A cross-platform social media summary generation method based on large language model self-iteration

Abstract

The application provides a cross-platform social media summary generation method based on large language model self-iteration, which is used for topic clustering and high-quality summary generation of cross-platform and multi-field social media data. The specific technical scheme is as follows: firstly, text data of multiple fields is obtained from multiple social media platforms, and data preprocessing is carried out through synonym expansion and topic word screening. Then, based on the text similarity and the LDA model, the data is subjected to topic clustering to form a multi-document data set. Finally, the generative large language model is used to generate an initial summary, and the large model and the multi-dimensional auxiliary evaluation network are combined to evaluate and feedback in four dimensions of consistency, coherence, fluency and theme relevance, and iterative optimization is carried out until the multi-dimensional quality standard is met. The method uses the powerful generation capability of the generative large model, can automatically and iteratively generate high-quality summaries, reduces the artificial burden, and thus provides a reliable solution for efficient summary generation of cross-platform social media.

Description

Technical Field

[0001] This invention relates to the field of natural language processing, and more particularly to a cross-platform social media summary generation method based on a large language model with self-iteration. Background Technology

[0002] With the rapid growth of social media platforms, the amount of text data generated daily is enormous. Users often feel overwhelmed by such a massive information flow and urgently need effective technical means to extract and summarize this information scattered across multiple platforms. Against this backdrop, text summarization technology has emerged, aiming to automatically extract the core meaning from text and generate concise summaries.

[0003] Existing summary generation methods are mainly divided into two categories: extractive and generative. Extractive methods generate summaries by extracting key sentences from text. While simple to operate, extractive summaries often lack logical coherence and semantic fluency because they are merely concatenated with key sentences, resulting in poor readability. Generative methods utilize models such as deep learning to generate natural and fluent summary text. However, their single-generation model determines the uncontrollability of the result quality. If the generated content has significant deficiencies in consistency, topic relevance, and fluency, this single-generation method cannot iteratively optimize based on the generated results, thus failing to effectively improve the quality of the final summary. In addition, traditional methods often lack effective quality evaluation mechanisms, mostly focusing only on a single indicator of the generated result (such as grammar or fluency), failing to comprehensively evaluate the generated summary, and even more so failing to use the evaluation results to optimize and iterate the generation model. This static generation model is prone to information bias or omission in complex scenarios, leading to unstable quality of the generated summary and failing to meet the actual needs of multi-platform, multi-domain social media scenarios.

[0004] To address the aforementioned issues, this invention proposes a cross-platform social media summarization method based on a large language model with self-iteration. Through a complete document collection, clustering, and generation optimization process, it significantly improves the quality and efficiency of summarization generation. In the data preprocessing stage, this method first crawls text content from multiple social media platforms based on domain-specific trigger word groups and generates a broader set of trigger words through a synonym expansion mechanism to address the heterogeneity of language styles and expressions across platforms. Subsequently, topic modeling is performed on the original dataset based on the LDA model, providing a structured multi-document dataset for subsequent summarization. During the generation process, this invention utilizes a generative large language model to generate initial summaries, clarifying the generation logic through instruction templates to ensure the summaries maintain basic performance in terms of content coverage and language fluency.

[0005] To further improve summary quality, this method introduces the multi-dimensional auxiliary evaluation network EvalNet, which comprehensively evaluates the generated summaries from four dimensions: consistency, coherence, fluency, and relevance. The evaluation results serve as feedback scores, dynamically guiding the generation model to optimize the instruction template (Prompt_Iterate), adjust the generation strategy, and gradually improve the summary quality until it reaches the preset quality standard. The importance of EvalNet lies in its ability to effectively alleviate the "overconfidence" problem of large language models in generating their own content. By objectively examining the generated content through multi-dimensional evaluation, it improves the generation quality and ensures the reliability and relevance of the content. Through rigorous evaluation of the consistency, coherence, fluency, and relevance of the generated summaries, this invention significantly reduces the generation of false information while ensuring a high degree of relevance between the content and the topic. Compared to traditional methods, the multi-dimensional evaluation and self-iterative optimization mechanism of this invention overcomes the limitations of the single-generation mode, resulting in more accurate, coherent, and reliable generation results. It provides an innovative and efficient solution for summary generation in complex cross-platform and multi-topic scenarios. Summary of the Invention

