, such as the node of the input layer in FIG4, and the output layer node of the neural network in FIG4 can represent the song category to which it belongs.

[0079] After the network is trained, the input layer and multiple feature vector extraction layers can be used as feature vector learning networks. By inputting at least one song feature of unpopular songs into the input layer of the feature vector learning network, the feature vector output by the last feature vector extraction layer in the multiple feature vector extraction layers can be used as the risk feature vector of the unpopular songs. At least one song feature of each popular song can be input into the input layer of the feature vector learning network, and the feature vector output by the last feature vector extraction layer in the multiple feature vector extraction layers can be used as the risk feature vector of each popular song. Thus, by learning the risk feature vectors of songs, data support is provided for searching songs based on the similarity of risk feature vectors.

[0080] In an exemplary embodiment, obtaining the risk coefficient of an unpopular song based on the risk coefficient of at least one target popular song may include the following steps: performing weighted averaging processing on the risk coefficient of at least one target popular song to obtain the risk coefficient of the unpopular song.

[0081] In a specific implementation, the risk coefficient of at least one target popular song can be processed by weighted average and used as the risk coefficient of the unpopular song, so as to achieve quantitative risk assessment of the unpopular song.

[0082] In an exemplary embodiment, the following steps may also be included:

[0083] The risk coefficient of each popular song and each unpopular song in the music library is weighted averaged to obtain the risk coefficient of the music library.

[0084] In one example, after obtaining the risk coefficient of each song (hot song, unpopular song) in the music library, a weighted average is performed based on the risk coefficient of each hot song and the risk coefficient of each unpopular song to obtain the risk coefficient of the music library. This can quantitatively measure the risk of each song, comprehensively calculate the risk distribution of the music library, and help comprehensively measure the risk coefficient of the entire music library.

[0085] In an exemplary embodiment, the following steps may also be included: when a need to improve the health of the music library is detected, the risk coefficient threshold of the music library is lowered; when a need to improve the richness of the music library is detected, the risk coefficient threshold of the music library is increased; wherein the risk coefficient threshold of the music library is determined based on the risk coefficient of each song in the music library, each song including popular songs and unpopular songs; and songs in the music library whose risk coefficient does not exceed the adjusted risk coefficient threshold of the music library are determined as songs to be put on the shelves.

[0086] In actual applications, a global risk factor threshold can be set for the entire music library, and then the quantified risk factor threshold can be used to dynamically adjust the song listing and delisting strategies. When the need to improve the health of the music library is detected, such as the overall risk factor of the music library has an increasing trend, the requirements for listing the music library become stricter, etc., the risk factor threshold can be lowered to select songs for listing using the lowered risk factor threshold; when the need to increase the richness of the music library is detected, the risk factor threshold can be raised to select songs for listing using the increased risk factor threshold.

[0087] For example, the risk factor of a song can be compared with a risk factor threshold, and songs that do not exceed the risk factor threshold will be put on the shelves, while songs that exceed the risk factor threshold will not be put on the shelves. Therefore, when deciding whether to put a song on the shelves, a risk quantitative assessment can be performed based on the risk factor of the song obtained, and the quantitative risk factor threshold can be used to dynamically adjust the song's on-shelf and off-shelf strategies.

[0088] In an exemplary embodiment, the following steps may also be included: for songs of a preset type in the music library, based on a specific risk factor threshold set for the songs of the preset type, determining whether the songs of the preset type are songs on the shelves; for songs other than the preset type in the music library, determining whether the other songs are songs on the shelves based on the risk factor threshold of the music library.

[0089] In one example, a specific risk factor threshold can be set for songs of preset types in the music library (e.g., songs with high value, which songs are high-value songs, such as some popular songs) so that high-value songs do not use the global risk factor threshold of the music library, but are adjusted for listing according to the specific risk factor threshold, and other songs other than the preset types of songs (e.g., non-high-value songs) can be adjusted for listing according to the global risk factor threshold of the music library. In this way, the risk factor threshold of the song can be set in combination with the value of the song, and the listing and delisting strategy of each song can be finely controlled to ensure that the overall risk of the music library is controllable, while also helping to maximize the value of the music library.

