Analysis method for analyzing post according to comment backtracking

Abstract

The invention relates to an analysis method for backtracking and analyzing a post according to comments, and the method comprises the following steps: 1, a data collection unit obtains the comment data of the post from a social media platform, and carries out the preprocessing of the post; step 2, a feature extraction unit performs multi-dimensional comprehensive feature extraction on the preprocessed post comment data, and evaluates a comprehensive feature weight; 3, a backtracking analysis unit comprehensively evaluates the post quality based on the high-weight comment set, and establishes a quantitative mapping relation from comment feedback to the post quality; step 4, a negative recognition unit adopts a hierarchical and multi-dimensional detection mechanism to comprehensively recognize and trace negative contents in the comments from three levels of a vocabulary layer, a semantic layer and a context layer; and step 5, a verification unit performs multi-dimensional verification and credibility evaluation on the results of the comment backtracking analysis and the negative identification, and generates corresponding differential output according to the combination of the credibility grade and the quality score.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of post analysis methods, more specifically, to an analysis method for analyzing posts based on comments. BACKGROUND

[0002] Existing analysis techniques for analyzing posts based on comments mainly include the following aspects: Text sequence modeling technology based on neural networks: using deep learning models such as recurrent neural networks (RNN) and long short-term memory networks (LSTM) to extract features and classify emotions from text content, this technology builds a multi-layer neural network structure to capture the context dependence and semantic information of the text, when processing social media comment data, the system first performs word segmentation and vectorization on the text, then gradually extracts text features through neural network layers, and finally outputs the emotion classification result, this method performs well in single text sentiment recognition tasks, with a classification accuracy of 82% to 85%, it can identify positive, negative or neutral sentiment tendencies of comments, this technology mainly focuses on independent analysis of single comments or posts, by learning the sentiment patterns in a large amount of labeled data, it establishes the mapping relationship between text features and emotion labels.

[0003] Negative information recognition technology based on content features: using keyword matching and rule template methods to detect negative information in social media content, this technology builds a negative vocabulary library and a sensitive vocabulary library, uses string matching algorithms to scan and identify text, the system predefines a dictionary resource containing negative vocabulary, sensitive topics, and rule expressions, when it detects that the text contains these words, it evaluates the negative degree according to the number and weight of the hit words, this method uses a fixed threshold judgment mechanism, when the negative features exceed the preset standard, it triggers an early warning, in actual application, the recognition accuracy of this technology is between 65% and 72%, it can quickly find obvious negative expressions and harmful content, this technology mainly relies on surface features at the vocabulary level, and realizes automatic identification and filtering of negative content through pattern matching.

[0004] However, as the application deepens, people have found that the existing analysis techniques for analyzing posts based on comments have the following shortcomings: 1. The lack of comment backtracking analysis capability and the lack of correlation modeling: existing text sequence modeling techniques mainly analyze single text independently, without establishing a correlation analysis mechanism between posts and their comments. In actual social media scenarios, the quality and characteristics of posts often need to be evaluated comprehensively through their comments, but existing technologies lack the ability to backtrack and analyze post quality based on comment data. The system uses equal weight processing for all comments, without considering the influence of commenters, the quality of comments, the time effect of publishing, and other important factors. The simple analysis method without weighting leads to insufficient reliability and accuracy of the evaluation results, and cannot accurately reflect the true quality of the post. The technology does not establish a causal reasoning model between comments and posts, and can only identify statistical correlations, but cannot distinguish between false correlations and real causal relationships. Traditional correlation analysis methods have a high causal identification error rate, and the method lacks a counterfactual verification mechanism and a confidence quantification system, which cannot effectively evaluate the reliability of the analysis results, resulting in poor consistency of the analysis results in different scenarios.

[0005] 2. Lack of depth recognition of negative content and lack of risk tracing capability: existing negative information recognition techniques mainly rely on keyword matching and surface feature analysis, and can only recognize obvious negative words, but cannot capture implicit negative expressions such as metaphors, sarcasm, and irony. The fixed threshold discrimination mechanism lacks context understanding ability, and simply triggers an early warning when the number of negative words exceeds the preset value, resulting in a large number of false positives and false negatives. More importantly, existing technologies do not establish a negative content tracing mechanism based on comment data, and cannot locate the true source of post negative attributes by analyzing comment feedback. Experiments show that traditional methods cannot effectively distinguish between posts with problems or comment-induced negative emotions. The technology lacks a multi-dimensional risk assessment system, only considers a single dimension of content harm, and does not comprehensively assess multiple risk factors such as transmission risk, social impact, and authenticity, resulting in insufficient comprehensiveness and accuracy of risk warnings.

[0006] Therefore, it is necessary to propose an analysis method for backtracking and analyzing posts based on comments to solve the above problems. SUMMARY

[0007] The present application provides an analysis method for backtracking and analyzing posts based on comments to overcome at least one of the deficiencies of the prior art.