[0006] To overcome the shortcomings of existing cross-platform, multi-domain social media summarization methods in terms of summarization quality, consistency, and automation, this invention proposes a cross-platform social media summarization method based on a large language model with self-iteration. This method achieves efficient processing and high-quality summarization of massive amounts of multi-domain text data from multiple social media platforms through the collaborative work of modules such as data collection, topic clustering, initial summarization, summarization feedback evaluation, and summarization optimization and verification.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A cross-platform social media summary generation method based on self-iteration of a large language model is characterized by comprising a data collection module, a topic clustering module, an initial summary generation module based on a generative large model, a summary feedback evaluation module, and a summary optimization and verification module based on a generative large model.

[0009] Data collection module: Based on trigger word groups set in multiple fields, collect relevant data across social media platforms to obtain the original dataset D;

[0010] Topic Clustering Module: Performs topic clustering on the original dataset D to form a multi-document dataset D. m ={T0,T1,...,T n}. Each topic T i ={d1,d2,...,dn} is a group containing multiple social text data related to this topic. i ;

[0011] The initial summary generation module based on the generative large model: Based on the constructed instruction prompt template Prompt_Init, the topic dataset T is input into the generative large language model M to generate the initial summary S0;

[0012] Summary Feedback Evaluation Module: Combining the constructed feedback evaluation prompt template, the module utilizes a large language model M and a multi-dimensional auxiliary evaluation network EvalNet to evaluate the generated summary S. i A joint evaluation was conducted to obtain a weighted multi-dimensional feedback score, Feedback_Score. i ;

[0013] The summary optimization and validation module is based on a generative large model: referencing the joint scoring Feedback_Score. i-1 Let the large language model M summarize its S i-1 Optimize and generate a new version of the summary Si. Repeat the summary feedback evaluation and optimization process until the iteration stopping condition (Stop) is met; otherwise, let the large language model M output the final Si. final .

[0014] Based on the above technical solution, the preferred data collection module is as follows:

[0015] The data collection module first sets up multiple domains, then constructs a trigger word group W for each domain, and then maps the words in the trigger word group W to the extended thesaurus of synonyms, thus obtaining the expanded trigger word group W. E According to W E Collect public opinion texts from various fields from multiple social media platforms;

[0016] Based on the above technical solution, the preferred topic clustering module is specifically as follows:

[0017] The original dataset D was modeled using the LDA (Latent Dirichlet Allocation) model to model the text topics, and the model with the highest Coherence score was selected as the optimal number of topics.

[0018] Extract and sort the keywords for each topic from the best LDA model, and select the top N keywords as the topic representatives;

[0019] The resulting topic structure includes keyword groups for each topic, used for categorizing document topic tags.

[0020] Based on the above technical solution, the preferred initial summary generation module based on the generative large model is as follows:

[0021] Design initial templates for abstracts in various fields, design specific initial prompt word templates for each field (Prompt_{Domain}), and define the structure and content requirements for generating abstracts.

[0022] Generate initial summary: Combine the text content with the corresponding domain cue template Prompt_{Domain}, input the cue into the generative large language model M, and generate the initial summary S0, expressed by the formula:

[0023] S0 = M(Prompt_{Domain}, T)

[0024] Where T is a set of multi-document data, and Prompt_Init is a text summary prompt.

[0025] Based on the above technical solution, the preferred feedback scoring module based on the generative large model is as follows:

[0026] The multi-dimensional auxiliary evaluation network EvalNet evaluates the summary Si from four dimensions: consistency, coherence, fluency, and relevance. The formula for EvalNet is as follows:

[0027] Feedback_Score EvalNet =E(S) i ,T)

[0028] Where S i Let T be the source multi-document data of the summary to be evaluated, and E be the EvalNet network model.

[0029] Correspondingly, the large model M scores the generated summary using input scoring examples combined with feedback scoring templates. M then uses these examples to score the summary across multiple dimensions: consistency, coherence, fluency, and relevance, resulting in Feedback_ScoreM, expressed by the formula:

[0030] Feedback_Score M =M(Prompt_Feedback,Example,S i ,T)

[0031] Example is a provided rating reference example, Prompt_Feedback is a feedback rating prompt template, and S i T represents the summary to be scored, and T represents the multi-document data.