[0090] In an exemplary embodiment, as shown in FIG5 , a method for training a song risk comparison model is provided. This embodiment uses the method applied to a terminal as an example. It is understandable that the method can also be applied to a server, and can also be applied to a system including a terminal and a server, and is implemented through the interaction between the terminal and the server. The server can be implemented as an independent server or a server cluster composed of multiple servers; the terminal can be, but is not limited to, various personal computers, laptops, smart phones, tablet computers, portable wearable devices, etc.

[0091] In this embodiment, the method includes the following steps S501 to S502. Among them:

[0092] In step S501, any two popular songs in the music library are obtained to construct multiple sample pairs; the two popular songs in the sample pairs are different songs.

[0093] In step S502, a song risk comparison model is trained based on the song risk comparison labels of each sample pair and the song features of the popular songs in each sample pair; the song risk comparison labels are used to characterize the level of content risk of the two popular songs in the corresponding sample pair; the song features include at least one of song content features, song creator features, and song interaction features in addition to security algorithm features.

[0094] Among them, the song risk comparison model can be used in any song content risk assessment method.

[0095] In a specific implementation, a small batch of representative songs can be selected from the popular songs in the music library to construct multiple song pairs (i.e., sample pairs). By auditing and comparing the risks of each song pair, multiple audit data can be obtained; such as <song 1, song 2, risk identifier>, where risk identifier (i.e., song risk comparison label) 0 can indicate that the risks of the two songs are equivalent, risk identifier 1 can indicate that the risk of song 1 is greater than the risk of song 2, risk identifier 2 can indicate that the risk of song 2 is greater than the risk of song 1, and risk identifier 3 can indicate that the risk of the two songs cannot be determined.

[0096] In one example, song features can be collected for each popular song in multiple sample pairs, which may include at least one of song content features, song creator features, and song interaction features in addition to security algorithm features. Thus, by using security algorithm features and combining user interaction behavior, lyricist / composer portrait information, and song content features for comprehensive risk assessment, new illegal content can be identified more effectively, improving the risk assessment effect of song content.

[0097] In another example, based on the song risk comparison labels of multiple sample pairs and the song features of each popular song in the multiple sample pairs, a model can be built to predict the risk level of two songs; specifically, supervised machine learning methods can be used for modeling, such as regression algorithms, decision tree algorithms, neural networks, etc., or multiple models can be integrated to improve the performance of the model, and then a song risk comparison model can be trained. Thus, by simplifying the direct prediction of risk coefficients or levels to predicting the risk level of two songs, the complexity of risk prediction modeling can be reduced.

[0098] In one embodiment, as shown in FIG6 , a flow chart of another song content risk assessment method is provided. In this embodiment, the method includes the following steps:

[0099] In step S601, song features of multiple popular songs in the music library are obtained. The song features include at least one of song content features, song creator features, and song interaction features in addition to security algorithm features. In step S602, according to a preset comparison and sorting method, two popular songs that participate in the comparison of song risks are determined from multiple popular songs for multiple times, and the song features of the two popular songs determined each time are input into a pre-trained song risk comparison model to obtain the risk comparison results of the two popular songs determined each time. The comparison and sorting method sorts multiple popular songs according to the risk comparison results of the two popular songs determined each time. In step S603, according to the exponential distribution partitioning rule, the sorted multiple popular songs are divided into multiple intervals, and the risk coefficient of the popular songs in each interval is set. In step S604, a pre-trained deep neural network for determining the song category of a song is obtained, and the input layer and multiple feature vector extraction layers in the deep neural network are used as feature vector learning networks. In step S605, at least one song feature of unpopular songs in the music library is input into the feature vector learning network to obtain the risk feature vector of the unpopular songs, and at least one song feature of each popular song is input into the feature vector learning network to obtain the risk feature vector of each popular song. In step S606, at least one target popular song is determined among the popular songs based on the similarity between the risk feature vector of the unpopular songs and the risk feature vector of each popular song. In step S607, the risk coefficient of at least one target popular song is weighted averaged to obtain the risk coefficient of the unpopular song. It should be noted that the specific limitations of the above steps can be found in the specific limitations of a song content risk assessment method above, which will not be repeated here.