[0008] To solve the above technical problems, the technical solution of the present application is as follows: an analysis method for backtracking and analyzing posts based on comments, comprising the following steps: Step 1: a data acquisition unit acquires comment data of a post from a social media platform and pre-processes the post; Step 2: a feature extraction unit performs multi-dimensional comprehensive feature extraction on the pre-processed post comment data and evaluates the comprehensive feature weights. Step 3, the backtracking analysis unit conducts a comprehensive evaluation of the post quality based on the high-weight review set, and establishes a quantitative mapping relationship from review feedback to post quality; Step 4, the negative recognition unit adopts a hierarchical and multi-dimensional detection mechanism to comprehensively identify and trace the negative content in the review from three levels of vocabulary layer, semantic layer and context layer; Step 5, the verification unit conducts multi-dimensional verification and credibility evaluation on the results of review backtracking analysis and negative recognition, and generates corresponding differentiated output according to the combination of credibility level and quality score; Further, the step 1 includes the following steps: Step 11, the data acquisition unit obtains the comment data of the post from the social media platform, and the preprocessing unit uses the local sensitive hashing (LSH) algorithm to identify duplicate reviews on the original comment data collected by the data acquisition unit; Step 12, a hybrid method based on rules and machine learning is used to identify spam reviews; Step 13, the encoding and full / half-width of the comment text are unified, and the conversion of the comment text between traditional and simplified Chinese and the normalization processing of special characters are performed; Step 14, the metadata of the comment text is checked for integrity and completed, including the publishing time, user information and comment level; Step 15, an explicit corresponding relationship record is established for each comment and post.

[0009] Further, in the step 12, the spam reviews are advertisements and reviews without actual meaning or content.

[0010] Further, in the step 2, the multi-dimensional comprehensive features include user influence dimension, review quality dimension, interaction feedback dimension, time effect dimension and user behavior credibility dimension; The user influence dimension includes fan size, account authentication level, account activity and historical influence; The review quality dimension includes text length, sentence complexity, information richness, logical coherence, professionalism and emotional tendency; The interaction feedback dimension includes the number of likes, replies, forwards and comprehensive interaction score; The time effect dimension includes the delay in publishing, the position in the post life cycle, and the comment flow density; The user behavior credibility dimension includes comment frequency, domain expertise, historical accuracy and credibility score.

[0011] Further, in the step 2, the comprehensive feature weight is evaluated by the following formula: The comprehensive feature weight = w1 x F (user influence) + w2 x F (review quality) + w3 x F (time effect) + w4 x F (interaction feedback) + w5 x F (user behavior credibility); Wherein the mapping function of each dimension is: F (user influence) = [log (fan size + 1) / 20 + account authentication level / 5 + account activity + historical influence] / 4, wherein, in the account authentication level, ordinary account =1, personal authentication =3, enterprise authentication =4, professional authentication =5; F (review quality) = [min (text length / 300) x0.3 + sentence complexity x0.2 + information richness x0.3 + logical coherence x0.1 + professionalism x0.1]; F (time effect) = exp (-release delay / decay half-time); F (interaction feedback) = comprehensive interaction score; F (credibility) = credibility score.

[0012] Further, the step 3 comprises the following steps: Step 31, the backtracking analysis unit calculates the initial score based on the post text itself characteristics, and obtains the initial score result: The initial score result = w_author x publisher historical score + w_length x content length + w_structure x structure integrity + w_time x release time effect, wherein the weight is set as w_author=0.4, w_length=0.3, w_structure=0.2, w_time=0.1; Step 32, the sentiment distribution score of the high weight review set is scored, and the sentiment distribution score of the high weight review set is (positive comment proportion-negative comment proportion) x comment intensity factor, wherein the evidence range of the comment intensity factor is 0-1.0; Step 33, the keyword feature score is scored, and the keyword score =Σ (keyword weight x keyword sentiment polarity) / total keyword quantity, wherein the value range of the sentiment polarity of the keyword is-1 to 1; Step 34, the support degree of the review to the post quality is calculated, and the driven quality score adjustment is fed back, the adjustment range is about ±0.5 points, and the post quality score is ensured to be between 1.0-5.0; Step 35, the credibility of the post quality score is evaluated, and the credibility score = (comment number density x0.3) + (weight distribution balance degree x0.3) + (comment coverage time span x0.4); Step 36, according to the credibility score, the final post quality score is scored; Step 37, the verification unit verifies the score stability by simulating the change of key factors, removes the high-weight comments ranked in the top 5% of the weight, recalculates the sentiment distribution score and the final post quality score of the high-weight comment set, and observes the score change amplitude. If the score change amplitude is < 0.15, it means that the score is not easily affected by a single comment and has high stability. Step 38, the scale stability is verified, assuming that the number of comments is reduced or increased by 50%, the overall score is recalculated using weight compensation, and the score change amplitude is observed. If the score change amplitude is < 0.2, it means that the score has good robustness. Step 39, the time window is verified, only using the comments within 24 hours after the post is published for quality evaluation, and comparing with the result of using all comments. If the comparison result is < 0.1, it means that early comments can fully reflect the post quality.