[0032] The final weighted multi-dimensional feedback scoring formula is defined as follows:

[0033] Feedback_Score i =W1×Feedback_Score EvalNet +W2×Feedback_Score M Where W1 and W2 are weighting coefficients, and W1 + W2 = 1;

[0034] Based on the above technical solution, the preferred feedback scoring module based on a generative large model has a multi-dimensional auxiliary evaluation network as follows:

[0035] The loss function for each dimension in the multi-dimensional auxiliary evaluation network EvalNet is defined as follows:

[0036]

[0037] in Generate scores for each dimension of the i-th training data for EvalNet prediction. (i) represents the actual score for each individual dimension in the i-th training data.

[0038] The final total loss is defined as:

[0039] L total =α×L c +β×L co +γ×L f +δ×L r

[0040] Among them, L c For consistency loss, L co For continuity loss, L f For smoothness loss, L r The loss is for topic relevance; α, β, γ, and δ are the weights of the loss function for each dimension, and satisfy α+β+γ+δ=1.

[0041] Based on the above technical solutions, the preferred iterative optimization module based on a generative large model is expressed by the following formula:

[0042] S i =M(Prompt_Iterate,Feedback_Score i-1 ,S i-1 ,T)

[0043] Among them, S i-1 The abstract to be optimized is T, where S is the sum of its components. i-1 The data consists of multiple documents, feedback score i-1, and optimization prompt templates, where M represents the large language model.

[0044] The formula for the iteration termination condition is:

[0045] Stop=(i≥N max )∨(Feedback_Score i ≥θ)

[0046] The meaning is as follows: Condition 1 – Maximum Iteration Limit: Set a maximum number of iterations Nmax. If the current iteration number i exceeds Nmax, the iteration stops; Condition 2 – Joint Feedback Score Threshold: Set a joint feedback score threshold θ. If the joint feedback score Feedback_Scorei obtained in the current iteration reaches or exceeds the threshold θ, the iteration stops.

[0047] The technical effects and advantages of the cross-platform social media summary generation method based on large language model self-iteration of the present invention are as follows:

[0048] This method combines trigger word expansion and topic clustering techniques to efficiently process heterogeneous text data from multiple social media platforms and multiple fields, providing structured multi-document input data for generative large language models, generating coherent, accurate and highly relevant summaries, and significantly improving the efficiency and quality of information extraction.

[0049] This method achieves adaptive iterative optimization of the summary generation process by introducing a joint feedback mechanism of a multi-dimensional auxiliary evaluation network and a generative large language model. It dynamically adjusts the generation strategy across multiple dimensions, including consistency, coherence, fluency, and topic relevance. This mechanism not only reduces human intervention but also effectively improves the system's intelligence level and summary quality, ensuring that the generated results meet expected standards and adapt to diverse needs in different scenarios. Attached Figure Description

[0050] Figure 1 This is a flowchart of the cross-platform social media summarization method based on a generative large model according to the present invention. Detailed Implementation

[0051] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0052] This invention provides a cross-platform social media summary generation method based on a large language model with self-iteration. For example... Figure 1 As shown, Figure 1 The flowchart of this invention includes five main modules: data collection, topic clustering, initial summary generation, summary feedback evaluation, and summary optimization and verification.

[0053] Data collection module: Based on pre-defined trigger word groups from multiple fields, collect cross-domain data from social media platforms to form the original dataset D.

[0054] Topic Clustering Module: Performs topic clustering on the original dataset D to form a multi-document dataset D. m ={T0,T1,...,T n}

[0055] Initial summary generation module: Based on the constructed instruction prompt template Prompt_Init, the topic dataset T is input into the generative large language model M to generate an initial summary S0.

[0056] Abstract Feedback Evaluation Module: Combining the constructed feedback evaluation prompt template, the module uses a large language model M and a multi-dimensional auxiliary evaluation network EvalNet to jointly evaluate the generated abstract S0, obtaining a feedback score (Feedback_Score) to provide a reference for subsequent abstract optimization.

[0057] Abstract optimization and validation module: This module iteratively optimizes the initial abstract S0 using the joint score Feedback_Score until a preset termination condition is met. Through this module, the abstract optimization and validation progressively improves the abstract quality, ensuring that the final abstract meets the expected standards in terms of consistency, coherence, fluency, and topic relevance.

