Comment processing method, comment selection display method, electronic device, and storage medium

Abstract

The application discloses a comment processing method, a comment selection display method, an electronic device and a storage medium. The comment processing method comprises the following steps: obtaining a plurality of first comment sentences by dividing a plurality of comments corresponding to a target object, wherein the target object is content, a product and / or a service provided by an Internet platform; pre-filtering the plurality of first comment sentences to obtain a plurality of second comment sentences; inputting the plurality of second comment sentences into a comment sentence classification sub-model to obtain a plurality of third comment sentences classified; and inputting the plurality of third comment sentences into a comment statistical language sub-model for correlation degree scoring, and selecting a third comment sentence with a correlation degree score lower than a given first threshold value or at least one third comment sentence with the lowest correlation degree score as a comment selection. Such a comment processing scheme can extract a comment selection sentence with beautiful and unique expression from a large number of comments of content, products and services provided by an Internet platform.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to comment processing methods, comment selection display methods, electronic devices, and storage media. Background Technology

[0002] With the development of internet technology, internet platforms have become widely used by users. Users can leave comments on different internet platforms to evaluate the content, products, and services offered by these platforms, such as applications. For example, users can express their listening experience about songs or music videos on music streaming platforms.

[0003] With the explosive growth in the number of comments, users want to quickly understand other users' evaluations of relevant content, products, and services, such as their listening experiences with songs or music videos. Therefore, extracting concise, comprehensive, and elegant comment tips (Tips), such as selected music reviews, can stimulate current user interest, help users choose content they like, and effectively improve their user experience on the internet platform. However, online language is not formal text and may contain inappropriate expressions, as well as negative information such as insults and plagiarism that needs to be removed. Furthermore, many comments may lack elegant language, be unoriginal, or even irrelevant to the topic. These factors make extracting high-quality comments difficult.

[0004] The background description is provided for the purpose of understanding the relevant technologies in this field and is not intended as an admission of prior art. Summary of the Invention

[0005] Therefore, it is necessary to provide a solution for processing comments to address the aforementioned technical problems. This solution can extract a curated selection of comments that are both aesthetically pleasing and unique from a large volume of comments, especially music comments. This allows users of internet platforms to quickly understand other users' evaluations of related products, services, or content, such as their listening experiences with songs or music videos. This can stimulate the user's interest and help them choose content they like, such as songs or music videos. Furthermore, by presenting the curated selection of comments obtained according to this application, such as curated music comments, it can be widely applied to the core functions of content-based platforms such as internet platforms and applications (APPs), including search, recommendation, and ranking functions.

[0006] In a first aspect, a comment processing method is provided, which may include:

[0007] Multiple first comment sentences are obtained by dividing multiple comments corresponding to a target object, wherein the target object is the content, products and / or services provided by the Internet platform;

[0008] The plurality of first comment sentences are pre-filtered to obtain a plurality of filtered second comment sentences;

[0009] The multiple second comment sentences are input into the comment sentence classification sub-model to obtain multiple classified third comment sentences;

[0010] The multiple third comment sentences are input into the comment statistics language sub-model for relevance scoring. The third comment sentences with relevance scores below a given first threshold or at least one third comment sentence with the lowest relevance score are selected as the best comments.

[0011] In this embodiment of the application, the comment processing method is specifically a comment selection and extraction method.

[0012] In some embodiments, the process of dividing multiple comments corresponding to a target object into multiple first comment sentences includes:

[0013] Obscure all original punctuation marks in the aforementioned multiple comments;

[0014] The multiple comments that obscure the original punctuation are input into the punctuation prediction sub-model to obtain the predicted punctuation at the corresponding obscuring position, so as to re-punctuate the multiple comments;

[0015] Based on the punctuation marks at the end of the sentences, the multiple comments after re-punctuation are divided into the multiple first comment sentences.

[0016] In some embodiments, the pre-filtering of the plurality of first comment sentences to obtain a plurality of filtered second comment sentences includes:

[0017] Obtain the attention feature value of each comment corresponding to the plurality of first comment sentences, wherein the attention feature value includes the number of likes, the number of replies, the number of reposts, the number of favorites, the number of reads and / or the number of dislikes;

[0018] Based on the attention feature values, a text summarization algorithm is used to determine the general score of each of the multiple first comment sentences;

[0019] The first comment phrase with a generalization score higher than a given second threshold is selected as the filtered plurality of second comment sentences.

[0020] In some embodiments, determining the general score of each of the plurality of first comment sentences using a text summarization algorithm based on the attention features includes:

[0021] The node copy number of each of the multiple first comment sentences is determined based on the attention feature value of each comment corresponding to the multiple first comment sentences.

[0022] Using the text summarization algorithm, a sentence graph is constructed by taking the first comment sentence as a node and copying the corresponding node according to the number of copies of each node of the first comment sentence. The weighted edges between nodes represent the similarity between the corresponding two nodes.

[0023] The sum of the weights of all weighted edges connected to the nodes in the sentence graph is used to determine the generalization score of the first comment sentence corresponding to each node.

[0024] In some embodiments, determining the generalized score of each of the plurality of first comment sentences using a text summarization algorithm based on the attention features further includes:

[0025] Before determining the general score of the first comment sentence corresponding to the node, remove weighted edges whose weights are less than a given third threshold.

[0026] In some embodiments, determining the generalized score of each of the plurality of first comment sentences using a text summarization algorithm based on the attention features further includes:

[0027] Before determining the general score of the first comment sentence corresponding to the node, the sentence graph is processed using a graph convolutional network and / or a graph attention network.

[0028] In some embodiments, the pre-filtering step may include filtering out the first comment sentence that has more than a predetermined maximum word count threshold, such as more than 20 words.

[0029] In some embodiments, the pre-filtering step may include filtering out the first comment sentence with fewer than a predetermined minimum word count threshold, such as fewer than 3 words.

[0030] In some embodiments, the step of filtering out first comment sentences that are too long or / or too short can be independent of or in combination with other pre-filtering steps, such as pre-filtering steps implemented using text summarization algorithms. For example, the step of filtering out first comment sentences that are too long or / or too short can be performed before the pre-filtering steps implemented using text summarization algorithms.

[0031] In some embodiments, inputting the plurality of third comment sentences into a comment statistics language sub-model for relevance scoring includes:

[0032] For each third comment sentence, segment the bytes to be evaluated sequentially;

[0033] The co-occurrence probability between the currently segmented byte fragment to be evaluated and the previous byte fragments is determined sequentially by the comment statistics language sub-model;

[0034] The relevance score of each third comment sentence is determined based on the co-occurrence probability of all the byte segments to be evaluated in the third comment sentence.

