Chinese rebuttal speech generation method based on multi-strategy generation and adversarial criticism rewriting

Abstract

The application provides a Chinese refutation speech generation method based on multi-strategy generation and counter-critic revision, and belongs to the technical field of natural language processing. The method sequentially performs five steps. First, Chinese hate speech is collected. A set of different types of refutation strategies is built. Then, the hate speech is matched with each type of refutation strategy to input the refutation generation model to output multi-perspective candidate refutation speeches. Then, relying on the counter-critic model, the opposite user's perspective is simulated to excavate the defects of the candidate speeches to obtain criticism opinions. The hate speech, strategy and criticism opinion are combined and input into the revision model to complete constraint optimization to generate revised refutation speeches. Finally, the comprehensive scores of the revised refutation speeches are calculated through multi-dimensional quantitative weighted scoring, and the highest score speech is selected as the final refutation speech output. The application can simultaneously improve the relevance, persuasiveness, scene appropriateness and sentence fluency of the refutation content, and provides a flexible management scheme for the moderate and controllable hate speech of the network platform.

Description

Technical Field

[0001] This invention relates to the fields of language processing and content governance, specifically to a method for generating Chinese rebuttal statements based on multi-strategy generation and adversarial criticism rewriting. Background Technology

[0002] With the booming development of social media and online communities, online platforms have become important spaces for the public to express their opinions and exchange information. However, the widespread dissemination of hate speech on online platforms has become a serious social problem, significantly impacting public discourse order, the rights of target groups, and platform governance. Compared to simply deleting or blocking hate speech, rebuttal is a more moderate approach that respects freedom of speech. It weakens the spread of hate speech through rational, constructive, and non-aggressive responses, guiding publishers to re-examine their original viewpoints and minimizing controversies surrounding freedom of speech.

[0003] Currently, methods for generating automatic rebuttal statements mainly focus on the following three directions: The first approach employs a sequence-to-sequence model. For example, Chinese patent CN107832310A discloses a structured argument generation method and system based on a seq2seq model. This method receives the input topic through an encoder-decoder structure and generates related main and secondary arguments. In rebuttal generation tasks, such methods typically use hate speech as encoder input and generate corresponding rebuttal text through the decoder. However, traditional Seq2Seq models based on recurrent neural network architectures are limited by their structure and struggle to effectively capture the complex semantic relationships and implicit biases in hate speech. This often results in rebuttals that deviate from the original claims, have loose argumentation structures, and are expressed in a stiff manner, lacking strategic focus.

[0004] The second approach employs pre-trained language models. With the emergence of pre-trained language models such as BERT and GPT, researchers have begun to leverage large-scale pre-trained knowledge to improve the quality of generated rebuttals. This type of method fine-tunes the pre-trained model to adapt it to the rebuttal generation task. For example, related research has used pre-trained language models to fine-tune manually annotated corpora of rebuttals to generate rebuttals against hate speech. However, while pre-trained language models improve the fluency of the generated text, the models themselves lack a deep understanding of the specific claims of hate speech. The generated rebuttals often exhibit template-like characteristics and insufficient relevance to the original hate claims. Furthermore, when faced with hate attack targets not covered in the training data, the rebuttals generated by the model show a significant decrease in specificity.

[0005] The third approach, based on large language model-based prompting, involves researchers exploring how prompting engineering can guide large language models to generate rebuttals. For example, some studies employ large language model-driven position-aware methods to generate dialectical statements refuting the original statement. Additionally, some research uses role-playing prompting strategies to guide large language models in generating more empathetic rebuttals. However, existing research indicates that rebuttals generated by large language models are often lengthy and less readable. Furthermore, in large-scale field experiments, context-sensitive rebuttals generated by large language models have proven ineffective and may even have counterproductive effects. Moreover, while the safety barriers of large language models aim to prevent the generation of harmful content, research suggests that safety constraints may actually impair the argumentative quality and effectiveness of rebuttals.

