Short comment generation method and device, equipment and storage medium

Abstract

The invention discloses a short comment generation method and device, equipment and a storage medium, and relates to the technical field of computers. The method comprises the following steps: acquiring multiple pieces of comment information of first multimedia content; extracting a plurality of pieces of target comment information from the plurality of pieces of comment information, wherein the target comment information refers to comment information with recommendation significance for the first multimedia content; segmenting the plurality of pieces of target comment information to obtain a plurality of segmented short sentences; generating a short comment recommendation library of the first multimedia content through a language model according to the information of the first multimedia content and the plurality of segmented short sentences, the short comment recommendation library including a plurality of recommended short comments; non-unique short reviews in the multiple recommended short reviews are removed, a final recommended short review of the first multimedia content is selected from the remaining recommended short reviews, and the non-unique short reviews refer to recommended short reviews associated with multiple other multimedia contents except the first multimedia content. According to the method, the quality of the final recommendation short comments and the recommendation effect of the final recommendation short comments are improved.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a method, device, equipment and storage medium for generating short reviews. Background Art

[0002] When an application displays multimedia content, it usually displays the reason for recommending the multimedia content next to the name of the multimedia content.

[0003] In the related technology, the original comments of the multimedia content are directly obtained, and traditional NLP (Natural Language Processing) methods such as key phrase extraction and summary generation are used for the original comments of the multimedia content. The word frequency and the number of times the phrases appear in the original comments are used as weights, and short comments are extracted from the original comments of the multimedia content as the reasons for recommending the multimedia content.

[0004] However, the short reviews extracted by the above method are prone to contain semantically incomplete fragments, general short reviews, negative short reviews, and other text descriptions that are not suitable as recommendations, and the quality of the short reviews is low. Summary of the invention

[0005] The embodiment of the present application provides a method, device, equipment and storage medium for generating a short review. The technical solution provided by the embodiment of the present application is as follows:

[0006] According to one aspect of an embodiment of the present application, a method for generating a short review is provided, the method comprising:

[0007] Acquire multiple comment information of the first multimedia content;

[0008] Extracting a plurality of target comment information from the plurality of comment information, wherein the target comment information refers to comment information having a recommendation significance for the first multimedia content;

[0009] Segmenting the plurality of target review information respectively to obtain a plurality of segmented short sentences;

[0010] Generate a short review recommendation library of the first multimedia content according to the information of the first multimedia content and the plurality of segmented short sentences through a language model, wherein the short review recommendation library includes a plurality of recommended short reviews;

[0011] Non-unique short reviews from the plurality of recommended short reviews are eliminated, and a final recommended short review of the first multimedia content is selected from the remaining recommended short reviews, wherein the non-unique short review refers to a recommended short review associated with a plurality of other multimedia contents except the first multimedia content.

[0012] According to one aspect of an embodiment of the present application, a short review generating device is provided, the device comprising:

[0013] A comment acquisition module, used to acquire multiple comment information of the first multimedia content;

[0014] A comment extraction module, used to extract a plurality of target comment information from the plurality of comment information, wherein the target comment information refers to comment information having a recommendation significance for the first multimedia content;

[0015] A comment segmentation module, used to segment the plurality of target comment information respectively to obtain a plurality of segmented short sentences;

[0016] A short review generation module, configured to generate a short review recommendation library of the first multimedia content according to the information of the first multimedia content and the plurality of segmented short sentences through a language model, wherein the short review recommendation library includes a plurality of recommended short reviews;

[0017] The short review selection module is used to eliminate non-unique short reviews from the multiple recommended short reviews and select a final recommended short review of the first multimedia content from the remaining recommended short reviews, wherein the non-unique short review refers to a recommended short review that is associated with multiple other multimedia contents except the first multimedia content.

[0018] According to one aspect of an embodiment of the present application, a computer device is provided, comprising a processor and a memory, wherein a computer program is stored in the memory, and the computer program is loaded and executed by the processor to implement the above-mentioned short review generating method.

[0019] According to one aspect of an embodiment of the present application, a computer-readable storage medium is provided, in which a computer program is stored. The computer program is loaded and executed by a processor to implement the above-mentioned short review generating method.

[0020] According to one aspect of an embodiment of the present application, a computer program product is provided. The computer program product includes a computer program. The computer program is loaded and executed by a processor to implement the above-mentioned short review generating method.

[0021] The technical solution provided in the embodiments of the present application can bring the following beneficial effects:

[0022] By extracting multiple target review information with recommendation significance for the first multimedia content from multiple review information, the final recommended short review is prevented from being presented as a negative short review or other short review that has no recommendation significance, thereby ensuring the recommendation quality of the final recommended short review. In addition, the recommended short review library is generated through the extraction and rewriting capabilities of the language model, thereby avoiding the appearance of semantically incomplete fragments in the generated recommended short reviews, reducing data noise, significantly improving the quality of the recommended short reviews, and improving the generation efficiency of the recommended short reviews. In addition, by eliminating non-unique short reviews from multiple recommended short reviews, reducing the proportion of general short reviews in the recommended short reviews, avoiding the appearance of sentences that generally praise the first multimedia content, further improving the quality of the final recommended short review, thereby improving the recommendation effect of the final recommended short review. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a schematic diagram of an implementation environment of a solution provided by an embodiment of the present application;

[0024] Figure 2 is a schematic diagram of a song display interface provided by an embodiment of the present application;

[0025] Figure 3 is a flow chart of a method for generating a short review provided by an embodiment of the present application;

[0026] Figure 4 is a flow chart of a second vocabulary extraction process provided by an embodiment of the present application;

[0027] Figure 5 is a flowchart of a process for extracting first comment information provided by an embodiment of the present application;

[0028] Figure 6 is a flowchart of a process for extracting second comment information provided by an embodiment of the present application;

[0029] Figure 7 This is a flow chart for eliminating non-unique short reviews provided by an embodiment of the present application;

[0030] Figure 8 is a schematic diagram of a short review generation process provided by an embodiment of the present application;

[0031] Fig. 9 is a block diagram of a short review generating device provided by an embodiment of the present application;

[0032] Fig.10 It is a structural block diagram of a computer device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0033] In order to make the objectives, technical solutions and advantages of the present application clearer, the implementation methods of the present application will be further described in detail below with reference to the accompanying drawings.

[0034] Please refer to Figure 1 , which shows a schematic diagram of a solution implementation environment provided by an embodiment of the present application. The solution implementation environment can be implemented as a short review generation system. The solution implementation environment may include: a terminal device 10 and a server 20.

[0035] The number of terminal devices 10 can be one or more. The terminal device 10 can be an electronic device such as a mobile phone, a tablet computer, a laptop computer, a desktop computer, a game console, an e-book reader, a multimedia player, a wearable device, an intelligent voice interaction device, a smart home appliance, a vehicle terminal, an aircraft, etc.