[0035] In some embodiments, the comment statistical language sub-model is an N-Gram-based statistical language model.

[0036] In some embodiments, the comment statistical language sub-model is a statistical language model trained on comment training data that originates from the same source as the plurality of comments.

[0037] In some embodiments, the comment sentence classification sub-model is a binary classification model trained on labeled comment sentence training data that are from the same source as the plurality of comments.

[0038] In some embodiments, the punctuation prediction sub-model is a pre-trained language model trained on training data from articles or essays that are not from the same source as the plurality of comments.

[0039] In some embodiments, the target object is music content, and the comment is a music comment.

[0040] In a second aspect, a method for displaying featured comments is provided, which may include:

[0041] The comment processing method according to any embodiment of this application is used to obtain a selection of comments for the corresponding target object;

[0042] The featured comments are displayed in association with the target object.

[0043] Optionally, the display of featured comments may be in response to a user accessing or retrieving the target object or may be implemented through proactive push notifications, depending on the functionality involved.

[0044] In some embodiments, displaying the featured comments in association with the target object includes:

[0045] In response to a user's search for the target object, the search results area simultaneously displays search results including information about the target object and selected comments.

[0046] In some embodiments, displaying the featured comments in association with the target object includes:

[0047] The system proactively pushes and displays recommended content related to the target object, including information about the target object and selected comments.

[0048] In some embodiments, displaying the featured comments in association with the target object includes:

[0049] A leaderboard containing the target object is displayed, and the target object information and the selected comments are displayed simultaneously in the target object ranking position of the leaderboard.

[0050] In some embodiments, displaying the featured comments in association with the target object includes:

[0051] In response to a user entering the display page of the target object, multiple comments on the target object and a selection of the comments are displayed simultaneously on the display page.

[0052] In a third aspect, a comment processing apparatus is provided, comprising:

[0053] The segmentation unit is configured to divide multiple comments corresponding to a target object into multiple first comment sentences, wherein the target object is content, products and / or services provided by the Internet platform;

[0054] A pre-filtering unit is configured to pre-filter the plurality of first comment sentences to obtain a plurality of filtered second comment sentences;

[0055] The classification unit is configured to input the plurality of second comment sentences into the comment sentence classification sub-model to obtain a plurality of classified third comment sentences;

[0056] The relevance scoring unit is configured to input the plurality of third comment sentences into the comment statistics language sub-model for relevance scoring, and select the third comment sentences with relevance scores below a given first threshold or at least one third comment sentence with the lowest relevance score as the selected comments.

[0057] In a fourth aspect, a comment selection display device is provided, comprising:

[0058] The comment processing unit is configured to obtain a selection of comments for a corresponding target object according to the comment processing method described in any embodiment of this application;

[0059] The display unit is configured to display the selected comments in association with the target object.

[0060] Optionally, the comment processing unit is a comment processing apparatus according to any embodiment of this application.

[0061] In a fifth aspect, an electronic device is provided, comprising: a processor and a memory storing a computer program, the processor being configured to implement the method described in any embodiment of the present application when running the calculator program.

[0062] In a sixth aspect, a storage medium is provided, characterized in that the storage medium stores a computer program configured to be executed to implement the methods described in any embodiment of the present application.

[0063] In this embodiment, multiple first comment sentences are obtained by dividing multiple comments corresponding to the target object; the multiple first comment sentences are pre-filtered to obtain multiple filtered second comment sentences; the multiple second comment sentences are input into a comment sentence classification sub-model to obtain multiple classified third comment sentences; the multiple third comment sentences are input into a comment statistical language sub-model for relevance scoring, and the third comment sentences with relevance scores below a given first threshold or at least one third comment sentence with the lowest relevance score are selected as the best comments, thus realizing the identification of the elegance and uniqueness of comment sentences (short sentences). Currently, some solutions exist that determine whether a language fragment is a complete sentence based on language value scores, or score language fragments to extract language fragments representing comments, or automatically generate summary text that is not verbatim identical to the comment based on the scores. In contrast, the comment processing method of this embodiment, in particular, processes complete sentences (short sentences) through classification and statistical relevance to make them both elegant and unique. As an explanation rather than a limitation, the inventors recognize that, preferably when classifying and statistically analyzing correlations using data from the same source, this embodiment of the application employs a seemingly contradictory but unexpected approach: classifying using positive categorical values ​​(affirmative values) while simultaneously filtering selected short sentences from comments through a negative correlation with the correlation score. The former achieves effective classification through positive categorical values, yielding elegant short sentences and filtering out potentially meaningless, abusive, or plagiarized sentences; the latter, through negative correlation filtering of the correlation score, ensures the selection of more unique short sentences. In other words, this embodiment of the application believes that the lower the value of the statistical language model for comments, the more unique the expression the sentence may contain. This configuration is particularly suitable for extracting selected short sentences from content-based targets, such as music content, like song comments.

[0064] In a further embodiment, the comments are re-punctuated during the step of segmenting multiple comments. Specifically, re-punctuating comments using a punctuation prediction sub-model, such as a pre-trained language model, increases the number of potentially recallable sentences, improving the success rate and quality of extracting comment tips. The inventors recognized that current online language suffers from inconsistent punctuation, frequently exhibiting either misuse or absence of punctuation, making it difficult to extract elegant and unique comment tips. This further embodiment effectively solves the problem of difficulty in extracting comment tips caused by improper punctuation.

[0065] In another embodiment, during the pre-screening step, the attention feature values ​​of each comment corresponding to the plurality of first comment sentences are obtained. Based on the attention feature values, a text summarization algorithm is used to determine the general score of each of the plurality of first comment sentences. Based on the general score, a plurality of filtered second comment sentences are obtained. This effectively improves the efficiency of obtaining subsequent comment tips by fully integrating importance weights into the text summarization algorithm, and also helps to obtain higher quality comment tips. The inventors have realized that comments, such as music comments, contain additional attention information besides the comment text itself, such as the number of likes and replies. This attention information plays a crucial role in judging the quality of comments and characterizing the interests of a large number of users. Currently, text summarization algorithms are commonly used to extract summaries or keywords from comments or to generate summary text based on these. In contrast, this other embodiment achieves superior technical results by incorporating attention information beyond the text into the text summarization algorithm for pre-filtering comment sentences.