[0058] Let's take a specific example to illustrate:

[0059] The design involves multiple domains, each with its own trigger words. Taking the sports domain as an example, a trigger word group is constructed for that domain, and then expanded using a thesaurus to obtain an extended trigger word group W. E .

[0060]

[0061] Based on the data crawled from the above trigger word groups, the data items collected in the example and a sample data point are shown in the table below: Specifically, they include: Post ID (ID), Author ID (User_id), Post Content (content), Posting Time (Created_at), Number of Likes (likes), Number of Comments (comments), and Number of Retweets (retweets).

[0062]

[0063] The collected raw data undergoes preprocessing. First, important posts are filtered based on the number of likes, comments, and shares. If all three numbers (likes, comments, and shares) are zero, the post is deleted to retain only informational data. Second, a text similarity algorithm is applied. This example uses TF-IDF combined with cosine similarity to identify and remove duplicate posts; that is, if the similarity of multiple posts exceeds a set threshold of 90%, only one post is retained. Finally, meaningless symbols and garbled characters, such as consecutive punctuation marks and random characters, are removed from the text.

[0064] After the above processing, we obtain the preprocessed sample data:

[0065]

[0066] For the example above, using the LDA model to model the text's topics, the optimal number of topics is 2. This results in two topics:

[0067] Theme 1 (2022 FIFA World Cup Final): Argentina, France, Messi, penalty, Mbappé, final;

[0068] Theme 2 (2023 Wimbledon Tennis Championships Final): Alcaraz, Wimbledon, match, championship, final, history

[0069] Keywords for each topic are extracted and ranked from the best LDA model, and the top 5 keywords are selected as topic representatives:

[0070] Theme 1 (2022 FIFA World Cup Final): Argentina, France, Messi, penalty, Mbappé

[0071] Theme 2 (2023 Wimbledon Tennis Championship Final): Alcaraz, Wimbledon, match, championship, final

[0072] After topic clustering using the LDA model, a multi-document dataset D is obtained. m = {T0, T1}, which includes T0 = {Id0, Id1, Id4, Id5} and T1 = {Id2, Id3}.

[0073] In one specific embodiment, the initial summary generation based on the large language model includes:

[0074] Initial template for design domain summary: Taking the sports domain as an example, the prompt word template is as follows: "As an expert in summarizing text abstracts, the text below relates to the SPORT field, which may contain specific event names, match outcomes, and key individuals. Here is the multi-document data: {T}. Please generate a summary based on the above-mentioned multi-text data. Summary:"

[0075] The text content is combined with corresponding domain cue word templates, with documents separated using "|||". This is then input into a generative large language model M to generate an initial summary S. 0。

[0076] Initial summary S0: "Argentina won the 2022 FIFA World Cup in a thrilling final against France, with Messi leading his team to victory in a penalty shootout. Mbappé's hat-trick also made history in the match, which ended 3-3 after extratime."

[0077] In one specific embodiment, the training implementation of the multi-dimensional auxiliary evaluation network EvalNet includes:

[0078] The multi-dimensional auxiliary evaluation network EvalNet in the example uses a pre-trained model for fine-tuning training and selects the RoBERTa model as the backbone model.

[0079] The dataset used is the SummEval dataset, which was created in 2020 by researchers at Cornell University and the Allen Institute for AI. It is suitable for multi-dimensional evaluation tasks of abstract quality. The dataset has four manually labeled evaluation dimensions: Consistency, Coherence, Fluency, and Relevance.

[0080] The RoBERTa model was chosen for feature extraction, followed by a fully connected layer and an output layer for scoring output. During the initial training phase, the low-level weights of the RoBERTa model were frozen, and only the high-level parameters and the scoring head were trained. The loss function for each dimension during training was defined as follows:

[0081]

[0082] in To generate scores for each dimension of the i-th training data for EvalNet prediction, y (i) represents the actual score for each individual dimension in the i-th training data.

[0083] The final total loss is defined as:

[0084] L=α×L c +β×L co +γ×L f +δ×L r

[0085] Where Lc is the consistency loss, Lco is the coherence loss, Lf is the fluency loss, and Lr is the topic relevance loss; α, β, γ, and δ are the weights of the loss functions for each dimension, and satisfy α + β + γ + δ = 1. In the example, all weights are set to 0.25, i.e., α = β = γ = δ = 0.25.