[0036] The terminal device 10 may be installed with a client of a target application, and the target application has the function of displaying multimedia content and comment information, and the user can view the multimedia content and at least one comment information corresponding to each multimedia content in the target application. The present application does not limit the type of the target application, including but not limited to music applications, video applications, reading applications, web applications, etc. Optionally, the target application may be an application that needs to be downloaded and installed, or an application that can be used instantly, and the present application does not limit this.

[0037] The server 20 is used to provide background services for the client of the target application installed and running in the terminal device 10. For example, the server 20 can be the background server of the above application. The server 20 can be an independent physical server, or a server cluster composed of multiple servers, or a cloud computing service center. Optionally, the server 20 provides background services for the applications in multiple terminal devices 10 at the same time. The terminal device 10 and the server 20 can communicate with each other through the network.

[0038] In an embodiment of the present application, it is necessary to first obtain multiple comment information of the first multimedia content, extract multiple target comment information with recommendation significance for the first multimedia content from the multiple comment information, and then segment the multiple target comment information respectively to obtain multiple segmented short sentences. A short review recommendation library of the first multimedia content is generated based on the information of the first multimedia content and the multiple segmented short sentences through a language model, and the short review recommendation library contains multiple recommended short reviews. Non-unique short reviews are eliminated from the multiple recommended short reviews, and the final recommended short review of the first multimedia content is selected from the remaining recommended short reviews.

[0039] The first multimedia content can be displayed on a multimedia content display interface, which includes but is not limited to a multimedia content display list interface, a multimedia content recommendation interface, a multimedia content search interface, etc. The final recommendation short comment is usually displayed next to the name of the first multimedia content. For example, the final recommendation short comment can be displayed below, above, to the right, to the left, at the bottom right, at the top right, etc. of the name of the first multimedia content. Figure 2 As shown, Figure 2 A schematic diagram of a song display interface is shown, and the multimedia content is songs. Figure 2 Figure (1) shows a song display list interface 210, which displays the names of multiple songs in the song display list, and the final recommended short review 201 is displayed in the upper right corner of the name of each song. Figure 2 Figure (2) shows a song recommendation interface 220, which displays the names of multiple recommended songs, and a final recommendation short review 201 is displayed in the upper right corner of the name of each recommended song. Figure 2 Figure (3) shows a song search interface 230, which displays the names of multiple songs matching the search content, and the final recommended short review 201 is displayed below the name of each song.

[0040] Please refer to Figure 3 , which shows a flow chart of a method for generating a short review provided by an embodiment of the present application. The execution subject of each step of the method may be a computer device. The method may include at least one of the following steps 310 to 350:

[0041] Step 310: Acquire multiple comment information of the first multimedia content.

[0042] The first multimedia content may be any multimedia content, including but not limited to songs, videos, novels, articles, etc. Comment information is text information published by a user on the first multimedia content, used to indicate the user's feelings after watching / listening to the first multimedia content.

[0043] Step 320 , extracting multiple target comment information from the multiple comment information, where the target comment information refers to comment information having a recommendation significance for the first multimedia content.

[0044] The multiple comment information includes comment information that has a recommendation significance for the first multimedia content, and comment information that has no recommendation significance for the first multimedia content. Comment information that has a recommendation significance for the first multimedia content generally refers to comment information that is highly relevant to the first multimedia content, or is expressed as a noun. Comment information that has no recommendation significance for the first multimedia content generally refers to comment information that is less relevant to the first multimedia content, and comment information that does not meet content requirements, such as comment information used to deceive likes, comment information with multiple rounds of dialogue sentences, and comment information that is too long or too short.

[0045] Step 330 , segment the plurality of target comment information respectively to obtain a plurality of segmented short sentences.

[0046] Each target comment information is segmented to obtain at least one segmented short sentence corresponding to each target comment information.

[0047] Optionally, the target comment information may be segmented according to the punctuation marks contained in the target comment information to obtain a plurality of segmented short sentences. For example, if the target comment information is "This song has a beautiful melody and rich emotions, and XXX sings it very well.", then according to the punctuation marks contained in the target comment information, the target comment information may be segmented into two segmented short sentences, "This song has a beautiful melody and rich emotions" and "XXX sings it very well."

[0048] Optionally, a comment segmentation model may be used to segment the target comment information according to the semantic information contained in the target comment information to obtain a plurality of segmentation short sentences. For example, if the target comment information is "This song has a beautiful melody and rich emotions, and XXX sings it very well." Then, the comment segmentation model may be used to segment the target comment information into three segmentation short sentences, "This song has a beautiful melody," "Rich emotions," and "XXX sings very well," according to the semantic information contained in the target comment information.

[0049] Step 340 , generating a short review recommendation library of the first multimedia content according to the information of the first multimedia content and the plurality of segmented short sentences through a language model, wherein the short review recommendation library includes a plurality of recommended short reviews.

[0050] The language model refers to a large language model (LLM). The large language model here can be any publicly available large language model, such as a natural language model based on a transformer structure trained with a large amount of data. The large amount of data can reach a sample level of more than 100 million, and this application does not limit this.

[0051] In some embodiments, the information of the first multimedia content includes N sub-information, each sub-information is used to indicate the configuration of a dimension of the first multimedia content, and N is a positive integer. The information of the first multimedia content may also include content information of the first multimedia content, which is a textual representation of the first multimedia content.

[0052] Exemplarily, if the first multimedia content is a song, the information of the first multimedia content is used to indicate the production information of the song, including but not limited to the singer name, song name, album name, lyricist name, composer name, lyrics information, etc. If the first multimedia content is a video, the information of the first multimedia content is used to indicate the production information of the video, including but not limited to the video title, names of people included in the video, video director, video editor, video lines, etc.

[0053] According to the information of the first multimedia content and the multiple segmented short sentences, input prompt information is formed, and the input prompt information is input into the language model to generate a recommended short review of the first multimedia content. Through multiple inputs to the language model, a short review recommendation library of the first multimedia content is generated, and the short review recommendation library contains multiple recommended short reviews.

[0054] Optionally, the input prompt information may include N sub-information and multiple segmented short sentences. For example, if the first multimedia content is a song, the input prompt information includes the singer's name, the song's name, the album's name, the lyricist's name, the composer's name, and multiple segmented short sentences. Optionally, the input prompt information may also include at least one sub-information among the N sub-information and multiple segmented short sentences. For example, if the first multimedia content is a song, the input prompt information may include the singer's name, the song's name, and multiple segmented short sentences.

[0055] For example, the input prompt information can be specifically expressed as "You are a song promoter. I input singer-song-song review. Please extract 3 eye-catching positive evaluation phrases from the perspectives of song popularity, influence, theme, emotional resonance, listening scene, listening experience, song style, etc. described in these reviews as selling points for recommending to others. It is best to be interesting, fun, and exaggerated. The phrases must appear directly in the review without line breaks: {0}-{1}-{2}." Among them, 0, 1, and 2 are substituted for the singer name, song name, and recommendation short review respectively. The generated recommendation short review can be shown in Table 1 below.