[0066] In short, in the embodiments of this application, the sentence beauty and uniqueness recognition method based on pre-trained and statistical language models selects more beautiful sentences while ensuring sentence conciseness and comprehensiveness, making the selected comments highly usable, especially in the core functions of content-based platforms such as recommendation and search. In a further embodiment, re-punctuating the comments solves the problem of low sentence extraction success rate caused by non-standard punctuation, improving the extraction quality of selected comments. In another embodiment, the text summarization algorithm that integrates importance weights, especially comment attention features, adds text importance features compared to traditional text summarization methods, making highly praised or highly replied comments more representative, improving the extraction efficiency and quality of selected comments. This application is intended to introduce some concepts in a simplified form, which will be further described in the following description. This application does not intend to identify the essential features of the claimed technical solutions. Furthermore, the claimed technical solutions are not limited to implementations that address all the shortcomings pointed out in any part of this document.

[0067] Other optional features and technical effects of the embodiments of this application are partly described below and partly apparent from reading this document. Attached Figure Description

[0068] The embodiments of this application will now be described in detail with reference to the accompanying drawings. The elements shown are not limited to the scale shown in the drawings, and the same or similar reference numerals in the drawings denote the same or similar elements, wherein:

[0069] Figure 1 An exemplary flowchart of a comment processing method according to an embodiment of this application is shown;

[0070] Figure 2 An exemplary flowchart of a comment processing method according to an embodiment of this application is shown;

[0071] Figure 3 An exemplary flowchart of a comment processing method according to an embodiment of this application is shown;

[0072] Figure 4 An exemplary flowchart of a comment processing method according to an embodiment of this application is shown;

[0073] Figure 5 An exemplary flowchart of a comment processing method according to an embodiment of this application is shown;

[0074] Figure 6A An exemplary flowchart of a comment processing method according to an embodiment of this application is shown;

[0075] Figure 6B Another exemplary flowchart of the comment processing method according to an embodiment of this application is shown;

[0076] Figure 7 An exemplary flowchart of a comment highlighting display method according to an embodiment of this application is shown;

[0077] Figure 8 A schematic diagram of the structure of a comment processing apparatus according to an embodiment of this application is shown;

[0078] Figure 9 A schematic diagram of the structure of a comment selection display device according to an embodiment of this application is shown;

[0079] Figure 10 A schematic diagram of the structure of an electronic device capable of implementing the methods of the embodiments of this application is shown;

[0080] Figure 11 This illustration shows a schematic diagram of masking and predicting comment punctuation in an embodiment of this application;

[0081] Figure 12 A schematic diagram of a sentence graph obtained using a text summarization algorithm in an embodiment of this application is shown;

[0082] Figure 13 This illustration shows a schematic diagram of determining the co-occurrence probability of byte segments using an N-Gram-based statistical language model in an embodiment of this application;

[0083] Figures 14A to 14D Several examples are shown of displaying a selection of music reviews in a music app. Detailed Implementation

[0084] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to specific embodiments and accompanying drawings. Here, the illustrative embodiments and descriptions of this application are used to explain this application, but are not intended to limit this application.

[0085] The term "comprising" and its variations as used herein signify open inclusion, i.e., "including but not limited to". Unless otherwise stated, the term "or" means "and / or". The term "based on" means "at least partially based on". The terms "one example embodiment" and "one embodiment" mean "at least one example embodiment". The term "another embodiment" means "at least one additional embodiment". The terms "first", "second", etc., may refer to different or the same objects. Other explicit and implicit definitions may also be included below.

[0086] In the embodiments of this application, "sentence" or "phrase" will have the conventional meaning understood in the art. A comment may contain multiple sentences. In the embodiments of this application, "sentence" encompasses "short sentence," and both may be used interchangeably in certain suitable contexts herein. Preferably, a short sentence can be a sentence with relatively few words; for example, in one embodiment, a short sentence can be a sentence with no more than 20 words, preferably no more than 15 words.

[0087] In the embodiments of this application, the selected comments (Tips) refer to short phrases chosen from comments. These may also be referred to herein as selected short phrases or selected sentences; that is, the selection described herein refers to sentences (short phrases), and may sometimes be used interchangeably with selected sentences or selected short phrases. It will be understood that such selected comments often aim to be concise, comprehensive, and elegant, especially for content-based objects such as music content, where uniqueness is also desired.

[0088] Featured comments, such as selected music reviews, can stimulate user interest and help them choose content they like, effectively improving their experience on the internet platform. However, online language is not formal text and may contain inappropriate expressions, as well as negative information such as insults and plagiarism that needs to be removed. Furthermore, many comments may lack elegant language, originality, or even be irrelevant to the topic. These factors make selecting featured comments difficult.

[0089] The comment processing method of this application embodiment, more preferably a comment selection (short sentence) extraction method, involves inputting pre-filtered comment sentences (classified from comments, such as music comments) into a comment sentence classification sub-model to obtain classified comment sentences. These classified comment sentences are then input into a comment statistical language sub-model for relevance scoring. Comment sentences with relevance scores below a given threshold, or one or more comment sentences with the lowest relevance scores, are selected as the featured comments. This method efficiently and effectively extracts elegant and unique short sentences from comments, overcoming current problems.

[0090] The specific embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0091] Figure 1 A comment processing method according to an embodiment of this application is illustrated. More specifically, the comment processing method described in this embodiment is a comment tips extraction method.

[0092] The comment processing method may include steps S110 to S140:

[0093] S110: Multiple first comment sentences are obtained by dividing multiple comments of the corresponding target object.

[0094] In this application embodiment, the target object is the content, products, and / or services provided by an internet platform. In this application embodiment, the internet platform can have various implementation forms, including but not limited to, applications (APPs) implemented on different types of devices and operating systems, and websites presented in a browser. The applications described in this application embodiment can have the normal meaning understood by those skilled in the art, and are sometimes also referred to as software or software applications. The applications described in this application embodiment can be various types of applications, including but not limited to music applications, video / short video / live streaming applications, social applications, shopping applications, or game applications, etc.

[0095] In a preferred embodiment, the internet platform is a music application. In another preferred embodiment, the target object is music content, including but not limited to songs, albums, music videos, instrumental tracks, karaoke tracks, etc., provided on a music platform or in a music application (APP). Accordingly, the comments are music comments, such as song comments.

[0096] In the illustrated example, the processing of user comments on songs will be described using songs as an example, but it is conceivable that the target object described according to the embodiments of this application may be other content, products and / or services provided by Internet platforms.

[0097] As a simplified example, multiple first comment sentences can be obtained by dividing the sentences based on the punctuation marks at the end of the sentences themselves, such as periods, question marks, and exclamation marks.

[0098] As an alternative preferred embodiment, the plurality of first comment sentences can be obtained by retyping the punctuation of the plurality of comments and then dividing them based on the retyped punctuation.

[0099] For example Figure 2 In the illustrated embodiment, step S110 may include steps S111 to S113:

[0100] S111: Mask all original punctuation marks in the multiple comments.