[0006] To address the aforementioned problems, there is an urgent need to design an automatic rebuttal generation method for Chinese hate speech, which can generate highly relevant, persuasive, naturally expressed Chinese rebuttals suitable for publication in real online scenarios. Summary of the Invention

[0007] To address the shortcomings of existing technologies, the present invention aims to provide a method for generating Chinese rebuttal statements based on multi-strategy generation and adversarial criticism rewriting. This method addresses the practical problems of current automatic rebuttal generation technologies for Chinese hate speech, such as generalized irrelevant answers, insufficient argumentation, and expressions that easily trigger conflict. It generates multi-angle candidate rebuttals through multi-type differentiated strategies, simulates the perspectives of opposing users to uncover flaws in candidate responses and rewrites them with targeted constraints, and establishes a multi-dimensional quantitative scoring mechanism to automatically select the optimal response, thereby improving the relevance, persuasiveness, contextual appropriateness, and fluency of the generated rebuttal statements.

[0008] This invention provides a method for generating Chinese rebuttal statements based on multi-strategy generation and adversarial criticism rewriting, the method comprising: The data collection step involves collecting Chinese hate speech, which refers to text containing discriminatory or offensive content; The steps for constructing a set of rebuttal strategies involve obtaining different types of rebuttal strategies based on pre-defined role descriptions, generation requirements, and output constraints, thus acquiring a set of rebuttal strategies. This step involves building a multi-dimensional and differentiated rebuttal logic system to provide multi-angle rebuttal benchmarks for the same hate speech. Each strategy is equipped with a dedicated role, generation rules, and output restrictions to constrain the direction of model generation and avoid rebuttals that are singular or one-sided, thus addressing the shortcomings of existing technologies with a single rebuttal perspective. The candidate rebuttal statement generation step involves combining the Chinese hate speech with the set of rebuttal strategies and inputting the results into the rebuttal statement generation model to obtain candidate rebuttal statements. This step combines hate speech with various rebuttal strategies to generate differentiated candidate rebuttals, forming a multi-angle alternative response pool. This not only improves the rebuttal coverage but also differs from the traditional method of generating similar texts through random sampling. The constraint rewriting step of candidate rebuttal statements involves combining the candidate rebuttal statements with the corresponding Chinese hate speech and the rebuttal strategy, inputting them into the adversarial criticism generation model to obtain corresponding criticism opinions, and then combining the Chinese hate speech, the rebuttal strategy, the candidate rebuttal statements and the corresponding criticism opinions, and inputting them into the rewriting model to obtain candidate corrected rebuttal statements. An adversarial criticism generation model is used to identify the flaws and vulnerabilities in the initial candidate rebuttals from an adversarial perspective. Then, a rewriting model is used to correct the criticisms based on the initial hate speech, corresponding strategies, and current rebuttals to ensure that the criticisms are closely aligned with the target of this round of rebuttal. The step of selecting and outputting the final rebuttal statement is to calculate the comprehensive score of each candidate revised rebuttal statement, and then select the candidate revised rebuttal statement with the highest comprehensive score as the final rebuttal statement and output it. By establishing a multi-dimensional quantitative scoring system, all candidate rebuttal statements are scored in a standardized manner, and the best ones are evaluated and selected to replace subjective human judgment.

[0009] Furthermore, the types of rebuttal strategies include, but are not limited to, factual rebuttal strategies, logical rebuttal strategies, emotional rebuttal strategies, rhetorical rebuttal strategies, and protective rebuttal strategies; The factual rebuttal strategy is used to identify factual claims, misrepresentations, or inappropriate generalizations in the Chinese hate speech. The logical rebuttal strategy is used to analyze the premises, causal relationships, reasoning processes, and conclusions in the Chinese hate speech. The emotional rebuttal strategy is used to employ a calm, restrained, and non-confrontational mode of expression; The rhetorical rebuttal strategy is used to enhance the clarity, readability, and expressive effect of the rebuttal content through rhetorical means; The protective rebuttal strategy is used to identify the individuals or groups attacked by the Chinese hate speech.

[0010] By refining the specific types of each rebuttal strategy and clarifying the task boundaries of each strategy, clear and actionable rules are provided for generating rebuttal statements to differentiate attack patterns against different Chinese hate speech, resulting in targeted and matching rebuttal logic.