[0056] Table 1

[0057] Singer name - Song name - "Roses are still romantic, youth is always bright, life is as bitter as a song" Singer name - Song name - "Tears in the eye, the depth of missed love, XXX's affectionate interpretation" Singer name - Song name - "I'm still listening to it in XXX years, my heart is beating fast, and the craze is coming"

[0058] Optionally, the recommended short review can be a short review that the language model rewrites and generates based on the information of the first multimedia content and the multiple segmented short sentences included in the input prompt information. Optionally, the recommended short review can also be a short review that the language model extracts from the multiple segmented short sentences based on the multiple segmented short sentences included in the input prompt information.

[0059] In some embodiments, the recommended short comment has a character limit. The recommended short comment needs to be displayed next to the name of the first multimedia content without interfering with the display of the name of the first multimedia content. Therefore, the recommended short comment has an upper limit and a lower limit of characters. The size of the upper limit and the lower limit of characters is related to display information such as the display font size of the recommended short comment, the display screen size of the terminal device, and the character length of the name of the first multimedia content, and this application does not limit this. For example, if the upper limit of the character number of the recommended short comment is 10 and the lower limit of the character number is 4, the language model is used to generate a recommended short comment of 4-10 characters.

[0060] Step 350 , remove non-unique short reviews from the plurality of recommended short reviews, and select a final recommended short review of the first multimedia content from the remaining recommended short reviews, wherein the non-unique short review refers to a recommended short review that is associated with a plurality of other multimedia contents except the first multimedia content.

[0061] If the non-unique short comments are related to multiple multimedia contents other than the first multimedia content, the non-unique short comments cannot produce a targeted recommendation effect on the first multimedia content, which is easy to cause aesthetic fatigue of users and reduce the conversion rate of users to the first multimedia content. Therefore, it is necessary to eliminate the non-unique short comments from the multiple recommended short comments and select the final recommended short comments of the first multimedia content from the remaining recommended short comments.

[0062] The remaining recommended comments contain at least one recommended comment, and a recommended comment is randomly selected from the remaining recommended comments as the final recommended comment. Optionally, each time the name of the first multimedia content is displayed, a recommended comment can be randomly selected from the remaining recommended comments as the final recommended comment displayed next to the name of the first multimedia content. Optionally, each time the name of the first multimedia content is displayed, a corresponding recommended comment can be selected from the remaining recommended comments for different display interfaces of the first multimedia content as the final recommended comment displayed next to the name of the first multimedia content. For example, in the display list interface of the first multimedia content, recommended comment 1 is selected from the remaining recommended comments as the final recommended comment; in the recommendation interface of the first multimedia content, recommended comment 2 is selected from the remaining recommended comments as the final recommended comment; in the search interface of the first multimedia content, recommended comment 3 is selected from the remaining recommended comments as the final recommended comment.

[0063] The technical solution provided in the embodiment of the present application extracts multiple target review information that has recommendation significance for the first multimedia content from multiple review information, thereby avoiding the final recommended short review from being a negative short review or other short review that has no recommendation significance, thereby ensuring the recommendation quality of the final recommended short review. And by generating a recommended short review library through the extraction and rewriting capabilities of the language model, it avoids the appearance of semantically incomplete fragments in the generated recommended short reviews, reduces data noise, significantly improves the quality of the recommended short reviews, and improves the generation efficiency of the recommended short reviews. And by eliminating non-unique short reviews from multiple recommended short reviews, reducing the proportion of general short reviews in the recommended short reviews, avoiding the appearance of sentences that generally praise the first multimedia content, further improving the quality of the final recommended short review, thereby improving the recommendation effect of the final recommended short review.

[0064] In some embodiments, step 320 includes at least one of sub-steps 321 to 323 (not shown).

[0065] Sub-step 321 , extracting at least one first comment information from the plurality of comment information according to the words contained in the comment information.

[0066] In some embodiments, step 321 includes at least one of sub-steps 3211 to 3214 (not shown in the figure).

[0067] Sub-step 3211, extracting multiple keywords from multiple comment information.

[0068] Keywords refer to words with practical meaning in the comment information, and a word segmentation tool is used to extract keywords from the comment information. For example, the word segmentation tool may refer to word segmentation tools such as jieba and hanlp.

[0069] In some embodiments, the keywords can be extracted from the comment information by introducing vocabulary in the corresponding field according to the type of the first multimedia content. For example, if the first multimedia content is a song, relevant vocabulary in the field of music can be introduced to extract keywords from the comment information to improve the accuracy of keyword extraction.

[0070] Sub-step 3212, extracting at least one first word from the multiple keywords according to the part of speech of the keywords, the first word including keywords whose part of speech is adjective and keywords whose part of speech is noun.

[0071] Parts of speech include nouns, adjectives, verbs, quantifiers, pronouns, prepositions, etc. At least one first word is extracted from the plurality of keywords according to the preset part of speech to be extracted.

[0072] Optionally, if the pre-set parts of speech to be extracted include nouns and adjectives, the first vocabulary includes keywords whose part of speech is adjectives and keywords whose part of speech is nouns. Optionally, if the pre-set parts of speech to be extracted include only nouns, the first vocabulary includes keywords whose part of speech is nouns.

[0073] By limiting the part of speech of the keyword, the extracted first word can be made more suitable for the information description of the first multimedia content itself, so as to reduce the comment information without actual meaning and improve the quality of the first comment information.

[0074] Sub-step 3213, calculates the occurrence frequency corresponding to at least one first word, and determines the first word whose occurrence frequency is greater than or equal to the first threshold as the second word, and the occurrence frequency refers to the proportion of the number of occurrences of the first word in multiple comment information relative to the total number of occurrences of multiple keywords in multiple comment information.

[0075] Obtain the number of times the first word appears in multiple comment information, and obtain the total number of times multiple keywords appear in multiple comment information, the frequency of occurrence of the first word = the number of times the first word appears in multiple comment information / the total number of times multiple keywords appear in multiple comment information.

[0076] A first word with a higher frequency of occurrence among at least one of the first words is extracted as the second word, specifically, a first word with a frequency of occurrence greater than or equal to a first threshold is determined as the second word. The size of the first threshold can be set by the technician according to the extraction requirements of the first comment information, and this application does not limit it.

[0077] For example, the extraction process of the second vocabulary can refer to Figure 4 As shown, Figure 4 The figure (1) in FIG. 1 shows that there is at least one first word, and the first word with a frequency greater than or equal to the first threshold is determined as the second word, and the Figure 4 Figure (2) in FIG. 1 shows that there is at least one second vocabulary. Figure 4 As shown, “regret” and “nice to hear” are screened out because their occurrence frequencies are less than the first threshold value and are not determined as the second vocabulary.

[0078] Sub-step 3214, obtaining at least one first comment information according to the number of likes of the comment information corresponding to the second word, where the first comment information is comment information with the number of likes greater than or equal to the second threshold.