[0101] Here, a uniform mask can be used to cover the original punctuation.

[0102] S112: Input the multiple comments that obscure the original punctuation into the punctuation prediction sub-model to obtain the predicted punctuation at the corresponding obscuring position, so as to re-punctuate the multiple comments.

[0103] S113: Based on the punctuation at the end of the characterizing sentence, divide the multiple comments after re-punctuation into the multiple first comment sentences.

[0104] In this embodiment, punctuation marks can also be referred to as punctuation symbols, and may include various symbols used to indicate pauses and intonation. The punctuation marks may include Chinese punctuation marks, including but not limited to periods, question marks, exclamation marks, commas, pause marks, semicolons, colons, quotation marks, parentheses, dashes, ellipses, hyphens, separators, book titles, proper nouns, emphasis marks, etc. Depending on the language of the comment, the punctuation marks may also include English punctuation marks, including but not limited to periods, question marks, exclamation marks, commas, colons, ellipses, semicolons, hyphens, hyphens, dashes, parentheses (small brackets, square brackets, large brackets), quotation marks (double quotation marks, single quotation marks), abbreviations and possessive symbols, slashes, etc.; it is also conceivable that the punctuation marks may include punctuation marks from other languages.

[0105] In a preferred embodiment, the masked "original punctuation" may include punctuation marks, spaces, and line breaks in the conventional sense. The inventors recognize that although spaces and line breaks are not typically considered punctuation marks, especially with the rise of mobile internet, more and more users may be accustomed to using spaces to indicate pauses or sentence breaks, or using carriage returns to indicate pauses or short sentences. Therefore, in this preferred embodiment, the masked "original punctuation" explicitly includes spaces and line breaks.

[0106] In a preferred embodiment, the masked "original punctuation" can encompass punctuation marks consistent with the comment's language, or it can encompass "non-standard format" punctuation marks inconsistent with the comment's language. For example, in a comment expressed in Chinese, if English punctuation marks, such as a period (.), are used, or vice versa, they will still be masked as "original punctuation." The inventors recognize that with the current prevalence of colloquial language, more and more users may not be able to correctly use punctuation formats, resulting in inconsistencies between punctuation formats and the comment's language.

[0107] In another preferred embodiment, the "predicted punctuation" may include punctuation marks in the conventional sense and empty strings (strings of length 0), but excluding spaces and newlines. The inventors recognized that although empty strings are not generally considered punctuation marks, due to the lack of standardized punctuation in comments, some original punctuation marks in comments may not need to be marked. Therefore, in this preferred embodiment, the "predicted punctuation" explicitly includes empty strings. It will be understood that after the original punctuation is replaced by an empty string, the phrases before and after the original punctuation will be connected.

[0108] In a preferred embodiment, the "predicted punctuation" may cover all punctuation marks. However, it is conceivable that in a simplified embodiment, the "predicted punctuation" may cover specific punctuation mark types, such as only including periods indicating the end of a sentence, commas indicating sentence breaks, and empty strings, which also fall within the scope of this application.

[0109] In a preferred embodiment, the "predicted punctuation" refers to "standard format" punctuation marks (excluding empty strings). For example, for comments expressed in Chinese, only Chinese punctuation marks (excluding empty strings) are predicted and retyped.

[0110] Combination Figure 2 and Figure 11 As shown, in (a), the original punctuation of the exemplary original comment (including regular punctuation 1101, 1101', space 1102, and line break 1103) is uniformly masked (as shown in the image). Figure 11 (The slash is schematically shown) occlusion. In the illustrated embodiment, the Chinese comment contains non-canonical English punctuation marks 1101', spaces 1102, and line breaks 1103. Accordingly, in (b), the punctuation prediction submodel predicts that the predicted punctuation marks (covering regular punctuation marks 1111 and empty strings 1112) will be used to replace the original punctuation marks. Figure 11 In the example shown, the first comment sentence that can be recalled for subsequent use changes from the original two sentences to five sentences with re-punctuation.

[0111] In this embodiment, the punctuation prediction sub-model is a natural language processing (NLP) based model. Preferably, the punctuation prediction sub-model is a pre-trained language model, which may include, but is not limited to, ELMo, GPT, BERT, etc.

[0112] In some embodiments, the punctuation prediction sub-model is a pre-trained language model trained on training data from articles or essays that are not from the same source as the plurality of comments. Here, data from the same source as comments refers to data from or extracted from similar comments; correspondingly, data not from the same source as comments refers to data not originating from similar comments. For example, for music content, such as songs, data from the same source as the plurality of comments can be comment data from other music, such as songs, or data extracted from or processed from such comment data; examples will continue to be provided below. In this preferred embodiment, the data used to train the punctuation prediction sub-model is not any comment data, but rather training data from articles or essays, such as training data obtained from essays by famous writers or data obtained from other articles with relatively standardized expressions.

[0113] In an exemplary embodiment, training a punctuation prediction sub-model using article or prose training data includes: providing a pre-trained language model; providing article or prose training data, wherein punctuation marks in the article or prose training data are occluded; inputting the training data with occluded punctuation marks into the pre-trained language model to be trained, and predicting the occluded punctuation marks; using the occluded punctuation marks as labels, verifying the output of the pre-trained language model and iteratively updating the pre-trained language model until the iteration exit condition (such as convergence) is met.

[0114] S120: Pre-filter the plurality of first comment sentences to obtain a plurality of filtered second comment sentences.

[0115] In this embodiment of the application, the pre-filtering of the comment sentences can be achieved by performing a general score on the second comment sentence. For example, the general score can be obtained by constructing a sentence-level graph network to perform a general ranking of sentence nodes.

[0116] In such Figure 3 In the illustrated embodiment, step S120 may include:

[0117] S121: Obtain the attention feature value of each comment corresponding to the plurality of first comment sentences.

[0118] In this embodiment of the application, the level of attention can reflect the importance of the comment to a certain extent.

[0119] In some embodiments, attention characteristics may include the number of likes, replies, shares, favorites, reads, and / or dislikes. In some embodiments, attention characteristics may include positive (+) attention characteristics, such as the number of likes, replies, shares, favorites, and reads, and may also include negative (-) attention characteristics, such as the number of dislikes.

[0120] In some embodiments, the acquired attention feature values ​​can be one or more, such as the number of likes and replies, and the attention feature values ​​can be smoothed.

[0121] S122: Based on the attention feature value, use a text summarization algorithm to determine the general score of each of the multiple first comment sentences.

[0122] In some embodiments, various text summarization algorithms can be used to incorporate importance weights, such as attention levels, to achieve the generalization score processing of the embodiments of this application. In some embodiments, the text summarization algorithm can be the TextRank algorithm, but other text summarization algorithms are conceived.