[0011] Furthermore, the construction steps of the rebuttal speech generation model specifically include: Obtain a pre-trained large language model as a base model; Configure the corresponding generation prompt template according to the role, description, generation requirements and output constraint rules of each of the aforementioned rebuttal strategies; The Chinese hate speech is combined with the generated prompt template corresponding to the selected rebuttal strategy, and the base model is called to generate candidate rebuttal speech.

[0012] By configuring independent prompt templates for each rebuttal strategy, different rebuttal content with different focuses can be generated for the same hate speech based on the prompt templates of different strategies, thereby improving the coverage and comprehensiveness of the rebuttal content. Furthermore, different strategy templates set corresponding roles and generation boundaries, and strategic generation can ensure that the output content closely follows the specific aspects of hate speech, guiding the generated comments to be more targeted and relevant.

[0013] Furthermore, the output constraint rules include: The candidate rebuttals directly address the specific claims made in the Chinese hate speech; The rebuttal statements do not contain any thought process or reasoning process, nor do they contain any offensive content; The text of any rebuttal must comply with the requirements for public release.

[0014] The design of output constraint rules aims to set hard boundaries during the generation of rebuttal statements, filter out invalid, illegal, and unusable candidate texts in advance, eliminate text content that does not conform to the rules from the source of generation, improve the appropriateness and relevance of the content, and reduce the pressure of correction in subsequent rebuttal criticism and rewriting stages.

[0015] Furthermore, the construction steps of the adversarial criticism generation model are as follows: Obtain a pre-trained large language model as a base model; Configure an adversarial criticism prompt template, which includes fill spaces for the Chinese hate speech, the rebuttal strategy, and the candidate rebuttal statements, and set the generation perspective, which is the perspective of the hate speech publisher; The Chinese hate speech, the rebuttal strategy, and the candidate rebuttal statements are filled into the adversarial criticism prompt template, and the base model is called to generate corresponding criticism opinions.

[0016] By constructing an adversarial interactive perspective, the criticisms simulate the specific questions and counterattacks that hate speech publishers might raise against rebuttals in real-world online scenarios. This allows the model to accurately identify vulnerabilities from an adversarial interactive perspective, making the criticisms more relevant to users' actual concerns rather than empty general evaluations or superficial embellishments at the grammatical or phrasing level. Furthermore, each candidate rebuttal is independently generated with a unique evaluation, and this differentiated criticism mechanism ensures that subsequent rewritten statements maintain their own strategic differences, preventing multiple candidates from ultimately converging.

[0017] Furthermore, the specific construction steps of the rewriting model are as follows: Obtain a pre-trained large language model as a base model; Configure a rewrite prompt template, which includes the Chinese hate speech, the rebuttal strategy, the candidate rebuttals and the corresponding criticisms, and includes rewrite constraint rules; The Chinese hate speech, the rebuttal strategy, the candidate rebuttals and the corresponding criticisms are filled into the rewrite prompt template, and the base model is called to generate the corrected rebuttal.

[0018] By constructing a criticism-driven rewriting model, an independent correction process is set up after the generated comments, so that each round of candidate rebuttals has the opportunity to fix defects based on specific criticisms, effectively improving the upper limit and stability of output quality.

[0019] Furthermore, the rewrite constraint rules include: No new content unrelated to the aforementioned Chinese hate speech shall be introduced; no new claims without factual basis shall be added; and the rebuttal strategy corresponding to the candidate rebuttal statement shall not be changed. The revised rebuttal does not contain any rewritten explanations or reasoning processes; the revised rebuttal matches the claims of the Chinese hate speech and meets the requirements for public release.

[0020] By designing rewriting constraint rules, negative issues such as excessive rewriting, deviation from the topic, addition of false information, and change of rebuttal angle are prevented, thereby improving the credibility of the rebuttal content. Furthermore, the rewriting process strictly locks the original rebuttal strategy, ensuring the independence of the differentiated candidates of multiple strategies and preventing the convergence of responses from different strategies after rewriting.