[0079] Obtain the comment information corresponding to at least one second word, and then obtain the number of likes of these comment information, the number of likes of the comment information is used to indicate the user's recognition of the comment information. Extract the comment information with a higher number of likes, specifically extract the comment information with a number of likes greater than or equal to the second threshold, and obtain at least one first comment information. The size of the second threshold can be set by the technician according to the extraction requirements of the first comment information, and this application does not limit it.

[0080] By extracting comment information with the number of likes greater than or equal to the second threshold, the first comment information obtained is comment information with a generally high degree of recognition, so that the final recommended short review can be recognized by users, thereby improving the recommendation effect of the final recommended short review.

[0081] The extraction process of the first comment information can refer to Figure 5 As shown, first obtain multiple comment information, extract multiple keywords from the multiple comment information, perform part-of-speech tagging on the keywords, extract at least one first word based on the part-of-speech, then calculate the occurrence frequency of the first word, determine the first word with an occurrence frequency greater than or equal to a first threshold as the second word, and finally determine the comment information with a like amount greater than or equal to the second threshold as the first comment information.

[0082] Sub-step 322: extracting at least one second comment information from the multiple comment information according to the relevance between the comment information and the first multimedia content.

[0083] In some embodiments, step 322 includes at least one of sub-steps 3221 to 3225 (not shown in the figure).

[0084] Sub-step 3221: Use the first correlation model to score the relevance between the multiple comment information and the first multimedia content, and obtain first relevance scores corresponding to the multiple comment information.

[0085] The first relevance model is used to determine whether the comment information is related to the first multimedia content, and obtain a first relevance score corresponding to the comment information according to the relevance between the comment information and the first multimedia content. Exemplarily, if the first multimedia content is a song, the first relevance model is used to determine whether the comment information is related to the song, and obtain a first relevance score corresponding to the comment information according to the relevance between the comment information and the song.

[0086] The first relevance score ranges from 0 to 1, wherein a first relevance score of 0 indicates that the comment information is irrelevant to the first multimedia content, and a first relevance score of 1 indicates that the comment information is relevant to the first multimedia content. The first edge relevance score is used to indicate the relevance between the comment information and the first multimedia content.

[0087] Sub-step 3222: Use the second correlation model to score the relevance between the multiple comment information and the information of the first multimedia content, and obtain second relevance scores corresponding to the multiple comment information.

[0088] In some embodiments, for each of the multiple comment information, the relevance scores corresponding to the N sub-information are obtained based on the proportion of the text length of the N sub-information in the comment information; the second correlation model is used to calculate the text similarity between the comment information and the content information of the first multimedia content, and the similarity score corresponding to the content information is obtained; based on the relevance scores corresponding to the N sub-information and the similarity score corresponding to the content information, a second relevance score corresponding to the comment information is obtained.

[0089] If the comment information does not include the first sub-information, the relevance score corresponding to the first sub-information is 0.

[0090] Exemplarily, if the first multimedia content is a song, and the N sub-information include the singer's name, the song's name, and the album's name, then the association score corresponding to the singer's name is obtained based on the proportion of the singer's name's text length in the comment information, and the association score corresponding to the song's name is obtained based on the proportion of the song's text length in the comment information, and the association score corresponding to the album's name is obtained based on the proportion of the album's text length in the comment information.

[0091] The second correlation model is used to calculate the text similarity between the comment information and the content information of the first multimedia content. For example, if the first multimedia content is a song, the content information of the song is the lyrics information of the song, and the second correlation model is used to calculate the text similarity between the comment information and the lyrics information. If the first multimedia content is a video, the content information of the video is the lines information of the video, and the second correlation model is used to calculate the text similarity between the comment information and the lines information.

[0092] According to the weights corresponding to the relevance scores of the N sub-information and the similarity scores of the content information, the second relevance score corresponding to the comment information is calculated. The weights corresponding to the relevance scores of the N sub-information and the similarity scores of the content information are set by the technicians according to the extraction requirements of the second comment information, and are not limited in this application. Generally, the weight of the similarity score corresponding to the content information is greater than the weight of the relevance scores corresponding to the N sub-information.

[0093] By integrating the information of the first multimedia content and the correlation between the first multimedia content itself and the comment information, the correlation between the comment content of the comment information and the first multimedia content is fully considered, thereby ensuring the accuracy of the second comment information extraction, which helps to improve the recommendation quality of the short review.

[0094] Sub-step 3223, using a sentiment analysis model to analyze the favorable opinions of the plurality of review information on the first multimedia content, and obtaining favorable opinion scores corresponding to the plurality of review information.

[0095] The sentiment analysis model is used to analyze the degree of favorability of the review information for the first multimedia content. The favorability score ranges from 0 to 1, wherein a favorability score of 0 indicates that the review information is negative for the first multimedia content, and a favorability score of 1 indicates that the review information is positive for the first multimedia content.

[0096] Sub-step 3224, obtaining comprehensive scores corresponding to the plurality of review information respectively according to the first relevance scores, the second relevance scores and the favorable review scores respectively corresponding to the plurality of review information.

[0097] According to the weights corresponding to the first relevance scores, the second relevance scores and the favorable rating scores respectively corresponding to the plurality of review information, the comprehensive scores respectively corresponding to the plurality of review information are calculated. The comprehensive scores are used to indicate the relevance between the review information and the first multimedia content.

[0098] The weights corresponding to the first relevance score, the second relevance score and the favorable review score are set by the technical personnel according to the extraction requirements of the second comment information, and this application does not limit them.

[0099] Optionally, the weights corresponding to the first relevance score, the second relevance score, and the favorable review score may be the same, all 1 / 3. Optionally, the weights corresponding to the first relevance score, the second relevance score, and the favorable review score may be different, for example, the weight corresponding to the first relevance score and the weight corresponding to the favorable review score are greater than the weight corresponding to the second relevance score.

[0100] Sub-step 3225, determining the comment information whose comprehensive score is greater than or equal to the third threshold as the second comment information.

[0101] The size of the third threshold can be set by the technician according to the extraction requirements of the second comment information, and this application does not limit it.

[0102] The second comment information is determined through the above steps, so that the obtained second comment information is related to the entity characteristics of the first multimedia content itself and the sentiment analysis is positive, rather than comment information that generally praises the first multimedia content, thereby improving the quality of the second comment information and thus improving the quality of the final recommended short review.

[0103] The extraction process of the second comment information can refer to Figure 6 As shown, a first relevance score corresponding to the comment information is obtained based on the first relevance model, a second relevance score corresponding to the comment information is obtained based on the second relevance model, the second relevance score is obtained based on the relevance scores corresponding to the N sub-information and the similarity score corresponding to the content information, and a favorable review score corresponding to the comment information is obtained based on the sentiment analysis model. Based on the first relevance score, the second relevance score and the favorable review score corresponding to the comment information, a comprehensive score of the comment information is obtained, and the comment information with a comprehensive score greater than or equal to the third threshold is determined as the second comment information.

[0104] In some embodiments, after extracting at least one first comment information from multiple comment information according to the vocabulary included in the comment information, at least one second comment information is extracted from the remaining comment information according to the relevance between the comment information and the first multimedia content.