[0123] In such Figure 4 In the illustrated embodiment, step S122 may include steps S1221, S1222, and S1225:

[0124] S1221: Determine the number of node copies of each of the multiple first comment sentences based on the attention feature values ​​of their respective comments.

[0125] In some embodiments, the number of node copies can be the attention feature value (or its sum), or the attention feature value can be processed to obtain a number of node copies that are positively correlated with the attention feature value. For example, when the attention feature value includes the number of likes and comments, the sum of the number of likes and comments can be directly used as the number of node copies.

[0126] S1222: Using the text summarization algorithm, construct a sentence graph by taking the first comment sentence as a node and copying the corresponding node according to the number of node copies of each of the first comment sentences, wherein the weighted edges between nodes represent the similarity between the corresponding two nodes.

[0127] S 1225: Determine the generalization score of the first comment sentence corresponding to each node based on the sum of the weights of all weighted edges connected to the nodes of the sentence graph.

[0128] Combination Figure 12 As shown, multiple different first comment sentences are used as nodes 1201, 1202, 1203, and 1204 of sentence graph 1200, or sentence-level graph network, to construct the sentence graph. For example... Figure 12 In the embodiment shown, the sentence graph 1200, or sentence-level graph network, can be a weighted undirected graph.

[0129] At the same time, such as Figure 12It is also schematically shown that node 1201 can also replicate the corresponding node 1201' with its corresponding node copy number (3 copies are schematically represented by slashes). It will be understood that in the embodiments of this application, replicating a node with a node copy number (N) may exclude the original node (N+1 nodes) or include the original node (N nodes), preferably the former.

[0130] like Figure 12 As shown, the weight (W) of the edge 1211 between nodes 1201, 1202, 1203, and 1204, or the edge weight (weight of the edge), represents the similarity between the corresponding two nodes. Figure 12 As shown, the replicated node 1201' also has weighted edges 1211 with other nodes and with other replicated nodes.

[0131] Specifically, the formulas for word bag similarity or edge weight, node weight, and generalization score between two nodes are as follows:

[0132]

[0133]

[0134]

[0135] Wherein, in equation (1) Sim(·,·) represents the similarity between two sentences (nodes), i.e., the edge weight, S i Let |S| represent the i-th sentence. i | represents the number of words in the i-th sentence, w k Let represent the k-th word; d is a hyperparameter representing the probability of pointing to other nodes, set to 0.85 in one instance. In(V i ) and Out(V i ) represents node V i Let W(V) be the set of nodes representing the starting and receiving points. Equation (2) i ) is the weight obtained for node i. Equation (2) is an iterative formula that can be iterated until convergence. Equation (3) Score(i) is the sum of the weights (edge ​​weights) of the weighted edges connected to the node corresponding to sentence i after convergence, which is used as the general score of the sentence (node).

[0136] In the embodiments of this application, multiple nodes corresponding to the same sentence (e.g., duplicate nodes) typically have approximately the same sum of weights. Therefore, when the same sentence corresponds to multiple nodes, any one of the nodes can be selected to determine the summary score of the sentence. Alternatively, the average summary score of multiple nodes can be used as the summary score of the sentence.

[0137] In such Figure 5In the embodiment shown, step S122 may optionally also include S1223: removing weighted edges with weights less than a given third threshold before determining the general score of the first comment sentence corresponding to the node.

[0138] like Figure 12 This schematically illustrates that when the weight of weighted edge 1212 is less than a given third threshold, it will be removed.

[0139] Those skilled in the art will understand that step S1223 can be combined with step S1222 or independent of step S1222. For example, the weight of a weighted edge can be determined during graph construction in step S1222; if the edge does not meet the specified criteria, it will not be included in the constructed graph. Alternatively, the weight of a weighted edge can be determined after the graph has been constructed in step S1222, in order to delete certain weighted edges.

[0140] In such Figure 5 In the illustrated embodiment, step S122 may optionally further include S1224: processing the sentence graph using a graph convolutional network and / or a graph attention network before determining the general score of the first comment sentence corresponding to the node.

[0141] In the graph convolutional network of this embodiment, a sentence graph is taken as input, passes through the hidden layers of multiple intermediate convolutional networks, and the corresponding output graph is obtained. This output graph can be used for subsequent applications, such as for determining a general score. The graph convolutional network can be a network that has been trained iteratively.

[0142] Graph attention networks refer to graph neural networks that incorporate attention mechanisms into spatial domain-based graph neural networks. Similarly, a sentence graph can be used as input, fed into a graph neural network with an attention mechanism, and the resulting output graph can be used for subsequent applications, such as determining a general score. This graph attention network can be an iteratively trained network.

[0143] In some embodiments, steps S1223 and S1224 may be implemented individually or together. In some preferred embodiments, step S1224 may be implemented after step S1223.

[0144] S123: Select the first comment phrase with a general score higher than a given second threshold as the filtered plurality of second comment sentences.

[0145] As Figures 3 to 5As a supplement or alternative to the illustrated embodiment, the pre-filtering step may further include filtering out first comment sentences with a word count greater than a predetermined maximum word count threshold, such as greater than 20 words or greater than 15 words. Thus, all sentences (short sentences) after filtering will have a word count less than or equal to the predetermined word count threshold, thereby satisfying the requirement for conciseness.

[0146] As Figures 3 to 5 As a supplement or alternative to the illustrated embodiment, the pre-filtering step may include filtering out the first comment sentence with a word count less than a predetermined minimum word count threshold, such as less than 3 words. Thus, all sentences after filtering will meet the minimum word count requirement to form sentences (short phrases), rather than words or phrases.

[0147] Accordingly, the step of filtering out first comment sentences that are too long or / or too short can be independent of or in combination with other pre-filtering steps, such as pre-filtering steps implemented using text summarization algorithms. For example, the step of filtering out first comment sentences that are too long or / or too short can be performed before the pre-filtering steps implemented using text summarization algorithms.

[0148] However, in some preferred embodiments, the filtered second comment sentence can meet the requirements of conciseness and minimum text requirements by processing it with the text summarization algorithm according to the embodiments of this application.

[0149] S130: Input the multiple second comment sentences into the comment sentence classification sub-model to obtain multiple classified third comment sentences.

[0150] In some embodiments, the comment sentence classification sub-model includes a binary classification model. This classification model can be a machine learning model or a deep learning model.