[0021] Furthermore, the step of selecting and outputting the final rebuttal statement includes a comprehensive scoring strategy, specifically comprising: Obtain the scores for each of the corrected rebuttal statements across multiple evaluation dimensions, including but not limited to appropriateness, persuasiveness, relevance, and fluency; The appropriateness reflects the suitability of the revised rebuttal in real-world online scenarios; the persuasiveness reflects the persuasiveness of the revised rebuttal to the publishers of Chinese hate speech; the relevance reflects the matching relevance between the revised rebuttal and the Chinese hate speech; and the fluency reflects the natural coherence of the content expressed in the revised rebuttal. Each of the evaluation dimensions is assigned a preset weight, and the scores of each evaluation dimension are weighted and summed with their corresponding weights to calculate the corresponding comprehensive score.

[0022] The comprehensive scoring strategy designed considers four dimensions: appropriateness, persuasiveness, relevance, and fluency, in order to cover all the core requirements for evaluating rebuttal arguments, establish a standardized quantitative evaluation system, and achieve automatic selection of the best candidate rebuttal arguments.

[0023] Furthermore, the step of selecting and outputting the final rebuttal statement also includes a selection mechanism, which is triggered when two or more identical highest comprehensive scores are obtained. Specifically: The scores of each revised rebuttal statement are compared in the following order of priority: appropriateness, persuasiveness, relevance, and fluency. When there are differences in scores on any evaluation dimension, the revised rebuttal statement with the higher score on that dimension is selected as the final rebuttal statement.

[0024] The optimal selection mechanism is designed to address the boundary scenario where multiple candidate rebuttal arguments have the same overall score and cannot be automatically filtered. By using a priority fallback rule, it ensures that a unique response can be output.

[0025] The beneficial effects of this invention are: This invention establishes a multi-strategy differentiated generation mechanism, relying on multiple independent rebuttal strategies to generate candidate rebuttals with essential perspectives and differences for the same hate speech. Each strategy is equipped with a dedicated role, generation rules, and output restrictions to constrain the direction of rebuttal generation, thereby solving the problem of a single rebuttal perspective in the existing rebuttal generation process and completely overcoming the technical shortcomings of traditional single-strategy rebuttals that are one-sided and omit key attack points.

[0026] This invention designs a dual-loop closed-loop optimization chain for adversarial criticism linkage and constraint rewriting. From the perspective of adversarial interaction, it accurately locates various defects and loopholes in the candidate rebuttal statements generated in the first round. Without changing the original rebuttal strategy, fabricating false information, or introducing irrelevant topics, it completes localized targeted correction based on exclusive criticism opinions. This solves the problems of traditional single-round rebuttal statement generation methods lacking self-correction capabilities and rewriting only remaining at a superficial level of text polishing. Furthermore, in both the rebuttal statement generation process and the rewriting stage to obtain the corrected rebuttal statements, multiple layers of output constraints are set to achieve pre-control of text compliance, eliminating the need for additional security filtering modules, thereby improving the appropriateness and relevance of the content. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating the Chinese rebuttal statement generation method based on multi-strategy generation and adversarial criticism rewriting of the present invention.

[0028] Figure 2This is a schematic diagram of the multi-strategy candidate generation process in the Chinese rebuttal statement generation method based on multi-strategy generation and adversarial criticism rewriting of the present invention.

[0029] Figure 3 This is a schematic diagram of the criticism-driven rewriting process in the Chinese rebuttal speech generation method based on multi-strategy generation and adversarial criticism rewriting of the present invention.

[0030] Figure 4 This is a schematic diagram of the multi-dimensional scoring and candidate selection process in the Chinese rebuttal statement generation method based on multi-strategy generation and adversarial criticism rewriting of the present invention. Detailed Implementation

[0031] The technical solutions of the embodiments 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] It should be noted that when a component is described as "fixed to" another component, it can be directly on the other component or may have a component in between. When a component is considered "connected to" another component, it can be directly connected to the other component or may have a component in between. When a component is considered "set on" another component, it can be directly set on the other component or may have a component in between. The terms "vertical," "horizontal," "left," "right," and similar expressions used in this document are for illustrative purposes only.