[0105] In some embodiments, after extracting at least one second comment information from multiple comment information according to the relevance between the comment information and the first multimedia content, at least one first comment information is extracted from the remaining comment information according to the vocabulary included in the comment information.

[0106] Sub-step 323, screening at least one first comment information and at least one second comment information to obtain multiple target comment information.

[0107] At least one first comment information and at least one second comment information are integrated, and the integrated comment information is screened to obtain a plurality of target comment information.

[0108] In some embodiments, step 323 includes at least one of sub-steps 3231 to 3233 (not shown in the figure).

[0109] Sub-step 3231, removing the comment information satisfying the removal condition from the at least one first comment information and the at least one second comment information, to obtain at least one remaining comment information.

[0110] In some embodiments, the elimination conditions include at least one of the following: the comment information contains preset conflicting words, the comment information is a multi-round dialogue sentence structure, the text length of the comment information is less than or equal to the fifth threshold, and the text length of the comment information is greater than or equal to the sixth threshold.

[0111] The comment information that meets the exclusion criteria can be found in Table 2 below.

[0112] Table 2

[0113]

[0114]

[0115] Conflicting words are words that are irrelevant to the first multimedia content, and are usually words used by users to deceive likes, including but not limited to words related to celebrities, famous schools, companies, diseases, regions, etc. Comment information containing conflicting words needs to be filtered. Therefore, the comment information "Like XX, comment XX", "XXX, I miss you", and "Everyone pointed out by the kitten can be admitted to XXX next year" in Table 2 above needs to be removed from at least one first comment information and at least one second comment information.

[0116] The comment information of the multi-round dialogue sentence is usually irrelevant to the first multimedia content, and includes at least one round of dialogue between at least two objects. The comment information of the multi-round dialogue sentence needs to be filtered, so the comment information of "XXX: XXX, if you can guess how many candies I have, I will give you two of my candies, how about it? XXX: OK, I guess 5." in Table 2 above needs to be removed from at least one first comment information and at least one second comment information.

[0117] The comment information with a text length less than or equal to the fifth threshold is too short, and the content value of the comment information is low. It is necessary to filter the comment information with a text length less than or equal to the fifth threshold. The size of the fifth threshold can be set by the technician according to the extraction requirements of the target comment information, and this application does not limit it. For example, the fifth threshold can be 3. Therefore, the comment information of "passed the exam", "accept or not", and "to..." in Table 2 above needs to be removed from at least one first comment information and at least one second comment information.

[0118] Comments with a text length greater than or equal to the sixth threshold are too long and have low content value. Comments with a text length greater than or equal to the sixth threshold need to be filtered. The size of the sixth threshold can be set by the technician according to the extraction requirements of the target comment information, and this application does not limit it. For example, the sixth threshold can be 20.

[0119] Sub-step 3232: Use the first correlation model to score the relevance between the at least one remaining comment information and the first multimedia content, and obtain a first relevance score corresponding to the at least one remaining comment information.

[0120] Optionally, the first relevance score corresponding to the at least one remaining comment information obtained in the above sub-step 3221 may also be directly obtained.

[0121] Sub-step 3233, eliminating the comment information whose first relevance score is less than or equal to the fourth threshold from at least one remaining comment information, to obtain multiple target comment information.

[0122] The size of the fourth threshold can be set by the technician according to the extraction requirements of the target comment information, and this application does not limit it.

[0123] In some embodiments, the first relevance score corresponding to at least one remaining comment information is multiplied by -1 to obtain the water content corresponding to at least one remaining comment information, and the comment information whose first relevance score is greater than or equal to the negative value of the fourth threshold in at least one remaining comment information is eliminated to obtain multiple target comment information.

[0124] By screening and filtering the extracted at least one first comment information and at least one second comment information, low-quality comment information that has no recommendation significance can be avoided in the target comment information, and the quality of the corpus sent to the language model in the next round can be improved, thereby improving the generation quality of the recommended short comments.

[0125] In some embodiments, step 350 includes at least one of sub-steps 351 to 355 (not shown).

[0126] Sub-step 351 , obtaining the number of associated multimedia contents corresponding to the plurality of recommended short reviews, wherein the associated multimedia contents refer to multimedia contents in the database whose first association scores with the recommended short reviews are greater than or equal to a seventh threshold.

[0127] A plurality of multimedia contents are stored in a database. A first correlation model is used to score the correlation between the plurality of recommended short reviews and the plurality of multimedia contents in the database, respectively, to obtain first correlation scores between the plurality of recommended short reviews and the plurality of multimedia contents in the database.

[0128] For each of the multiple recommended short reviews, obtain the first relevance scores between the recommended short reviews and the multiple multimedia contents in the database, and determine the multimedia contents whose first relevance scores are greater than or equal to the seventh threshold as the associated multimedia contents corresponding to the recommended short reviews. Obtain the number of associated multimedia contents. The size of the seventh threshold can be set by the technician according to the screening requirements of the recommended short reviews, and this application does not limit it.

[0129] Exemplarily, if the first multimedia content is a song, the number of associated songs corresponding to each of the plurality of recommended short reviews is obtained.

[0130] The number of multimedia contents associated with each recommended short review is used to indicate the concentration of the multimedia contents of the recommended short review, which represents the number of multimedia contents with similar short reviews.

[0131] Sub-step 352, obtaining at least one quantile of the number of associated multimedia contents according to the numbers of associated multimedia contents corresponding to the plurality of recommended short reviews, wherein the quantile is used to indicate the distribution ratio of the number of associated multimedia contents arranged in descending order.

[0132] The quantile may be any quantile method such as quartile, median, percentile, etc. The numbers of associated multimedia content corresponding to the multiple recommended short reviews are arranged in descending order, and at least one quantile of the number of associated multimedia content is obtained according to the selected quantile method.

[0133] The numbers of associated multimedia contents arranged in descending order may be shown in the following Table 3.

[0134] Table 3

[0135] Recommended short review Number of associated multimedia contents One song, one story 123 Hit Song 78 Never get tired of listening 76 A feast for the ears 62 Daily Song Recommendations 56 Childhood memories 55 Youth Memories 51 I fell in love with it as soon as I heard it 49

[0136] If the selected quantile method is quartiles, the first quartile in at least one quartile indicates that among the numbers of associated multimedia contents corresponding to the multiple recommended short reviews, 25% of the numbers are below it, the second quartile indicates that among the numbers of associated multimedia contents corresponding to the multiple recommended short reviews, 50% of the numbers are below it, the third quartile indicates that among the numbers of associated multimedia contents corresponding to the multiple recommended short reviews, 75% of the numbers are below it, and the fourth quartile indicates that among the numbers of associated multimedia contents corresponding to the multiple recommended short reviews, 100% of the numbers are below it.

[0137] If the selected quantile method is percentile, then the first percentile in at least one quantile point indicates that among the numbers of associated multimedia contents corresponding to the multiple recommended short reviews, 1% of the numbers are below it, the second percentile indicates that among the numbers of associated multimedia contents corresponding to the multiple recommended short reviews, 2% of the numbers are below it, the third percentile indicates that among the numbers of associated multimedia contents corresponding to the multiple recommended short reviews, 3% of the numbers are below it, and so on.