[0151] In some embodiments, the comment sentence classification sub-model is a binary classification model trained on labeled comment sentence training data that originates from the same source as the plurality of comments. When the target object of the comment is, for example, music, the labeled comment sentence training data is, for example, sentence data obtained from music reviews, which can be labeled, for example, manually. The binary classification labels include, for example, positive / yes labels (selected by classification) and negative / no labels (rejected by classification). In this embodiment, the positive / yes label indicates that the sentence meets the requirements of elegance and is not a meaningless, abusive, or plagiarized sentence. Alternatively, other classification models may also be used.

[0152] In one example, a binary classification dataset can be obtained by manually annotating multiple (e.g., 8000) short music review phrases, and the binary classification model can be obtained by adjusting it on a pre-trained language model. This pre-trained language model used for the binary classification model can be the one described in this paper or another pre-trained language model.

[0153] S140: Input the plurality of third comment sentences into the comment statistics language sub-model for relevance scoring, and select the third comment sentences with relevance scores lower than a given first threshold or at least one third comment sentence with the lowest relevance score as the selected comments.

[0154] In such Figure 6A In the illustrated embodiment, inputting the plurality of third comment sentences into the comment statistics language sub-model for relevance scoring may include:

[0155] S141: For each third comment sentence, segment the byte fragments to be evaluated sequentially;

[0156] S142: The co-occurrence probability between the currently segmented byte fragment to be evaluated and the previous byte fragments is determined sequentially by the comment statistics language sub-model;

[0157] S143: Determine the relevance score of each third comment sentence based on the co-occurrence probability of all the byte segments to be evaluated in each third comment sentence.

[0158] In some embodiments, the comment statistical language sub-model is an N-gram-based statistical language model. N-gram is an algorithm based on statistical language models, specifically involving a sliding window operation of size n bytes through the text (e.g., the third comment sentence), forming a sequence of byte segments of length n. Depending on the size of N, a unigram (1-gram or unigram), a 2-gram (2-gram or bigram), or a 3-gram (3-gram or trigram) model can be used, such as... Figure 13 A ternary model is used. The byte fragment can include phrases, and the two can be interchanged in certain appropriate positions.

[0159] Combination Figure 6A and Figure 13 As shown, a sliding window 1301 with n=3 slides along an exemplary comment sentence to segment each n=3 byte fragment to be evaluated, and accordingly determines the co-occurrence probability between the current byte fragment / phrase (n=3) and the phrases preceding the sentence. In some alternative embodiments, the first byte fragment / phrase to be evaluated may be padded with several bytes.

[0160] In some embodiments, the co-occurrence probability can be calculated based on chain rules or Markov chains.

[0161] Although the sliding step size is also 3 in the illustrated embodiment, which is equal to the size of the byte fragment, it is conceivable that the sliding compensation is smaller than the size of the byte fragment.

[0162] Although in the illustrated embodiment the byte fragment size, or sliding window, is fixed, such as n=3, it is conceivable that it can be varied, for example, varying between 2, 3, and 4 according to the phrase n.

[0163] Although the statistical language submodel for commenting in the illustrated embodiment is an N-Gram-based statistical language model, other statistical language models are conceivable.

[0164] In some embodiments, the comment statistical language sub-model is a statistical language model trained on comment training data that is from the same source as the plurality of comments. For example, when the comments being processed are music comments, a large number (e.g., 1 million) of music comments (from the same source) can be used to train the statistical language model.

[0165] The statistical language model used in this application embodiment is not used to predict the most likely (highest co-occurrence probability) byte segment (e.g., to predict the next byte segment / phrase) as currently used. Instead, it uses the statistical language model to determine the co-occurrence probability between the currently evaluated byte segment (phrase) in the third comment sentence and previous byte segments. More importantly, unlike the current method of using the statistical language model to sequentially predict the byte segments with the highest probability, in this application embodiment, the correlation determined by the co-occurrence probability of all byte segments is negatively correlated and used to select featured comments. That is, the byte segments (phrase) in the featured comments will have a low overall co-occurrence probability.

[0166] In step S144, the co-occurrence probability of all byte segments to be evaluated for each third comment sentence can be obtained by multiplying (logarithmic sum) the co-occurrence probabilities, thereby determining the relevance score of the third comment sentence. Here, this relevance score can also be referred to as the combined co-occurrence probability.

[0167] In a preferred embodiment, the length of sentences can also be considered when determining the relevance score. The inventors have discovered that by considering the length of sentences, a more effective selection of unique comments can be obtained. For example... Figure 6B In the illustrated embodiment, inputting the plurality of third comment sentences into the comment statistics language sub-model for relevance scoring may include:

[0168] S141: For each third comment sentence, segment the byte fragments to be evaluated sequentially;

[0169] S142: The co-occurrence probability between the currently segmented byte fragment to be evaluated and the previous byte fragments is determined sequentially by the comment statistics language sub-model;

[0170] S143: Determine the baseline co-occurrence probability of each third comment sentence based on the number of words in that sentence;

[0171] S144': Determine the relevance score of the third comment sentence based on the co-occurrence probability of all byte segments to be evaluated and the benchmark combination co-occurrence probability of each third comment sentence.

[0172] Therefore, in step S143, the baseline co-occurrence probability can be obtained based on the number of words in the sentence, which is, for example, the average co-occurrence probability of sentences with that number of words (the product of byte fragment co-occurrence probabilities or the logarithmic sum). Accordingly, in step S144', the relevance score can be determined based on both the co-occurrence probability of all byte fragments to be evaluated and the baseline co-occurrence probability. For example, the relevance score can be determined by dividing the product of byte fragment co-occurrence probabilities by the baseline co-occurrence probability (logarithmic difference).

[0173] Furthermore, such as Figure 7 As shown in the embodiments of this application, a method for displaying featured comments is also provided, which may include:

[0174] S710: Use comment processing methods to obtain a selection of comments for the corresponding target object;

[0175] In the embodiments of this application, the comment processing method is preferably the comment processing method described in any embodiment of this application.

[0176] S720: Display the featured comments in association with the target object.

[0177] Optionally, the display of featured comments may be in response to a user accessing or retrieving the target object or may be implemented through proactive push notifications, depending on the functionality involved.

[0178] In some embodiments, the curated comments obtained by the comment processing method according to this application can be combined with the search function of an internet platform, such as an app. Here, step S720 may include: in response to a user searching for the target object, simultaneously displaying search results including target object information and the curated comments in the search results area. Figure 14A In the user interface 1400 of the music application shown, the search results area simultaneously displays information about the searched song (such as the song title and song image) and a selection of comments for that song.

[0179] In some embodiments, the selected comments obtained by the comment processing method according to this application can be combined with the recommendation function of an internet platform, such as an app. Here, step S720 may include: actively pushing and displaying recommended content related to the target object, the recommended content including target object information and the selected comments. Figure 14B In the user interface 1400 of the music application shown, songs are displayed in the recommendation area, along with song information (such as song title and song image) and a selection of comments for that song.