[0086] In one specific embodiment, the generated summary is jointly evaluated using a large language model M and a multi-dimensional auxiliary evaluation network EvalNet to obtain a weighted multi-dimensional feedback score, specifically:

[0087] Design a feedback rating prompt template. The example template is constructed as follows: Prompt_Feedback: "You are an AI model for evaluating summaries, and we aim to improve the quality of summaries by evaluating their content. To achieve this, we provide several examples and score them based on multiple quality-related dimensions. Below are these examples:"

[0088]

[0089] Here is also an example from the sample data:

[0090]

[0091]

[0092] Regular expressions are used to extract the scoring data from the example file, along with the summary S to be evaluated and its source multitext data, and these are then populated into the Prompt_Feedback. The score for the example is then obtained.

[0093]

[0094]

[0095] Next, the trained auxiliary evaluation network EvalNet is used for inference to obtain multi-dimensional scores for the example data:

[0096]

[0097] Feedback_Score i =W1×Feedback_Score EvalNet +W2×Feedback_Score M

[0098] Using the combined weighting formula above, W1 = W2 = 0.5 in this example, resulting in the final feedback score for the example data:

[0099]

[0100] In one specific embodiment, the joint score Feedback_Score is referenced. i-1 Let the large language model M summarize its S i-1 Optimize and generate a new version of the summary S i Specifically:

[0101] Design feedback rating prompt templates. The example iterative rating prompt template is constructed as follows:

[0102] Pompt_Iterate: "You are an AI model for generating summaries, and your task is to iteratively improve the provided summary based on the givenfeedback. The summary must adhere to the following scoring criteria:relevance, coherence, fluency, consistency, and overall score.

[0103] Document content:{T}

[0104] Generated summary:{S}

[0105] Scores:{Feedback_All}

[0106] Please use the above feedback to improve the original summary.

[0107] Now,please provide the refined summary:”

[0108] The summary S to be evaluated and its source multitext data T are combined, and regular expressions are used to match the scores of each dimension in the scoring file. The results are then populated into Prompt_Iterate.

[0109] The optimized summary S1 is: "In the 2022 FIFA World Cup final, Argentina aced France in a penalty shootout after a 3-3 draw, with Messi and Mbappé delivering historic performances."

[0110] Repeat the summary feedback evaluation and optimization process randomly until the iterative optimization conditions are met:

[0111] Stop=(i≥N max )∨(Feedback_Score i ≥θ)

[0112] Condition 1 – Maximum Iteration Limit: A maximum number of iterations Nmax is set. If the current iteration number i exceeds Nmax... max In this example, N max If the score is 5, then stop the iteration; Condition 2 – Joint Feedback Score Threshold: Set the joint feedback score threshold θ. If the joint feedback score obtained in the current iteration is θ... i If the threshold θ is reached or exceeded, in this example θ = 0.95, the iteration stops.

[0113] The above description is merely one embodiment of the present invention, and the specific embodiments described herein are only for explaining the present invention and are not intended to limit the present invention. Any simple modifications, equivalent changes, and alterations made by those skilled in the art to the above embodiments based on the technical essence of the present invention without departing from the principles and spirit of the present invention should be included within the protection scope of the present invention.