[0138] Sub-step 353: determining at least one recommended short review corresponding to a first quantile of at least one quantile as a common short review, wherein the first quantile is used to indicate the number of associated multimedia contents of a first proportion ranked at the top.

[0139] The first quantile is a quantile determined according to the quantile method of the quantile, and is used to indicate the number of the first proportion of the associated multimedia content ranked at the top. Exemplarily, if the selected quantile method is quartile, the first quantile can be a quantile among the first quartile, the second quartile, the third quartile and the fourth quartile. For example, if the first quantile is the third quartile, the third quartile is used to indicate the number of the top 25% of the associated multimedia content, and the recommended short comments corresponding to the number of the top 25% of the associated multimedia content are determined as general short comments.

[0140] Combined with Table 3 above, if the number of recommended short reviews is 8 and the first quartile is the third quartile, "One Song, One Story" and "Best-selling Divine Song" are determined as common short reviews.

[0141] The size of the first ratio can be set by the technical personnel according to the screening requirements of the recommended short reviews, and this application does not limit it.

[0142] Sub-step 354 , based on the common short reviews, remove non-unique short reviews from the plurality of recommended short reviews to obtain the remaining recommended short reviews.

[0143] In some embodiments, the semantic similarities between the general short reviews and the multiple recommended short reviews are calculated; the recommended short reviews whose semantic similarity is greater than or equal to an eighth threshold are determined as non-unique short reviews; and the non-unique short reviews among the multiple recommended short reviews are eliminated to obtain the remaining recommended short reviews.

[0144] By using the common short comments as identification seeds for non-unique short comments, and determining the recommended short comments whose semantic similarity with the common short comments is greater than or equal to the eighth threshold as non-unique short comments, at least one non-unique short comment corresponding to each common short comment can be obtained. The non-unique short comments corresponding to each common short comment in the plurality of recommended short comments are eliminated to obtain the remaining recommended short comments.

[0145] For example, the semantic similarity between the general short review "It sounds great" and multiple recommended short reviews can be shown in Table 4 below.

[0146] Table 4

[0147]

[0148]

[0149] The size of the eighth threshold can be set by the technical personnel according to the screening requirements of the recommended short reviews, and this application does not limit it.

[0150] For example, if the eighth threshold is 0.985 for the above Table 4, the non-unique short comments corresponding to the general short comment "It's so nice" include "It's so nice", "It's so nice", "It's really so nice" and "It's so nice".

[0151] Figure 7 The invention shows a process of removing non-unique short reviews from a plurality of recommended short reviews. First, based on a plurality of recommended short reviews included in a short review recommendation library, the number of associated multimedia contents respectively corresponding to the plurality of recommended short reviews is obtained, and according to the number of associated multimedia contents respectively corresponding to the plurality of recommended short reviews, a first quantile in at least one quantile of the number of associated multimedia contents is determined, and according to the first quantile, common short reviews from the plurality of recommended short reviews are obtained, and according to the semantic similarities between the common short reviews and the plurality of recommended short reviews, non-unique short reviews from the plurality of recommended short reviews are obtained, and the non-unique short reviews from the plurality of recommended short reviews are removed to obtain the remaining recommended short reviews.

[0152] Sub-step 355 , selecting a final recommended short review of the first multimedia content from the remaining recommended short reviews.

[0153] By identifying common reviews in recommended reviews and proposing non-unique reviews in recommended reviews, the proportion of common reviews in recommended reviews can be reduced, which helps to improve the quality of the final recommended reviews.

[0154] Figure 8 A schematic diagram of the short review generation process is shown. First, based on multiple review information to be input, the first review information and the second review information are extracted from the multiple review information. Then, the review information that meets the elimination conditions and the review information with a low first correlation score are eliminated from at least one first review information and at least one second review information to obtain multiple target review information. The multiple target review information is segmented to obtain multiple segmented short sentences. Based on the information of the first multimedia content and the multiple segmented short sentences, the input prompt information of the language model is constructed to generate a short review recommendation library containing multiple recommended short reviews. Finally, the non-unique short reviews in the multiple recommended short reviews are eliminated, and the final recommended short review of the first multimedia content is selected from the remaining recommended short reviews.

[0155] The following is an embodiment of the device of the present application, which can be used to execute the embodiment of the method of the present application. For details not disclosed in the embodiment of the device of the present application, please refer to the embodiment of the method of the present application.

[0156] Please refer to Fig. 9 , which shows a block diagram of a short review generation device provided by an embodiment of the present application. The device has the function of implementing the short review generation method described above, and the function can be implemented by hardware, or by hardware executing corresponding software. The device can be the computer device described above, or it can be set in a computer device. Fig. 9As shown, the device 900 may include: a comment acquisition module 910 , a comment extraction module 920 , a comment segmentation module 930 , a short comment generation module 940 and a short comment selection module 950 .

[0157] A comment acquisition module 910 is used to acquire multiple comment information of the first multimedia content;

[0158] A comment extraction module 920 is used to extract a plurality of target comment information from the plurality of comment information, wherein the target comment information refers to comment information having a recommendation significance for the first multimedia content;

[0159] A comment segmentation module 930 is used to segment the plurality of target comment information respectively to obtain a plurality of segmented short sentences;

[0160] A short review generation module 940, configured to generate a short review recommendation library of the first multimedia content according to the information of the first multimedia content and the plurality of segmented short sentences through a language model, wherein the short review recommendation library includes a plurality of recommended short reviews;

[0161] The short review selection module 950 is used to eliminate non-unique short reviews from the multiple recommended short reviews and select the final recommended short review of the first multimedia content from the remaining recommended short reviews. The non-unique short review refers to a recommended short review that is associated with multiple other multimedia contents except the first multimedia content.

[0162] In some embodiments, the comment extraction module 920 includes:

[0163] The first extraction unit is configured to extract at least one first comment information from the plurality of comment information according to words contained in the comment information.

[0164] The second extraction unit is configured to extract at least one second comment information from the multiple comment information according to the relevance between the comment information and the first multimedia content.

[0165] The comment screening unit is used to screen the at least one first comment information and the at least one second comment information to obtain the multiple target comment information.

[0166] In some embodiments, the first extraction unit is used to:

[0167] Extracting multiple keywords from the multiple comment information;

[0168] Extracting at least one first vocabulary from the plurality of keywords according to the part of speech of the keyword, the first vocabulary including keywords whose part of speech is adjective and keywords whose part of speech is noun;

[0169] Calculate the occurrence frequency corresponding to the at least one first word respectively, and determine the first word whose occurrence frequency is greater than or equal to the first threshold as the second word, wherein the occurrence frequency refers to the proportion of the number of occurrences of the first word in the multiple comment information relative to the total number of occurrences of the multiple keywords in the multiple comment information;

[0170] The at least one first comment information is obtained according to the number of likes of the comment information corresponding to the second word, where the first comment information is comment information whose number of likes is greater than or equal to a second threshold.