[0180] In some embodiments, the selected comments obtained by the comment processing method according to this application can be combined with the ranking function of an internet platform, such as an app. Here, step S720 may include: displaying a ranking list containing the target object, and simultaneously displaying the target object information and the selected comments in the target object's ranking position in the ranking list. Figure 14C The user interface 1400 of the music application shown displays a song chart, which can be displayed directly in the ranking position of the corresponding song or by clicking on the song (not shown in the figure) to display the selected comments of the song.

[0181] In some embodiments, the selected comments obtained by the comment processing method according to the embodiments of this application can be combined with or supplemented by comments / comment sections provided for a target object on an internet platform, such as an app. Here, step S720 may include: in response to a user entering the display page of the target object, simultaneously displaying multiple comments of the target object and the selected comments on the display page. Figure 14D The music application's user interface 1400 shown includes a song page, which can be accessed in areas other than the comment section of that song page, such as... Figure 14D The top area (featured comments section) displays selected comments for the song, allowing the current user to quickly understand other users' listening experiences without having to scroll through numerous comments. This can stimulate the current user's interest and help them understand the song. Alternatively, the featured comments could be integrated into the comment section (not shown).

[0182] In this application embodiment, a comment selection and extraction model is also provided, which may include a punctuation prediction sub-model, a text summarization sub-model (or text summarization algorithm module), a comment sentence classification sub-model, and a comment statistical language sub-model.

[0183] In this embodiment of the application, a training method for a comment selection and extraction model is also provided, which may include the steps described in other embodiments for training a punctuation prediction sub-model, a comment sentence classification sub-model, and a comment statistical language sub-model.

[0184] like Figure 8A comment processing apparatus 800, also known as a comment selection and extraction apparatus, is shown according to an embodiment of this application. The comment processing apparatus 800 may include a segmentation unit 810, a pre-filtering unit 820, a classification unit 830, and a relevance scoring unit 840. The segmentation unit 810 is configured to segment multiple comments corresponding to a target object to obtain multiple first comment sentences, where the target object is content, products, and / or services provided by an internet platform. The pre-filtering unit 820 is configured to pre-filter the multiple first comment sentences to obtain multiple filtered second comment sentences. The classification unit 830 is configured to input the multiple second comment sentences into a comment sentence classification sub-model to obtain multiple classified third comment sentences. The relevance scoring unit 840 is configured to input the multiple third comment sentences into a comment statistical language sub-model for relevance scoring, selecting third comment sentences with relevance scores below a given first threshold or at least one third comment sentence with the lowest relevance score as featured comments (Tips).

[0185] In some embodiments, the comment processing device 800 may be installed on a server.

[0186] like Figure 9 A comment selection display apparatus 900 according to an embodiment of this application is shown. The comment selection display apparatus 900 may include a comment processing unit 910 and a display unit 920. The comment processing unit 910 is configured to obtain a selection of comments for a corresponding target object according to the comment processing method described in any embodiment of this application. Preferably, the comment processing unit 910 is a comment processing apparatus 800 according to an embodiment of this application. The display unit 920 is configured to display the comment selection in association with the target object.

[0187] In some embodiments, the display unit 920 may be installed on a client.

[0188] Under the teachings of this application, features of method embodiments may be incorporated in a non-contradictory manner into apparatus, device, or storage medium embodiments or other method embodiments to obtain new embodiments, and features of apparatus embodiments may also be incorporated in a non-contradictory manner into method, device, or storage medium embodiments or other apparatus embodiments to obtain new embodiments, which falls within the scope of this application.

[0189] In some embodiments, an electronic device is provided that may include a processor and a memory storing a computer program, the processor being configured to perform the method of any embodiment of the present application when running the computer program. Figure 10A schematic diagram of an exemplary electronic device 1000 that can implement the methods of embodiments of this application is shown. In some embodiments, it may include more or fewer electronic devices than shown. In some embodiments, it may be implemented using a single or multiple electronic devices. In some embodiments, it may be implemented using cloud-based or distributed electronic devices.

[0190] like Figure 10 As shown, the electronic device 1000 includes a processor 1001, which can perform various appropriate operations and processes based on programs and / or data stored in read-only memory (ROM) 1002 or programs and / or data loaded from storage portion 1008 into random access memory (RAM) 1003. The processor 1001 can be a single-core or multi-core processor, or may contain multiple processors. In some embodiments, the processor 1001 may include a general-purpose main processor (such as a CPU) and one or more special coprocessors, such as a graphics processing unit (GPU), a neural network processor (NPU), a digital signal processor (DSP), or other general-purpose or application-specific integrated circuits. Various programs and data required for the operation of the electronic device 1000 are also stored in RAM 1003. The processor 1001, ROM 1002, and RAM 1003 are interconnected via a bus 1004. An input / output (I / O) interface 1005 is also connected to the bus 1004.

[0191] The processor and memory described above are used together to execute the program stored in the memory. When the program is executed by the computer, it can implement the steps or functions of the model generation method and recognition method described in the above embodiments.

[0192] The following components are connected to I / O interface 1005: an input section 1006 including a keyboard, mouse, etc.; an output section 1007 including a display and speakers, etc.; a storage section 1008 including a hard disk, etc.; and a communication section 1009 including a network interface card such as a LAN card and a modem, etc. The communication section 1009 performs communication processing via a network such as the Internet. Drive 1010 is also connected to I / O interface 1005 as needed. Removable media 1011, such as a hard disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 1010 as needed so that computer programs read from them can be installed into storage section 1008 as needed.

[0193] Figure 10 The electronic device shown is merely illustrative, but the electronic device according to embodiments of this application may include more than [other components]. Figure 10 The electronic device shown has more or fewer components or has more or fewer components than the one shown. Figure 10 The embodiments shown have the same, partially the same, or different architectures.

[0194] Although not shown, some embodiments also provide a computer-readable storage medium storing a computer program configured to be executed to perform the methods of any of the embodiments of this application. The computer program includes various program modules / units constituting the apparatus according to the embodiments of this application, and when executed, the computer program comprised of the various program modules / units can perform the functions corresponding to the various steps in the methods described in the above embodiments. The computer program can also run on electronic devices as described in the embodiments of this application.

[0195] The storage medium in embodiments of this application includes non-volatile and / or volatile articles that can store information by any method or technology. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0196] Those skilled in the art will understand that the embodiments of this specification can be implemented in various forms, such as methods, systems, or computer program products. Therefore, those skilled in the art will realize that the functional modules / units or controllers and related method steps described in the above embodiments can be implemented in software, hardware, or a combination of software and hardware.