[0100] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0101] Based on the same inventive concept, the embodiment of the present application also provides a song content risk assessment device for implementing the song content risk assessment method involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more song content risk assessment device embodiments provided below can refer to the limitations of the song content risk assessment method above, and will not be repeated here.

[0102] In one embodiment, a song content risk assessment device is provided, comprising:

[0103] A song feature acquisition module, used to acquire song features of multiple popular songs in the music library, wherein the song features include at least one of song content features, song creator features, and song interaction features in addition to security algorithm features;

[0104] A song risk ranking module is used to determine two popular songs that participate in the comparison of song risks from the multiple popular songs multiple times according to a preset comparison and ranking method, and input the song features of the two popular songs determined each time into a pre-trained song risk comparison model to obtain the risk comparison result of the two popular songs determined each time; the comparison and ranking method is used to sort the multiple popular songs according to the risk comparison result of the two popular songs determined each time;

[0105] The song risk coefficient determination module is used to divide the sorted multiple popular songs into multiple intervals and set the risk coefficient of the popular songs in each interval, wherein the risk coefficients of the popular songs in the same interval are the same.

[0106] In one embodiment, the song risk coefficient determination module includes:

[0107] The interval division submodule is used to divide the sorted multiple popular songs into the multiple intervals according to the exponential distribution division rule.

[0108] In one embodiment, the apparatus further comprises:

[0109] a risk feature vector determination module, configured to input at least one song feature of unpopular songs in the music library into a feature vector learning network to obtain a risk feature vector of the unpopular songs, and input at least one song feature of each of the popular songs into the feature vector learning network to obtain a risk feature vector of each of the popular songs; wherein the types of song features of the unpopular songs are the same as the types of song features of the popular songs, and include at least one of a security algorithm feature, a song content feature, a song creator feature, and a song interaction feature;

[0110] a target popular song determination module, configured to determine at least one target popular song from among the popular songs according to the similarity between the risk feature vectors of the unpopular songs and the risk feature vectors of the popular songs;

[0111] The module for obtaining the risk coefficient of unpopular songs is used to obtain the risk coefficient of the unpopular songs based on the risk coefficient of the at least one target popular song.

[0112] In one embodiment, the apparatus further comprises:

[0113] A training network acquisition module, used to obtain a pre-trained deep neural network for determining the song category of a song; wherein the deep neural network includes an input layer, a plurality of feature vector extraction layers and an output layer in sequence, and after at least one song feature of the popular songs in the same interval is input into the input layer, the song category of each of the popular songs output by the output layer is the same;

[0114] A feature vector learning network obtaining module, used for using the input layer and the plurality of feature vector extraction layers as the feature vector learning network;

[0115] The risk feature vector determination module includes:

[0116] A submodule for obtaining a risk feature vector of an unpopular song, for inputting at least one song feature of the unpopular song into an input layer of the feature vector learning network, and using a feature vector output by a feature vector extraction layer of the last layer among the multiple feature vector extraction layers as a risk feature vector of the unpopular song;

[0117] The submodule for obtaining risk feature vectors of popular songs is used to input at least one song feature of each of the popular songs into the input layer of the feature vector learning network, and use the feature vector output by the last feature vector extraction layer among the multiple feature vector extraction layers as the risk feature vector of each of the popular songs.

[0118] In one embodiment, the module for obtaining the risk coefficient of unpopular songs includes:

[0119] The weighted average processing submodule is used to perform weighted average processing on the risk coefficient of the at least one target popular song to obtain the risk coefficient of the unpopular song.

[0120] In one embodiment, the apparatus further comprises:

[0121] The risk coefficient obtaining module is used to perform weighted average processing on the risk coefficients of each of the popular songs and the risk coefficients of each of the unpopular songs in the music library to obtain the risk coefficient of the music library.