[0013] Further, the step 4 includes the following steps: Step 41, the negative recognition unit performs word segmentation processing on the comment text, searches in the negative vocabulary library using an exact string matching algorithm, identifies the sentiment state and specific meaning and sensitivity of the vocabulary from the vocabulary layer, uses an edit distance algorithm for fuzzy matching of transformed vocabulary, and calculates the vocabulary layer score according to the number, weight, and position distribution of the hit vocabulary; Step 42, the semantic layer recognition mechanism identifies sarcasm and irony, detects semantic reversal to obtain a sarcasm score, identifies implicit negative expressions to identify the actual negative meaning of surface neutrality, uses semantic similarity calculation to determine the implicit meaning using the distance of word vectors in semantic space to obtain an implicit negative score, uses a multi-layer neural network model for fine-grained sentiment judgment, scores the overall sentiment polarity of each sentence through natural language processing technology to obtain an explicit negative score, and integrates the semantic layer scores, semantic layer score = sarcasm score × 0.35 + implicit negative score × 0.35 + explicit negative score × 0.30; Step 43, based on the vocabulary layer and the semantic layer, the user's information is introduced for comprehensive judgment from the context layer, and based on the three-layer negative recognition results, the system establishes a multi-dimensional risk assessment framework to evaluate the overall risk degree of the content, social influence evaluation, transmission risk evaluation, credibility, and comprehensive risk score for evaluation; Step 44, the traceability unit traces the source of negative content and locates the post problem after completing negative recognition.

[0014] Further, the step 5 includes the following steps: Step 51, the verification unit uses the counterfactual verification method to test the stability of the analysis result, and observes the change range of the score by simulating the change of the key factor; Step 52, the stability of the weight is verified, the high-weight comments ranked in the top 5% of the weight are removed, and if the score change is <0.15, it indicates that the analysis result is stable; Step 53, the stability of the scale is verified: assuming that the number of comments changes by 50%, if the score change is <0.2, it indicates that the method has good robustness; Step 54, the time window verification: using the comment data in the previous 24 hours for analysis, if the score change is <0.1, it indicates that the early comments have fully reflected the quality of the post; Step 55, the results of the comment backtracking analysis and negative identification are classified according to the credibility level, and according to the combination of the credibility level and the quality score, the corresponding differentiated output is generated.

[0015] Compared with the prior art, the beneficial effects of the technical scheme of the present application are: The analysis method for analyzing posts according to comment backtracking disclosed in the present application significantly improves the accuracy of quality evaluation of posts through the backtracking analysis mechanism based on comment data, significantly improves the quality of feature representation through the self-adaptive five-dimensional feature fusion mechanism, realizes accurate evaluation of the importance of comments, and quickly and effectively evaluates the quality of posts through the simple quality evaluation model (initial score + comment feedback adjustment) and counterfactual verification mechanism. Through the multi-level negative content identification and tracing mechanism, the system can accurately locate the content elements that cause negative emotions in the post, improve the effectiveness of risk warning, and through the credibility level division and differentiated output mechanism, the system generates targeted reports and suggestions according to the credibility and quality score of the analysis result, providing clear credibility reference and operation guidance for different application scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 is the flowchart of the analysis method for analyzing posts according to comment backtracking in the present application. DETAILED DESCRIPTION

[0017] The drawings are only used for illustrative description and cannot be understood as a limitation of the present patent; in order to better illustrate the present embodiment, some components in the drawings may be omitted, enlarged or reduced, and do not represent the size of the actual product; for those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0018] In the description of the present application, it should be noted that unless otherwise expressly specified and limited, the terms "mounting", "connecting" should be understood broadly, for example, it can be fixed connection, or detachable connection, or integrally connected; it can be mechanical connection, or electrical connection; it can be directly connected, or indirectly connected through intermediate medium, or said to be connected inside two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances. The technical solutions of the present application will be further described below in combination with the drawings and examples.

[0019] As shown in Figure 1 An analysis method for analyzing post based on comments, comprising the following steps: Step 1, the data acquisition unit obtains the comment data of the post from the social media platform, and pre-processes the post; Step 2, the feature extraction unit performs multi-dimensional comprehensive feature extraction on the pre-processed post comment data, and evaluates the comprehensive feature weight; Step 3, the backtracking analysis unit comprehensively evaluates the post quality based on the high-weight comment set, and establishes a quantitative mapping relationship from comment feedback to post quality; Step 4, the negative recognition unit adopts a hierarchical and multi-dimensional detection mechanism to comprehensively identify and trace the negative content in the comments from three levels of vocabulary layer, semantic layer and context layer; Step 5, the verification unit performs multi-dimensional verification and credibility evaluation on the results of comment backtracking analysis and negative recognition, and generates corresponding differentiated output according to the combination of credibility level and quality score; In the present application, step 1 comprises the following steps: Step 11, the data acquisition unit obtains the comment data of the post from the social media platform, and the preprocessing unit uses the local sensitive hash (LSH) algorithm to identify duplicate comments for the original comment data collected by the data acquisition unit; Step 12, using a hybrid method based on rules and machine learning to identify spam comments; Step 13, unifying the encoding and full / half angle of the comment text, and converting the full / simplified Chinese of the comment text and normalizing the special characters; Step 14, integrity check and completion of the metadata of the comment text, the metadata including publishing time, user information and comment level; Step 15, establishing an explicit corresponding relationship record for each comment and post.

[0020] In step 12, the spam comments are advertisements and comments without actual meaning or content.

[0021] In step 2, the multi-dimensional comprehensive features include user influence dimension, comment quality dimension, interaction feedback dimension, time effect dimension, and user behavior credibility dimension; The user influence dimension includes fan size, account authentication level, account activity, and historical influence; The comment quality dimension includes text length, sentence complexity, information richness, logical coherence, professionalism, and emotional tendency; The interaction feedback dimension includes likes, replies, forwards, and comprehensive interaction score; The time effect dimension includes publication delay, position in post life cycle, and comment flow density; The user behavior credibility dimension includes comment frequency, domain expertise, historical accuracy, and credibility score.