[0171] In some embodiments, the second extraction unit is used to:

[0172] Using a first correlation model, scoring the relevance between the plurality of comment information and the first multimedia content respectively, to obtain first relevance scores corresponding to the plurality of comment information respectively;

[0173] Using a second correlation model, scoring the relevance between the plurality of comment information and the information of the first multimedia content, respectively, to obtain second relevance scores corresponding to the plurality of comment information respectively;

[0174] Using a sentiment analysis model to analyze the favorable opinions of the plurality of review information on the first multimedia content, respectively, to obtain favorable opinion scores corresponding to the plurality of review information;

[0175] Obtaining comprehensive scores corresponding to the plurality of review information respectively according to the first relevance scores, the second relevance scores, and the favorable review scores respectively corresponding to the plurality of review information;

[0176] The review information whose comprehensive score is greater than or equal to the third threshold is determined as the second review information.

[0177] In some embodiments, the information of the first multimedia content includes N sub-information, each of which is used to indicate the configuration of a dimension of the first multimedia content, and N is a positive integer; the second extraction unit is used to:

[0178] For each of the plurality of comment information, obtaining the relevance scores corresponding to the N sub-information respectively according to the proportion of the text length of the N sub-information in the comment information;

[0179] Calculating text similarity between the comment information and content information of the first multimedia content using the second correlation model to obtain a similarity score corresponding to the content information;

[0180] A second relevance score corresponding to the comment information is obtained according to the relevance scores corresponding to the N sub-information and the similarity score corresponding to the content information.

[0181] In some embodiments, the comment screening unit is used to:

[0182] Eliminate the comment information that meets the elimination condition from the at least one first comment information and the at least one second comment information to obtain at least one remaining comment information;

[0183] Using a first correlation model, scoring the relevance between the at least one remaining comment information and the first multimedia content, respectively, to obtain first relevance scores corresponding to the at least one remaining comment information;

[0184] The comment information whose first relevance score is less than or equal to a fourth threshold is eliminated from the at least one remaining comment information to obtain the plurality of target comment information.

[0185] In some embodiments, the rejection condition includes at least one of the following:

[0186] The comment information contains preset conflicting words;

[0187] The comment information is a multi-round dialogue sentence structure;

[0188] The text length of the comment information is less than or equal to a fifth threshold;

[0189] The text length of the comment information is greater than or equal to a sixth threshold.

[0190] In some embodiments, the short review selection module 950 is used to:

[0191] Acquire the number of associated multimedia contents corresponding to the plurality of recommended short reviews, wherein the associated multimedia contents refer to multimedia contents in a database whose first association score with the recommended short reviews is greater than or equal to a seventh threshold;

[0192] Obtaining, according to the numbers of associated multimedia contents respectively corresponding to the plurality of recommended short reviews, at least one quantile of the number of associated multimedia contents, the quantile being used to indicate a distribution ratio of the number of associated multimedia contents arranged in descending order;

[0193] Determine at least one recommended short review corresponding to a first quantile among the at least one quantile as a common short review, wherein the first quantile is used to indicate the number of associated multimedia contents with a first proportion ranked high;

[0194] According to the common short reviews, the non-unique short reviews in the plurality of recommended short reviews are eliminated to obtain the remaining recommended short reviews;

[0195] A final recommended short review of the first multimedia content is selected from the remaining recommended short reviews.

[0196] In some embodiments, the short review selection module 950 is used to:

[0197] Calculating semantic similarities between the general short review and the plurality of recommended short reviews respectively;

[0198] Determine the recommended short reviews whose semantic similarity is greater than or equal to an eighth threshold as the non-unique short reviews;

[0199] The non-unique short reviews are eliminated from the plurality of recommended short reviews to obtain the remaining recommended short reviews.

[0200] It should be noted that the device provided in the above embodiment, when implementing its functions, is only illustrated by the division of the above functional modules. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the content structure of the device is divided into different functional modules to complete all or part of the functions described above. In addition, the device and method embodiments provided in the above embodiment belong to the same concept, and the specific implementation process is detailed in the method embodiment, which will not be repeated here.

[0201] Please refer to Fig.10 , which shows a block diagram of a computer device 1000 provided in one embodiment of the present application. The computer device 1000 may be any electronic device with data calculation, processing and storage functions. The computer device 1000 may be used to implement the short review generation method provided in the above embodiment.

[0202] Typically, the computer device 1000 includes a processor 1001 and a memory 1002 .

[0203] The processor 1001 may include one or more processing cores, such as a 4-core processor, an 8-core processor, etc. The processor 1001 may be implemented in at least one hardware form of DSP (Digital Signal Processing), FPGA (Field Programmable Gate Array), and PLA (Programmable Logic Array). The processor 1001 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 1001 may be integrated with a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the display screen. In some embodiments, the processor 1001 may also include an AI (Artificial Intelligence) processor, which is used to process computing operations related to machine learning.

[0204] The memory 1002 may include one or more computer-readable storage media, which may be non-transitory. The memory 1002 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 some embodiments, the non-transitory computer-readable storage medium in the memory 1002 is used to store a computer program, which is configured to be executed by one or more processors to implement the above-mentioned short review generation method.

[0205] Those skilled in the art will understand that Fig.10 The structure shown in the figure does not constitute a limitation on the computer device 1000, and the computer device 1000 may include more or less components than those shown in the figure, or combine some components, or adopt a different component arrangement.

[0206] In an illustrative embodiment, a computer-readable storage medium is also provided, in which a computer program is stored, and when the computer program is executed by a processor of a computer device, the short review generation method is implemented. Optionally, the computer-readable storage medium may be a ROM (Read-Only Memory), a RAM (Random Access Memory), a CD-ROM (Compact Disc Read-Only Memory), a magnetic tape, a floppy disk, an optical data storage device, etc.

[0207] In an exemplary embodiment, a computer program product is also provided, the computer program product comprising a computer program, the computer program being stored in a computer-readable storage medium. A processor of a computer device reads the computer program from the computer-readable storage medium, and the processor executes the computer program, so that the computer device executes the above-mentioned short review generation method.

[0208] It should be understood that the "multiple" mentioned in this article refers to two or more. "And / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the objects associated before and after are in an "or" relationship. In addition, the step numbers described in this article only illustrate a possible execution sequence between the steps. In some other embodiments, the above steps may not be executed in the order of the numbers, such as two steps with different numbers are executed at the same time, or two steps with different numbers are executed in the opposite order to the diagram. The embodiments of the present application are not limited to this.