[0122] In one embodiment, the apparatus further comprises:

[0123] a risk coefficient threshold adjustment module, configured to lower the risk coefficient threshold of the music library when a need to improve the health of the music library is detected, and to increase the risk coefficient threshold of the music library when a need to improve the richness of the music library is detected; wherein the risk coefficient threshold of the music library is determined based on the risk coefficient of each song in the music library, each of which includes the popular songs and the unpopular songs;

[0124] The module for determining songs to be put on the shelves is used to determine songs in the music library whose risk coefficients do not exceed the adjusted risk coefficient threshold of the music library as songs to be put on the shelves.

[0125] In one embodiment, the apparatus further comprises:

[0126] A preset type song listing module is used to determine whether a song of a preset type in the music library is a song to be listed based on a specific risk factor threshold set for the song of the preset type;

[0127] The module for putting other songs on the shelves is used to determine whether other songs in the music library other than the preset type of songs are songs to be put on the shelves based on the risk coefficient threshold of the music library.

[0128] Each module in the above song content risk assessment device can be implemented in whole or in part by software, hardware, or a combination thereof. Each module can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute operations corresponding to each module.

[0129] Based on the same inventive concept, the embodiment of the present application also provides a song risk comparison model training device for implementing the song risk comparison model training method involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in the embodiments of the training device for one or more song risk comparison models provided below can refer to the limitations of the song risk comparison model training method above, and will not be repeated here.

[0130] In one embodiment, a training device for a song risk comparison model is provided, comprising:

[0131] A sample pair construction module is used to obtain any two popular songs in the music library and construct multiple sample pairs; the two popular songs in the sample pairs are different songs;

[0132] The model training module is used to train a song risk comparison model based on the song risk comparison labels of each of the sample pairs and the song features of the popular songs in each of the sample pairs; the song risk comparison labels are used to characterize the level of content risk of the two popular songs in the corresponding sample pairs; the song features include at least one of song content features, song creator features, and song interaction features in addition to security algorithm features.

[0133] Each module in the training device of the song risk comparison model can be implemented in whole or in part by software, hardware and a combination thereof. Each module can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute operations corresponding to each module.

[0134] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 7 As shown. The computer device includes a processor, a memory, an input / output interface, a communication interface, a display unit and an input device. The processor, the memory and the input / output interface are connected through a system bus, and the communication interface, the display unit and the input device are connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be implemented through WIFI, a mobile cellular network, NFC (near field communication) or other technologies. When the computer program is executed by the processor, a song content risk assessment method and / or a song risk comparison model training method is implemented.

[0135] Those skilled in the art will understand that Figure 7 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0136] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the following steps are implemented:

[0137] Acquire song features of multiple popular songs in the music library, wherein the song features include at least one of song content features, song creator features, and song interaction features in addition to security algorithm features;

[0138] According to a preset comparison and sorting method, two popular songs that participate in the risk comparison of songs are determined from the multiple popular songs for multiple times, and the song features of the two popular songs determined each time are input into the pre-trained song risk comparison model to obtain the risk comparison result of the two popular songs determined each time; the comparison and sorting method sorts the multiple popular songs according to the risk comparison result of the two popular songs determined each time;

[0139] The sorted plurality of popular songs are divided into a plurality of intervals, and risk coefficients of the popular songs in each interval are set, wherein the risk coefficients of the popular songs in the same interval are the same.

[0140] In one embodiment, when the processor executes the computer program, the steps in the other embodiments described above are also implemented.

[0141] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the following steps are implemented:

[0142] Obtain any two popular songs in the music library and construct multiple sample pairs; the two popular songs in the sample pairs are different songs;

[0143] Based on the song risk comparison labels of each sample pair and the song features of the popular songs in each sample pair, a song risk comparison model is trained; the song risk comparison labels are used to characterize the content risks of the two popular songs in the corresponding sample pair; the song features include at least one of the song content features, song creator features, and song interaction features in addition to the security algorithm features;

[0144] The song risk comparison model is used in any of the song content risk assessment methods described.

[0145] In one embodiment, when the processor executes the computer program, the steps in the other embodiments described above are also implemented.