[0022] In step 2, the comprehensive feature weight is evaluated by the following formula: Comprehensive feature weight = w1 × F(user influence) + w2 × F(comment quality) + w3 × F(time effect) + w4 × F(interaction feedback) + w5 × F(user behavior credibility); Where the mapping function of each dimension is: - F(user influence) = [log(fan size + 1) / 20 + account authentication level / 5 + account activity + historical influence] / 4, where in the account authentication level, ordinary account = 1, personal authentication = 3, enterprise authentication = 4, professional authentication = 5; - F(comment quality) = [min(text length / 300) × 0.3 + sentence complexity × 0.2 + information richness × 0.3 + logical coherence × 0.1 + professionalism × 0.1]; - F(time effect) = exp(-publication delay / decay half-time); - F(interaction feedback) = comprehensive interaction score; - F(credibility) = credibility score.

[0023] In step 3, the following steps are included: Step 31, the backtracking analysis unit calculates the initial score based on the post features, obtaining the initial score result: Initial score result = w_author × publisher historical score + w_length × content length + w_structure × structure integrity + w_time × publication time effect, where the weights are set as: w_author = 0.4, w_length = 0.3, w_structure = 0.2, w_time = 0.1; Step 32, score the sentiment distribution of the high-weight comment set, sentiment distribution score of high-weight comment set = (positive comment proportion - negative comment proportion) x comment intensity factor, wherein the value range of comment intensity factor is 0-1.0; Step 33, score the keyword feature, keyword score = Σ (keyword weight x keyword sentiment polarity) / total keyword quantity, wherein the value range of keyword sentiment polarity is -1 to 1; Step 34, calculate the support degree of comments to post quality, and feedback driven quality score adjustment, the adjustment range is about ±0.5 points, ensuring that the post quality score is between 1.0-5.0; Step 35, evaluate the credibility of post quality score, credibility score = (comment number density x 0.3) + (weight distribution balance x 0.3) + (comment coverage time span x 0.4); Step 36, score the final post quality according to the credibility score; Step 37, the verification unit verifies the stability of the score by simulating the change of key factors, removes the top 5% high-weight comments, recalculates the sentiment distribution score of the high-weight comment set and the final post quality score, and observes the score change amplitude. If the score change amplitude < 0.15, it means that the score is not easily affected by a single comment and has high stability; Step 38, verify the size stability, assume that the number of comments is reduced or increased by 50%, recalculate the overall score using weight compensation, observe the score change amplitude, if the score change amplitude < 0.2, it means that the score has good robustness; Step 39, verify the time window, only use the comments within 24 hours after the post is published for quality evaluation, compare with the result using all comments, if the comparison result < 0.1, it means that early comments can fully reflect the post quality.

[0024] Step 4 includes the following steps: Step 41, the negative recognition unit performs word segmentation processing on the comment text, uses accurate string matching algorithm to search in the negative vocabulary library, identifies the sentiment state and specific meaning and sensitivity of the vocabulary from the vocabulary layer, uses edit distance algorithm for fuzzy matching for variant vocabulary, and calculates the vocabulary layer score according to the number, weight and position distribution of hit vocabulary; Step 42, the semantic layer recognition mechanism identifies sarcasm and irony, detects semantic reversal, obtains a sarcastic score, identifies implicit negative expression mechanisms, identifies surface neutrality but actual negative meaning, uses semantic similarity calculation, uses word vectors to judge implicit meaning in semantic space, obtains implicit negative score, uses a multi-layer neural network model for fine-grained sentiment judgment, scores the overall sentiment polarity of each sentence through natural language processing technology, obtains explicit negative score, and integrates semantic layer scores. Semantic layer score = sarcastic score x 0.35 + implicit negative score x 0.35 + explicit negative score x 0.30; Step 43, based on the vocabulary layer and the semantic layer, the user's information is introduced, the comprehensive judgment is made from the context layer, and based on the three-layer negative recognition results, the system establishes a multi-dimensional risk assessment framework to assess the overall risk degree of the content, the social influence assessment, the transmission risk assessment, the credibility and the comprehensive risk score. Step 44, after completing the negative recognition, the source tracing unit traces the source of the negative content and locates the post problem.

[0025] Step 5 includes the following steps: Step 51, the verification unit uses counterfactual verification method to test the stability of the analysis results, and observes the change amplitude of the score by simulating the change of key factors; Step 52, verify the stability of the weight, remove the top 5% high-weight comments, if the score change < 0.15, the analysis result is stable; Step 53, verify the stability of the scale: assume that the number of comments changes by 50%, if the score change < 0.2, the method has good robustness; Step 54, time window verification: use the comment data in the previous 24 hours for analysis, if the score change < 0.1, the early comments have fully reflected the quality of the post; Step 55, divide the credibility level of the comment backtracking analysis and negative recognition results, and generate corresponding differentiated output according to the combination of credibility level and quality score.

[0026] Embodiment In this embodiment, a large number of comments are needed for a popular post published in a short time. The post was published on April 15, 2024 at 10:00, with the title "New product release evaluation", the publisher is a certified enterprise account (certification level 4), and the number of fans is 500,000. The post content is about 800 words, with 3 product pictures and 1 product demonstration video, and a total of 2547 comments were obtained within 2 hours of publication, with a total of 68,000 likes and 23,000 forwards.