[0197] Unless explicitly stated otherwise, the actions or steps of the methods and procedures described in the embodiments of this application do not necessarily have to be performed in a specific order and can still achieve the desired results. In some implementations, multitasking and parallel processing are also possible or may be advantageous.

[0198] This document describes several embodiments, but for the sake of brevity, the descriptions of the embodiments are not exhaustive, and identical or similar features or parts between the embodiments may be omitted. In this document, "one embodiment," "some embodiments," "example," "specific example," or "some examples" refers to at least one embodiment or example applicable to this application, but not all embodiments. The above terms do not necessarily mean referring to the same embodiment or example. Without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of the different embodiments or examples.

[0199] The exemplary systems and methods of this application have been specifically shown and described with reference to the above embodiments, which are merely examples of the best mode for implementing the systems and methods. Those skilled in the art will understand that various changes can be made to the embodiments of the systems and methods described herein without departing from the spirit and scope of this application as defined in the appended claims when implementing the systems and / or methods.

Claims

1. A comment processing method, characterized in that, include: Multiple first comment sentences are obtained by dividing multiple comments corresponding to a target object, wherein the target object is the content, products and / or services provided by the Internet platform; The plurality of first comment sentences are pre-filtered to obtain a plurality of filtered second comment sentences; The multiple second comment sentences are input into the comment sentence classification sub-model to obtain multiple classified third comment sentences, wherein the comment sentence classification sub-model is a classification sub-model based on elegance; The plurality of third comment sentences are input into the comment statistics language sub-model for relevance scoring. Third comment sentences with relevance scores below a given first threshold or at least one third comment sentence with the lowest relevance score are selected as the featured comments. The relevance score is determined based on the co-occurrence probability between byte segments of the third comment sentences.

2. The comment processing method according to claim 1, characterized in that, The process of dividing multiple comments from the corresponding target object to obtain multiple first comment sentences includes: Obscure all original punctuation marks in the aforementioned multiple comments; The multiple comments that obscure the original punctuation are input into the punctuation prediction sub-model to obtain the predicted punctuation at the corresponding obscuring position, so as to re-punctuate the multiple comments; Based on the punctuation marks at the end of the sentences, the multiple comments after re-punctuation are divided into the multiple first comment sentences.

3. The comment processing method according to claim 1, characterized in that, The pre-filtering of the plurality of first comment sentences to obtain a plurality of filtered second comment sentences includes: Obtain the attention feature value of each comment corresponding to the plurality of first comment sentences, wherein the attention feature value includes the number of likes, the number of replies, the number of reposts, the number of favorites, the number of reads and / or the number of dislikes; Based on the attention feature values, a text summarization algorithm is used to determine the general score of each of the multiple first comment sentences; The first comment sentence with a general score higher than a given second threshold is selected as the filtered plurality of second comment sentences.

4. The comment processing method according to claim 3, characterized in that, The step of determining the general score of each of the multiple first comment sentences based on the attention features using a text summarization algorithm includes: The node copy number of each of the multiple first comment sentences is determined based on the attention feature value of each comment corresponding to the multiple first comment sentences. Using the text summarization algorithm, a sentence graph is constructed by taking the first comment sentence as a node and copying the corresponding node according to the number of copies of each node of the first comment sentence. The weighted edges between nodes represent the similarity between the corresponding two nodes. The sum of the weights of all weighted edges connected to the nodes in the sentence graph is used to determine the generalization score of the first comment sentence corresponding to each node.

5. The comment processing method according to claim 4, characterized in that, The step of determining the general score of each of the multiple first comment sentences using a text summarization algorithm based on the attention features also includes: Before determining the general score of the first comment sentence corresponding to the node, remove weighted edges in the sentence graph whose weights are less than a given third threshold.

6. The comment processing method according to claim 4 or 5, characterized in that, The step of determining the general score of each of the multiple first comment sentences using a text summarization algorithm based on the attention features also includes: Before determining the general score of the first comment sentence corresponding to the node, the sentence graph is processed using a graph convolutional network and / or a graph attention network.

7. The comment processing method according to claim 1, characterized in that, The step of inputting the multiple third comment sentences into the comment statistics language sub-model for relevance scoring includes: For each third comment sentence, segment the bytes to be evaluated sequentially; The co-occurrence probability between the currently segmented byte fragment to be evaluated and the previous byte fragments is determined sequentially by the comment statistics language sub-model; The relevance score of each third comment sentence is determined based on the co-occurrence probability of all the byte segments to be evaluated in the third comment sentence.

8. The comment processing method according to claim 7, characterized in that, The comment statistical language sub-model is an N-Gram-based statistical language model.

9. The comment processing method according to claim 1, 7, or 8, characterized in that, The comment statistical language sub-model is a statistical language model trained on comment training data that shares the same source as the multiple comments.

10. The comment processing method according to claim 1, characterized in that, The comment sentence classification sub-model is a binary classification model trained on labeled comment sentence training data that are from the same source as the multiple comments.

11. The comment processing method according to claim 2, characterized in that, The punctuation prediction sub-model is a pre-trained language model trained on training data from articles or essays that are not from the same source as the multiple comments.

12. The comment processing method according to claim 1, characterized in that, The target object is music content, and the comment is a music review.

13. A method for displaying selected comments, characterized in that, include: The comment processing method according to any one of claims 1 to 12 is used to obtain a selection of comments for the corresponding target object; The featured comments are displayed in association with the target object.

14. The method for displaying selected comments according to claim 13, characterized in that, The display of the featured comments in association with the target object includes: In response to a user's search for the target object, the search results area simultaneously displays search results including information about the target object and selected comments.

15. The method for displaying selected comments according to claim 13, characterized in that, The display of the featured comments in association with the target object includes: The system proactively pushes and displays recommended content related to the target object, including information about the target object and selected comments.

16. The method for displaying selected comments according to claim 13, characterized in that, The display of the featured comments in association with the target object includes: A leaderboard containing the target object is displayed, and the target object information and the selected comments are displayed simultaneously in the target object ranking position of the leaderboard.

17. The method for displaying selected comments according to claim 13, characterized in that, The display of the featured comments in association with the target object includes: In response to a user entering the display page of the target object, multiple comments on the target object and a selection of the comments are displayed simultaneously on the display page.

18. An electronic device, characterized in that, include: A processor and a memory storing a computer program, the processor being configured to implement the method according to any one of claims 1 to 17 when running the computer program.

19. A storage medium, characterized in that, The storage medium stores a computer program configured to be executed to implement the method according to any one of claims 1 to 17.

Images