[0209] The above description is only an exemplary embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for generating a short review, characterized in that: The method comprises: Acquire multiple comment information of the first multimedia content; Extracting a plurality of target comment information from the plurality of comment information, wherein the target comment information refers to comment information having a recommendation significance for the first multimedia content; Segmenting the plurality of target review information respectively to obtain a plurality of segmented short sentences; Generate a short review recommendation library of the first multimedia content according to the information of the first multimedia content and the plurality of segmented short sentences through a language model, wherein the short review recommendation library includes a plurality of recommended short reviews; Non-unique short reviews from the plurality of recommended short reviews are eliminated, and a final recommended short review of the first multimedia content is selected from the remaining recommended short reviews, wherein the non-unique short review refers to a recommended short review associated with a plurality of other multimedia contents except the first multimedia content.

2. The method according to claim 1, characterized in that The extracting a plurality of target review information from the plurality of review information includes: extracting at least one first comment information from the plurality of comment information according to the vocabulary included in the comment information; extracting at least one second comment information from the plurality of comment information according to the relevance between the comment information and the first multimedia content; The at least one first comment information and the at least one second comment information are filtered to obtain the plurality of target comment information.

3. The method according to claim 2, characterized in that The step of extracting at least one first comment information from the plurality of comment information according to the vocabulary included in the comment information includes: Extracting multiple keywords from the multiple comment information; Extracting at least one first vocabulary from the plurality of keywords according to the part of speech of the keyword, the first vocabulary including keywords whose part of speech is adjective and keywords whose part of speech is noun; Calculate the occurrence frequency corresponding to the at least one first word respectively, and determine the first word whose occurrence frequency is greater than or equal to the first threshold as the second word, wherein the occurrence frequency refers to the proportion of the number of occurrences of the first word in the multiple comment information relative to the total number of occurrences of the multiple keywords in the multiple comment information; The at least one first comment information is obtained according to the number of likes of the comment information corresponding to the second word, where the first comment information is comment information whose number of likes is greater than or equal to a second threshold.

4. The method according to claim 2, characterized in that: The extracting at least one second comment information from the plurality of comment information according to the relevance between the comment information and the first multimedia content comprises: Using a first correlation model, scoring the relevance between the plurality of comment information and the first multimedia content, respectively, to obtain first relevance scores corresponding to the plurality of comment information respectively; Using a second correlation model, scoring the relevance between the plurality of comment information and the information of the first multimedia content, respectively, to obtain second relevance scores corresponding to the plurality of comment information respectively; Using a sentiment analysis model to analyze the favorable opinions of the plurality of review information on the first multimedia content, respectively, to obtain favorable opinion scores corresponding to the plurality of review information respectively; Obtaining comprehensive scores corresponding to the plurality of review information respectively according to the first relevance scores, the second relevance scores, and the favorable review scores respectively corresponding to the plurality of review information; The review information whose comprehensive score is greater than or equal to the third threshold is determined as the second review information.

5. The method according to claim 4, characterized in that The information of the first multimedia content includes N sub-information, each of which is used to indicate the configuration of one dimension of the first multimedia content, and N is a positive integer; The step of using the second correlation model to score the relevance between the plurality of comment information and the information of the first multimedia content to obtain second relevance scores corresponding to the plurality of comment information respectively includes: For each of the plurality of comment information, obtaining the relevance scores corresponding to the N sub-information respectively according to the proportion of the text length of the N sub-information in the comment information; Calculating text similarity between the comment information and content information of the first multimedia content using the second correlation model to obtain a similarity score corresponding to the content information; A second relevance score corresponding to the comment information is obtained according to the relevance scores corresponding to the N sub-information and the similarity score corresponding to the content information.

6. The method according to claim 2, characterized in that The screening of the at least one first comment information and the at least one second comment information to obtain the plurality of target comment information includes: Eliminate the comment information that meets the elimination condition from the at least one first comment information and the at least one second comment information to obtain at least one remaining comment information; Using a first correlation model, scoring the relevance between the at least one remaining comment information and the first multimedia content, respectively, to obtain first relevance scores corresponding to the at least one remaining comment information; The comment information whose first relevance score is less than or equal to a fourth threshold is eliminated from the at least one remaining comment information to obtain the plurality of target comment information.

7. The method according to claim 6, characterized in that The rejection condition includes at least one of the following: The comment information contains preset conflicting words; The comment information is a multi-round dialogue sentence structure; The text length of the comment information is less than or equal to a fifth threshold; The text length of the comment information is greater than or equal to a sixth threshold.

8. The method according to any one of claims 1 to 7, characterized in that: The step of removing non-unique short reviews from the plurality of recommended short reviews and selecting a final recommended short review of the first multimedia content from the remaining recommended short reviews includes: Acquire the number of associated multimedia contents corresponding to the plurality of recommended short reviews, wherein the associated multimedia contents refer to multimedia contents in a database whose first association score with the recommended short reviews is greater than or equal to a seventh threshold; Obtaining, according to the numbers of associated multimedia contents respectively corresponding to the plurality of recommended short reviews, at least one quantile of the number of associated multimedia contents, the quantile being used to indicate a distribution ratio of the number of associated multimedia contents arranged in descending order; Determine at least one recommended short review corresponding to a first quantile among the at least one quantile as a common short review, wherein the first quantile is used to indicate the number of associated multimedia contents with a first proportion ranked high; According to the common short reviews, the non-unique short reviews in the plurality of recommended short reviews are eliminated to obtain the remaining recommended short reviews; A final recommended short review of the first multimedia content is selected from the remaining recommended short reviews.

9. The method according to claim 8, characterized in that The step of removing the non-unique short reviews from the plurality of recommended short reviews according to the general short reviews to obtain the remaining recommended short reviews includes: Calculating semantic similarities between the general short review and the plurality of recommended short reviews respectively; Determine the recommended short reviews whose semantic similarity is greater than or equal to an eighth threshold as the non-unique short reviews; The non-unique short reviews are eliminated from the plurality of recommended short reviews to obtain the remaining recommended short reviews.

10. A short review generating device, characterized in that: The device comprises: A comment acquisition module, used to acquire multiple comment information of the first multimedia content; A comment extraction module, used to extract a plurality of target comment information from the plurality of comment information, wherein the target comment information refers to comment information having a recommendation significance for the first multimedia content; A comment segmentation module, used to segment the plurality of target comment information respectively to obtain a plurality of segmented short sentences; A short review generation module, configured to generate a short review recommendation library of the first multimedia content according to the information of the first multimedia content and the plurality of segmented short sentences through a language model, wherein the short review recommendation library includes a plurality of recommended short reviews; The short review selection module is used to eliminate non-unique short reviews from the multiple recommended short reviews and select a final recommended short review of the first multimedia content from the remaining recommended short reviews, wherein the non-unique short review refers to a recommended short review that is associated with multiple other multimedia contents except the first multimedia content.

11. A computer device, characterized in that: The computer device includes a processor and a memory, wherein a computer program is stored in the memory, and the computer program is loaded and executed by the processor to implement the short review generating method according to any one of claims 1 to 9.

12. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and the computer program is loaded and executed by a processor to implement the short review generating method according to any one of claims 1 to 9.

13. A computer program product, characterized in that The computer program product comprises a computer program, and the computer program is loaded and executed by a processor to implement the short review generating method according to any one of claims 1 to 9.

Images