[0027] The first step is to collect and preprocess the comment data. The collection and cleaning process involves the system collecting 2,547 comments for the post. According to the collection strategy, the data is collected every 10 minutes during the 12 hours when the post is popular, every hour for the next 24 hours, and then every 6 hours thereafter.

[0028] The specific cleaning process is as follows: Step 1 - Duplicate and Spam Comment Removal: Using the LSH algorithm (similarity threshold 0.95), 312 duplicate comments were identified and deleted (12.3% of the total). A hybrid method (rule base + SVM classifier) ​​was used to identify spam comments, detecting 156 advertising comments and 28 meaningless comments (such as "111", "like", etc.), with a spam detection rate of 12.5%. After processing, 2051 valid comments remained. Step 2 - Format Standardization: All 2051 comments underwent unified encoding, simplified / traditional character conversion (unified to simplified characters), full-width / half-width character unification, and special characters standardization. For example, the original comment "This product is good," was transformed into "This product is nice." Step 3 - Metadata Integrity Check: Key fields of each comment were checked. Eighteen comments lacking user avatar URLs were filled in using user ID lookup; two comments lacking posting times were filled in using system-collected timestamps. User authentication status and follower counts for all comments were successfully retrieved. Step 4 - Establishing the correspondence between comments and posts: Each comment was assigned a unique ID (format: post ID_comment number), and the following information was labeled: post ID, comment posting timestamp, commenter ID, commenter follower count, commenter authentication level, original comment content, cleaned comment content, and comment level (level 1 comment, level 2 reply, etc.). There were 1534 level 1 comments and 517 level 2 replies. Preprocessing results statistics: Data retention rate = 2051 / 2547 = 80.5% (below the target of 90% may be due to a large number of spam comments). Processing efficiency: The complete processing of 2547 comments was completed within 15 seconds.

[0029] The second step involves multi-dimensional feature extraction and weight evaluation of the comments. For the valid set of 2051 comments, the system extracts features in five dimensions and scores the comments based on the following dimensions: user influence (number of followers, certification level, account activity, historical influence), comment quality (text length, information richness, logical coherence, professionalism), interaction feedback (number of likes, number of replies, number of reposts), time effect (posting delay, lifecycle position), and credibility (account history, certification status).

[0030] Calculate the overall weight using the weighted fusion formula: Overall weight = w1 x F(user influence) + w2 x F(review quality) + w3 x F(time effect) + w4 x F(interaction feedback) + w5 x F(credibility); Where the weight parameters are set as: w1=0.25, w2=0.3, w3=0.15, w4=0.2, w5=0.1; Different quality reviews get different weight scores. High-quality reviews (detailed, logically clear, from authenticated users, and with more interactions) usually have a weight of more than 0.6; medium-quality reviews (short content, limited information) have a weight between 0.3 and 0.5; and low-quality reviews (spam reviews, meaningless content) have a weight of less than 0.2.

[0031] Weight ranking and core review selection: The overall 2051 reviews are evaluated and ranked by weight. The weight system has a discrimination degree of 0.87, which can effectively distinguish reviews of different quality.

[0032] The system calculates the cumulative distribution function (CDF) of the weight, and finds that: - The cumulative contribution of the top 100 reviews (about 4.9%) in the weight ranking = 28.5% - The cumulative contribution of the top 200 reviews (about 9.8%) in the weight ranking = 45.2% - The cumulative contribution of the top 300 reviews (about 14.6%) in the weight ranking = 58.3% - The cumulative contribution of the top 400 reviews (about 19.5%) in the weight ranking = 68.9% - The cumulative contribution of the top 500 reviews (about 24.4%) in the weight ranking = 76.8% - The cumulative contribution of the top 542 reviews (about 26.4%) in the weight ranking = 80.1% Therefore, the system selects the top 542 reviews (accounting for 26.4% of all reviews) in the weight ranking as the core review set for subsequent causal reasoning analysis. The average weight of these 542 reviews is 0.485, which is significantly higher than the average weight of all reviews, which is 0.298.

[0033] The evaluation mechanism has a discrimination degree of 0.87, indicating that the weight system can effectively distinguish reviews of different quality.

[0034] Third, review backtracking analysis and post quality evaluation, based on the 542 core review set, the system conducts a comprehensive evaluation of the post quality. The first step of the comprehensive evaluation of the post quality is the initial quality score calculation: based on the post's own characteristics, the post quality score (Q_initial) is calculated; Q_initial = w_author x score_author + w_length x score_length + w_structure x score_structure + w_time x score_time; wherein: - score_author: average quality score of the publisher's historical posts (0-1.0); - score_length: post content length score (500-2000 words is optimal, 1.0); - score_structure: post structure completeness score (contains text, pictures, videos, and other multimedia elements); - score_time: post time score (9-18 on weekdays is optimal, 1.0); - weight setting: w_author=0.4, w_length=0.3, w_structure=0.2, w_time=0.1; Step 2 of comprehensive evaluation of post quality - comment feedback-driven quality adjustment: adjust the score based on feedback data from the core comment set, comment sentiment comprehensive score (sentiment_score): statistics of sentiment distribution of core comments; sentiment_score = (p_pos - p_neg) x intensity_factor; where p_pos is the proportion of positive comments, p_neg is the proportion of negative comments, and intensity_factor is the comment intensity factor.