Claims

1. A cross-platform social media summary generation method based on large language model self-iteration, characterized in that, The method includes a data collection module, a topic clustering module, an initial summary generation module based on a generative large model, a summary feedback and evaluation module, and a summary optimization and validation module based on a generative large model. Data collection module: Based on trigger word groups set in multiple fields, collect relevant data across social media platforms to obtain the original dataset D; A subject clustering module: subject clustering is performed on the original data set D to form a multi-document data set D m = {T0, T1,..., T n} wherein each topic T i = { d1, d2,..., d n } is a set containing a plurality of social text data d i associated with the topic; The initial summary generation module based on the generative large model: Based on the constructed instruction prompt template Prompt_Init, the topic dataset T is input into the generative large language model M to generate the initial summary S0; The abstract feedback evaluation module: combined with the constructed feedback evaluation prompt template, using the large language model M and the multi-dimensional auxiliary evaluation network EvalNet to generate the abstract S i Conduct joint evaluation to obtain a weighted multi-dimensional feedback score Feedback_Score i ; Based on the generative large model, the summary optimization and verification module: reference joint score Feedback_Score i-1 , let the large language model M optimize its summary S i-1 , generate a new version of the summary S i ; repeat the summary feedback evaluation and optimization before meeting the iteration stop condition Stop; otherwise, let the large language model M output the last S final ; The generated summaries are jointly evaluated using a large language model M and a multi-dimensional auxiliary evaluation network to obtain a weighted multi-dimensional feedback score, specifically: The multi-dimension auxiliary evaluation network EvalNet evaluates the summary S from four dimensions of consistency, coherence, fluency and relevance. i The formula of the multi-dimension auxiliary evaluation network EvalNet is as follows: ； where S i is the summary to be evaluated, T is the source multi-document data of the summary to be evaluated, and E is the EvalNet network model. Correspondingly, the score of the large model M is obtained by inputting the score example into the feedback score template, so that M refers to the consistency, coherence, fluency and theme relevance of the generated summary, and obtains Feedback_Score M , which is expressed by the formula: ； Wherein, Example is a score reference example provided, Prompt_Feedback is a feedback score prompt template, S i is the abstract to be scored, and T is multi-document data. The final weighted multi-dimensional feedback scoring formula is defined as follows: ; where W1, W2 are weight coefficients, and W1+W2=1.

2. The method of claim 1, wherein, Based on trigger word groups set across multiple domains, relevant data is crawled from various social media platforms, specifically: The data collection module firstly sets multiple fields, secondly constructs a trigger word group W of each field, and then maps the words in the trigger word group W to a synonym word forest expansion word table to obtain an expanded trigger word group W E ; according to W E , collect public opinion text of each field from multiple social media platforms.

3. The method of claim 2, wherein, The original data D is subjected to topic clustering to form a multi-document dataset, specifically: The original dataset D was modeled using the LDA (Latent Dirichlet Allocation) model to model the text topics, and the model with the highest Coherence score was selected as the optimal number of topics. Extract and sort the keywords for each topic from the best LDA model, and select the top N keywords as the topic representatives; The resulting topic structure includes keyword groups for each topic, used for categorizing document topic tags.

4. The method of claim 3, wherein, Based on the constructed instruction prompt template Prompt_Init, a set of topic-related data T is input into the generative large language model M to generate an initial summary S0, specifically: Design initial templates for abstracts in various fields, design specific initial prompt word templates for each field (Prompt_{Domain}), and define the structure and content requirements for generating abstracts. Generate initial summary: Combine the text content with the corresponding domain cue template Prompt_{Domain}, input the cue into the generative large language model M, and generate the initial summary S0, expressed by the formula: ; where T is a set of multi-document data, and Prompt_Init is a text summarization prompt.

5. The method of claim 4, wherein, The generated summaries are jointly evaluated using a large language model M and a multi-dimensional auxiliary evaluation network EvalNet to obtain a weighted multi-dimensional feedback score, specifically: The loss function L of each dimension in the multi-dimensional auxiliary evaluation network EvalNet c , co , f , r are defined as: ; wherein is the EvalNet prediction of each uni-dimensional rating in the ith training set data, is the actual rating of each uni-dimensional rating in the ith training set data; The final total loss is defined as: ； where L c is the consistency loss, L co is the coherence loss, L f is the fluency loss, L r is the topic relevance loss; and a, b, g, and d are weights of the respective dimension loss functions and satisfy a + b + g + d = 1.

6. The method of claim 5, wherein, Reference joint score Feedback_Score i-1 , let the large language model M optimize its summary S i-1 to generate a new version of the summary S i , expressed as: ； Wherein, S i-1 is the multi-document data to be optimized, T is the summary of S i-1 , the feedback score Feedback_Score i-1 and the optimization instruction prompt template prompt_iterate, M represents a large language model; The formula for the iteration termination condition is: ； The meaning is: Condition 1 - Maximum iteration count limit: Set a maximum number of iterations N. max If the current iteration number i reaches N max If the current iteration's score is not met, then the iteration stops; Condition 2 – Joint Feedback Score Threshold: Set a joint feedback score threshold θ; if the current iteration's joint feedback score Feedback_Score is not met, then the iteration stops. i If the threshold θ is reached or exceeded, the iteration stops.

Images