[0146] In one embodiment, a computer readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the following steps are implemented:

[0147] Acquire song features of multiple popular songs in the music library, wherein the song features include at least one of song content features, song creator features, and song interaction features in addition to security algorithm features;

[0148] According to a preset comparison and sorting method, two popular songs that participate in the risk comparison of songs are determined from the multiple popular songs for multiple times, and the song features of the two popular songs determined each time are input into the pre-trained song risk comparison model to obtain the risk comparison result of the two popular songs determined each time; the comparison and sorting method sorts the multiple popular songs according to the risk comparison result of the two popular songs determined each time;

[0149] The sorted plurality of popular songs are divided into a plurality of intervals, and risk coefficients of the popular songs in each interval are set, wherein the risk coefficients of the popular songs in the same interval are the same.

[0150] In one embodiment, when the computer program is executed by a processor, the steps in the other embodiments described above are also implemented.

[0151] In one embodiment, a computer readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the following steps are implemented:

[0152] Obtain any two popular songs in the music library and construct multiple sample pairs; the two popular songs in the sample pairs are different songs;

[0153] Based on the song risk comparison labels of each sample pair and the song features of the popular songs in each sample pair, a song risk comparison model is trained; the song risk comparison labels are used to characterize the content risks of the two popular songs in the corresponding sample pair; the song features include at least one of the song content features, song creator features, and song interaction features in addition to the security algorithm features;

[0154] The song risk comparison model is used in any of the song content risk assessment methods described.

[0155] In one embodiment, when the computer program is executed by a processor, the steps in the other embodiments described above are also implemented.

[0156] In one embodiment, a computer program product is provided, comprising a computer program, which, when executed by a processor, implements the following steps:

[0157] Acquire song features of multiple popular songs in the music library, wherein the song features include at least one of song content features, song creator features, and song interaction features in addition to security algorithm features;

[0158] According to a preset comparison and sorting method, two popular songs that participate in the risk comparison of songs are determined from the multiple popular songs for multiple times, and the song features of the two popular songs determined each time are input into the pre-trained song risk comparison model to obtain the risk comparison result of the two popular songs determined each time; the comparison and sorting method sorts the multiple popular songs according to the risk comparison result of the two popular songs determined each time;

[0159] The sorted plurality of popular songs are divided into a plurality of intervals, and risk coefficients of the popular songs in each interval are set, wherein the risk coefficients of the popular songs in the same interval are the same.

[0160] In one embodiment, when the computer program is executed by a processor, the steps in the other embodiments described above are also implemented.

[0161] In one embodiment, a computer program product is provided, comprising a computer program, which, when executed by a processor, implements the following steps:

[0162] Obtain any two popular songs in the music library and construct multiple sample pairs; the two popular songs in the sample pairs are different songs;

[0163] Based on the song risk comparison labels of each sample pair and the song features of the popular songs in each sample pair, a song risk comparison model is trained; the song risk comparison labels are used to characterize the content risks of the two popular songs in the corresponding sample pair; the song features include at least one of the song content features, song creator features, and song interaction features in addition to the security algorithm features;

[0164] The song risk comparison model is used in any of the song content risk assessment methods described.

[0165] In one embodiment, when the computer program is executed by a processor, the steps in the other embodiments described above are also implemented.

[0166] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant laws, regulations and standards of relevant countries and regions.

[0167] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but are not limited to this.

[0168] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0169] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A song content risk assessment method, characterized in that: The method comprises: Acquire song features of multiple popular songs in the music library, wherein the song features include at least one of song content features, song creator features, and song interaction features in addition to security algorithm features; According to a preset comparison and sorting method, two popular songs that participate in the risk comparison of songs are determined from the multiple popular songs for multiple times, and the song features of the two popular songs determined each time are input into the pre-trained song risk comparison model to obtain the risk comparison result of the two popular songs determined each time; the comparison and sorting method sorts the multiple popular songs according to the risk comparison result of the two popular songs determined each time; The sorted plurality of popular songs are divided into a plurality of intervals, and risk coefficients of the popular songs in each interval are set, wherein the risk coefficients of the popular songs in the same interval are the same.