[0035] Keyword feature score (keyword_score): TF-IDF keyword extraction and sentiment polarity analysis on core comments; keyword_score = Σ(tf_idf_i x polarity_i) / total_keywords; where tf_idf_i is the keyword weight, and polarity_i is the sentiment polarity of the keyword.

[0036] Comment support degree index (support_degree): calculate the support degree of comments to post quality; support_degree = (number of core comments / total number of comments) x average comment weight x scale_factor; Feedback-driven quality score (Q_feedback) adjustment: Q_feedback = Q_initial + (sentiment_score × 0.4 + keyword_score × 0.3 + support_degree × 0.3); The adjustment range is limited to ±0.5 points, ensuring that the final score is between 1.0 and 5.0.

[0037] Step 3 of comprehensive evaluation of post quality - Confidence score (confidence_score) confidence_score = w1 × comment density + w2 × weight distribution balance + w3 × comment coverage time span; Where: - Comment density: The proportion of core comments in total comments (15%-25% is optimal); - Weight distribution balance: Whether the distribution of high, medium, and low weight comments is reasonable; - Comment coverage time span: The time coverage of comments after the post is published (comment coverage rate > 70% within 24 hours is sufficient); - Weight setting: w1 = 0.3, w2 = 0.3, w3 = 0.4; Final post quality score Q_final = Q_feedback, confidence level determined by confidence_score (> 0.7 is high confidence).

[0038] Counterfactual verification: Test the stability of the score by simulating changes in key factors.

[0039] Weight stability verification: Remove the top 5% high weight comments, recalculate sentiment distribution and quality score. If the score change |Δ_Q_w| < 0.15, it means that the score is not easily affected by a single high weight comment, and the stability is high.

[0040] Scale stability verification: Assume that the number of comments changes by 50%, use weight compensation mechanism to adjust, recalculate the overall score. If the score change |Δ_Q_n| < 0.2, it means that the score has good robustness.

[0041] Time window verification: Only use the comments within the first 24 hours after the post is published for quality evaluation, and compare the results with the full amount of comments. If the score change |Δ_Q_time| < 0.1, it means that early comments have fully reflected the quality characteristics of the post.

[0042] Comprehensive stability evaluation: counterfactual verification passed, indicating that the quality score obtained has strong stability and reliability.

[0043] Fourth step, multi-level recognition and tracing of negative content, three layers of negative recognition, 2051 full-amount comments are analyzed, lexical layer recognition: using precise matching and fuzzy matching (edit distance threshold 0.8) to scan in the negative vocabulary library, identifying the frequency and intensity of negative vocabulary. Calculate the lexical layer score (Negative_score_lexical): Negative_score_lexical = min(Σ weight_i × position_decay_i / 10,1.0); Where weight_i is the word weight, position_decay_i is the position decay coefficient.

[0044] Semantic layer recognition: through satire marker word recognition, semantic polarity reversal detection, implicit negative expression recognition, fine-grained sentiment analysis and other methods for deep semantic understanding, calculate the semantic layer score (Negative_score_semantic): Negative_score_semantic = satire score × 0.35 + implicit negative score × 0.35 + explicit negative score × 0.30; Context layer recognition: comprehensive consideration of time background, social background, user group characteristics, content relevance and other external information for comprehensive judgment, calculate the context layer score (Negative_score_context): Negative_score_context = (lexical layer score + semantic layer score + context adjustment coefficient × max score) / 3.

[0045] Multi-dimensional risk assessment: Comprehensive risk score = w1 × H_harm + w2 × S_impact + w3 × P_spread + w4 × (1-T_credibility); Where: - H_harm (degree of harm) = illegality × 0.35 + moral problem × 0.25 + hatred degree × 0.20 + other harm × 0.20; - S_impact (social impact) = (1 + user influence × 0.5) × (controversial potential × 0.4 + group exacerbation × 0.3 + misleading risk × 0.3); - P_spread (Spread Risk) = Time Context Coefficient x (Historical Retweet Rate + Spread Speed + Vulnerability) / 3; - T_credibility (Credibility) = Fact Verification x 0.4 + Source Reliability x 0.4 + Evidence Sufficiency x 0.2; - Weight Setting: w1=0.35, w2=0.30, w3=0.20, w4=0.15; Risk Level Determination: - High Risk (>0.75): Immediate Warning Triggered; - Medium Risk (0.50-0.75): Marked for Attention; - Low Risk (0.25-0.50): Data Recorded; - Extremely Low Risk (<0.25): No Warning Triggered.

[0046] Step 5, Analysis Result Verification and Differentiated Output, Counterfactual Verification: Based on multiple experimental scenarios of counterfactual verification, the stability of the verification score is comprehensively verified. If all verifications pass (the score change is within the specified threshold), it means that the obtained score has strong stability and reliability.

[0047] Confidence Assessment: Based on the counterfactual verification results, calculate the 95% confidence interval of the post quality score. The confidence interval width D_confidence reflects the stability of the evaluation results. If D_confidence < 0.5, it is high confidence; 0.5-1.0 is medium confidence; >1.0 is low confidence.

[0048] Comprehensive Confidence Score = 0.35 x Confidence Interval Factor + 0.35 x Counterfactual Verification Pass Rate + 0.15 x Confidence Interval Factor + 0.15 x Robustness Score. Confidence Level Classification: Determine the confidence level according to the comprehensive confidence score. Differentiated Output Report: The system generates a corresponding differentiated output report based on the combination of the confidence level and the quality score. High-quality posts (Q>4.0 and confidence>0.85): Output "Recommended Reference" report, showing the advantages of the post for other creators to learn from. The report includes post basic information, comprehensive quality score, key advantages, content strategy analysis, release timing reference, etc. Low-quality posts (Q<3.0 and confidence>0.85): Output "Problem Diagnosis" report, providing improvement solutions for the main problem points identified, and strongly suggesting optimization by the publisher. The report includes a problem list, severity score, and improvement suggestions. Medium-confidence posts (0.65≤confidence≤0.85): Output "Reference Analysis" report, marking uncertainty information and points that need further verification, and suggesting a comprehensive judgment combined with manual review.

[0049] Early warning and feedback mechanism: according to the risk score and quality score, the system automatically triggers the corresponding feedback and early warning mechanism, high-risk content: immediately push the early warning information to the content review personnel, including the problem key points and the disposal priority, medium-risk content: included in the artificial review queue, regularly summarized for review personnel to review, excellent content: push the positive feedback to the publisher, encourage him to continue to maintain the same creative style, this embodiment fully demonstrates the application effect of the technical scheme of the present application, and verifies the effectiveness of the core technologies such as multi-dimensional weight evaluation, simplified quality evaluation model, multi-level negative identification, counterfactual verification and differential output.

[0050] In the fifth step, the parallel computing implementation of the real-time stream processing module is also calculated, the parallel computing implementation is realized through multi-threaded computing: the system uses the Multiprocessing library of Python to realize the multi-threaded processing of data weighting and hot event generation process, so as to improve the computing efficiency. The division of parallel tasks is based on data sources, which ensures the independent and efficient processing of each data source, a distributed computing framework is adopted: Apache Spark is used for distributed processing of large-scale data, and the processing tasks of different data sources are allocated to different nodes in the cluster for execution, so as to speed up the computing process, and the implementation of real-time stream processing is realized, real-time data stream management: the system uses Apache Kafka to build a data stream pipeline to realize real-time transmission and processing of data, each data source corresponds to a Kafka topic, and data flows to the processing module through Kafka, wherein, the stream processing engine uses Apache Flink as the real-time stream processing engine to perform real-time calculation and analysis on the data transmitted from Kafka, so as to ensure that the hot event can be generated immediately when the data arrives and recommended to the user.

[0051] In the figure, the description of the position relationship is only used for exemplary illustration, and cannot be understood as a limitation on the patent; obviously, the above embodiments of the present application are only examples for clearly illustrating the present application, and are not a limitation on the embodiments of the present application. For ordinary skilled persons in the art, other different forms of changes or variations can be made on the basis of the above description. Here, it is not necessary and impossible to enumerate all the embodiments. Any modification, equivalent replacement and improvement made within the spirit and principles of the present application shall be included in the protection scope of the claims of the present application.

Claims

1. A method for analyzing posts based on comment backtracking, characterized in that: Includes the following steps: Step 1: The data acquisition unit obtains comment data for posts from social media platforms and preprocesses the posts; Step 2: The feature extraction unit performs multi-dimensional comprehensive feature extraction on the preprocessed post comment data and evaluates the comprehensive feature weights. Step 3: The retrospective analysis unit comprehensively evaluates the quality of posts based on the high-weight comment set and establishes a quantitative mapping relationship from comment feedback to post quality; Step 4: The negative identification unit adopts a hierarchical and multi-dimensional detection mechanism to comprehensively identify and trace the source of negative content in the comments from three levels: lexical layer, semantic layer, and contextual layer. Step 5: The verification unit performs multi-dimensional verification and credibility assessment on the results of the comment backtracking analysis and negative identification, and generates corresponding differentiated outputs based on the combination of credibility level and quality score.

2. The analysis method for analyzing posts based on comment backtracking according to claim 1, characterized in that: Step 1 includes the following steps: Step 11: The data acquisition unit obtains comment data of posts from social media platforms, and the preprocessing unit uses the Locality Sensitive Hash (LSH) algorithm to identify duplicate comments on the raw comment data collected by the data acquisition unit. Step 12: Identify spam comments using a hybrid approach combining rule-based and machine learning methods; Step 13: Unify the encoding and full-width / half-width characters of the comment text, convert between simplified and traditional Chinese characters, and standardize special characters. Step 14: Perform a completeness check and supplementation on the metadata of the comment text, including the publication time, user information, and comment level; Step 15: Establish a clear correspondence record for each comment and post.

3. The analysis method for analyzing posts based on comment backtracking according to claim 2, characterized in that: In step 12, spam comments are advertisements and comments that have no real meaning or content.

4. The analysis method for analyzing posts based on comment backtracking according to claim 1, characterized in that: In step 2, the multi-dimensional comprehensive features include user influence dimension, comment quality dimension, interaction feedback dimension, time effect dimension, and user behavior credibility dimension; The user influence dimensions include fan base size, account verification level, account activity, and historical influence. The dimensions of comment quality include text length, sentence complexity, information richness, logical coherence, professionalism, and emotional bias. The interactive feedback dimensions include the number of likes, replies, shares, and overall interaction score; The time-effect dimensions include posting delay, position in the post's lifecycle, and comment stream density; The user behavior credibility dimensions include comment frequency, domain expertise, historical accuracy, and credibility score.