2. The method according to claim 1, characterized in that: The step of dividing the sorted plurality of popular songs into a plurality of intervals includes: According to the exponential distribution division rule, the sorted plurality of popular songs are divided into the plurality of intervals.

3. The method according to claim 1, characterized in that After the step of setting the risk factor of the popular songs in each of the intervals, the method further includes: Inputting at least one song feature of unpopular songs in the music library into a feature vector learning network to obtain a risk feature vector of the unpopular songs, and inputting at least one song feature of each of the popular songs into the feature vector learning network to obtain a risk feature vector of each of the popular songs; wherein the types of song features of the unpopular songs are the same as the types of song features of the popular songs, and include at least one of a security algorithm feature, a song content feature, a song creator feature, and a song interaction feature; Determining at least one target popular song from among the popular songs according to the similarity between the risk feature vector of the unpopular song and the risk feature vector of each of the popular songs; The risk coefficient of the unpopular song is obtained based on the risk coefficient of the at least one target popular song.

4. The method according to claim 3, characterized in that The method further comprises: Obtaining a pre-trained deep neural network for determining the song category of a song; wherein the deep neural network sequentially comprises an input layer, a plurality of feature vector extraction layers, and an output layer, and after at least one song feature of the popular songs in the same interval is input into the input layer, the song category of each of the popular songs output by the output layer is the same; Using the input layer and the plurality of feature vector extraction layers as the feature vector learning network; The step of inputting at least one song feature of the unpopular songs in the music library into a feature vector learning network to obtain a risk feature vector of the unpopular songs, and inputting at least one song feature of each of the popular songs into the feature vector learning network to obtain a risk feature vector of each of the popular songs comprises: Inputting at least one song feature of the unpopular song into the input layer of the feature vector learning network, and using the feature vector output by the last feature vector extraction layer among the multiple feature vector extraction layers as the risk feature vector of the unpopular song; At least one song feature of each of the popular songs is input into the input layer of the feature vector learning network, and the feature vector output by the last feature vector extraction layer among the multiple feature vector extraction layers is used as the risk feature vector of each of the popular songs.

5. The method according to claim 3, characterized in that: The step of obtaining the risk coefficient of the unpopular song based on the risk coefficient of the at least one target popular song includes: The risk coefficient of the at least one target popular song is processed by weighted average to obtain the risk coefficient of the unpopular song.

6. The method according to claim 3, characterized in that The method further comprises: The risk coefficient of each popular song and the risk coefficient of each unpopular song in the music library are weighted averaged to obtain the risk coefficient of the music library.

7. The method according to claim 3, characterized in that The method further comprises: When a need to improve the health of the music library is detected, the risk coefficient threshold of the music library is lowered; when a need to improve the richness of the music library is detected, the risk coefficient threshold of the music library is increased; wherein the risk coefficient threshold of the music library is determined based on the risk coefficient of each song in the music library, each of which includes the popular songs and the unpopular songs; Songs in the music library whose risk coefficients do not exceed the adjusted risk coefficient threshold of the music library are determined as songs to be put on the shelves.

8. The method according to claim 3, characterized in that The method further comprises: For songs of a preset type in the music library, determining whether the songs of the preset type are songs on the shelves based on a specific risk factor threshold set for the songs of the preset type; For other songs in the music library except the preset type of songs, it is determined whether the other songs are listed songs based on the risk coefficient threshold of the music library.

9. A method for training a song risk comparison model, characterized in that: The method comprises: Obtain any two popular songs in the music library and construct multiple sample pairs; the two popular songs in the sample pairs are different songs; Based on the song risk comparison labels of each sample pair and the song features of the popular songs in each sample pair, a song risk comparison model is trained; the song risk comparison labels are used to characterize the content risks of the two popular songs in the corresponding sample pair; the song features include at least one of the song content features, song creator features, and song interaction features in addition to the security algorithm features; The song risk comparison model is used in the song content risk assessment method described in any one of claims 1 to 8.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 9 are implemented.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 9 are implemented.

12. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 9 are implemented.

Images