5. The analysis method for analyzing posts based on comment backtracking according to claim 1, characterized in that: In step 2, the comprehensive feature weights are evaluated using the following formula: Comprehensive feature weight = w1×F (user influence) + w2×F (comment quality) + w3×F (time effect) + w4×F (interaction feedback) + w5×F (user behavior credibility); The mapping functions for each dimension are as follows: - F (User Influence) = [log(Fan Scale + 1) / 20 + Account Verification Level / 5 + Account Activity + Historical Influence] / 4, where, in the account verification level, ordinary account = 1, personal verification = 3, enterprise verification = 4, professional verification = 5; - F (Comment Quality) = [min(Text Length / 300)×0.3 + Sentence Complexity×0.2 + Information Richness×0.3 + Logical Coherence×0.1 + Professionalism×0.1]; - F(time effect) = exp(-release delay / half-term decay time); - F (Interactive Feedback) = Overall Interaction Score; - F (credibility) = credibility score.

6. The analysis method for analyzing posts based on comment backtracking according to claim 1, characterized in that: Step 3 includes the following steps: Step 31: The backtracking analysis unit calculates the initial score based on the post's own characteristics, and obtains the initial score result: Initial rating result = w_author × publisher's historical rating + w_length × content length + w_structure × structural completeness + w_time × publication time effect, where the weights are set as follows: w_author=0.4, w_length=0.3, w_structure=0.2, w_time=0.1; Step 32: Score the sentiment distribution of the high-weighted comment set. The sentiment distribution score of the high-weighted comment set = (proportion of positive comments - proportion of negative comments) × comment intensity factor, where the range of the comment intensity factor is 0-1.

0. Step 33: Score the keyword features. Keyword score = Σ(keyword weight × keyword sentiment polarity) / total number of keywords, where the sentiment polarity of the keyword ranges from -1 to 1. Step 34: Calculate the degree to which comments contribute to the quality of the post and adjust the quality score accordingly, with an adjustment range of approximately ±0.5 points, to ensure that the post quality score is between 1.0 and 5.

0. Step 35: Evaluate the credibility of the post quality score. Credibility score = (comment density × 0.3) + (weight distribution balance × 0.3) + (comment coverage time span × 0.4); Step 36: Score the quality of the final post based on its credibility score; Step 37: The verification unit tests the stability of the rating by simulating changes in key factors. The top 5% of high-weight comments are removed, and the sentiment distribution score of the high-weight comment set and the final post quality score are recalculated. The magnitude of the rating change is observed. If the magnitude of the rating change is < 0.15, it means that the rating is not easily affected by a single comment and has high stability. Step 38: Verify the stability of the scale. Assuming the number of comments decreases or increases by 50%, recalculate the overall score using a weight compensation method and observe the magnitude of the score change. If the magnitude of the score change is < 0.2, it indicates that the score has good robustness. Step 39: Verify the time window by using only comments from the first 24 hours after the post is published for quality assessment and comparing them with the results of using all comments. If the comparison result is < 0.1, it means that the early comments can fully reflect the quality of the post.

7. The analysis method for analyzing posts based on comment backtracking according to claim 1, characterized in that: Step 4 includes the following steps: Step 41: The negative identification unit performs word segmentation on the comment text, searches the negative vocabulary database using an exact string matching algorithm, identifies the emotional state, specific meaning, and sensitivity of words at the vocabulary level, performs fuzzy matching using the edit distance algorithm for variant words, and calculates the vocabulary level score based on the number, weight, and position distribution of the matched words. Step 42: The semantic layer recognition mechanism identifies irony and sarcasm, detects semantic reversal to obtain an irony score, identifies implicit negative expressions, identifies seemingly neutral but actually negative meanings, calculates semantic similarity, uses the distance between word vectors in the semantic space to determine the implicit meaning, obtains an implicit negative score, uses a multi-layer neural network model for fine-grained sentiment judgment, and uses natural language processing technology to score the overall sentiment polarity of each sentence to obtain an explicit negative score, and performs a comprehensive semantic layer score: Semantic layer score = Irony score × 0.35 + Implicit negative score × 0.35 + Explicit negative score × 0.30; Step 43: Based on the lexical and semantic layers, user information is introduced, and a comprehensive judgment is made from the context layer. Based on the three-layer negative identification results, the system establishes a multi-dimensional risk assessment framework to evaluate the overall risk level of the content, social impact assessment, dissemination risk assessment, credibility and comprehensive risk score. Step 44: After completing the negative identification, the source tracing unit traces the source of the negative content and locates the problem in the post.

8. The analysis method for analyzing posts based on comment backtracking according to claim 1, characterized in that: Step 5 includes the following steps: Step 51: The verification unit uses counterfactual verification to test the stability of the analysis results, and observes the range of score changes by simulating changes in key factors. Step 52: Verify the stability of the weights by removing the top 5% of high-weight comments. If the rating change is less than 0.15, the analysis results are stable. Step 53, verify the stability of the scale: Assuming the number of comments changes by 50%, if the change in rating is < 0.2, it indicates that the method has good robustness; Step 54, verify the time window: Analyze the comment data from the first 24 hours. If the rating change is less than 0.1, it means that the early comments have fully reflected the quality of the post. Step 55: Classify the credibility levels of the results of the review backtracking analysis and negative identification, and generate corresponding differentiated outputs based on the combination of credibility level and quality score.

Images