[0033] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0034] Example 1 To overcome the shortcomings of existing methods for generating Chinese rebuttal statements, such as over-reliance on a single rebuttal strategy, vague results, lack of targeted correction, and difficulty in closely adhering to the original hate claims, this invention proposes a method for generating Chinese rebuttal statements based on multi-strategy conditional generation and adversarial criticism rewriting. For example... Figure 1 As shown, the method for generating rebuttal statements specifically includes the following: The data collection step is used to collect Chinese hate speech, which is specifically represented by text containing discrimination or attacks. It can be derived from social media comments, online forum posts, content moderation platforms, manually submitted text, or pre-built text datasets.

[0035] Furthermore, this step includes a text processing mechanism to standardize the format of the collected text content, including removing invalid whitespace characters, standardizing text encoding, preserving necessary contextual information, and marking the targets of attacks in the input text. This processing does not alter the main claims expressed in the original text. The processed Chinese hate speech serves as the common input for subsequent generation stages, ensuring that different candidate rebuttals are generated targeting the same original claim.

[0036] The step of constructing a set of rebuttal strategies involves a set of rebuttal strategies of multiple different types. Each rebuttal strategy is defined with a corresponding role description, generation requirements, and output constraints. Specifically, a rebuttal strategy refers to a text generation guidance prompt rule established based on a pre-defined set of roles with independent rebuttal logic, descriptions of generation requirements, and output constraints.

[0037] By building a multi-dimensional and differentiated rebuttal strategy, we can provide a multi-angle rebuttal benchmark for the same Chinese hate speech, thereby solving the shortcoming of the single rebuttal perspective in existing technologies.

[0038] In one embodiment of the present invention, the types of rebuttal strategies include, but are not limited to, factual rebuttal strategies, logical rebuttal strategies, emotional rebuttal strategies, rhetorical rebuttal strategies, and protective rebuttal strategies; specifically: Factual rebuttal strategies are used to identify factual claims, misrepresentations, or inappropriate generalizations in hate speech and to correct the original claims using specific, verifiable facts. Logical rebuttal strategies are used to analyze the premises, causal relationships, reasoning processes, and conclusions of hate speech, pointing out overgeneralizations, inverted causal relationships, insufficient evidence, or internal contradictions, and constructing rebuttal content with coherent reasoning. Emotionally charged rebuttal strategies are used to maintain a rebuttal stance while employing a calm, restrained, and non-confrontational mode of expression to reduce the resistance of hate speech publishers and guide them to re-examine their original viewpoints. Rhetorical rebuttal strategies are used to generate rebuttal content through metaphors, parallelism, rhetorical questions, interrogative sentences, antithesis, or other rhetorical devices, thereby enhancing the clarity, readability, and expressive effect of the rebuttal content. At the same time, it is required that the rhetorical content should not replace the substantive rebuttal. Protective rebuttal strategies are used to identify individuals or groups targeted by hate speech, emphasize the dignity, safety, and legitimate rights of the targets, and generate rebuttal content by publicly supporting and regulating boundary expressions in order to mitigate the impact of discriminatory content.

[0039] It should be noted that the number and type of rebuttal strategies can be adjusted according to the actual application scenario. Different strategies have different rebuttal focuses, thus enabling the generation of differentiated candidate rebuttal content for the same Chinese hate speech, rather than simply obtaining multiple similar candidate content through random repetition.

[0040] The candidate rebuttal statement generation step combines Chinese hate speech with a set of rebuttal strategies and inputs them into the rebuttal statement generation model to obtain candidate rebuttal statements, such as... Figure 2 As shown.

[0041] Specifically, the construction steps of the rebuttal argument generation model are as follows: Select a high-performance Chinese pre-trained large language model as the base model, such as the qwen3-4b model; After selecting the base model, configure the corresponding prompt templates based on the role descriptions, generation requirements, and output constraints of various rebuttal strategies. The role descriptions are used to define the professional role that the model should play during the generation process; the generation requirements are used to clarify the specific rebuttal tasks that need to be completed under the strategy; and the output constraints are used to restrict the model's generation behavior. After the template design is completed, in the real-time generation stage, the input Chinese hate speech text is combined with the generated prompt template corresponding to the selected rebuttal strategy to form a complete model input. Then, by calling the base model and inputting it, candidate rebuttal statements are obtained.

[0042] For the k-th rebuttal strategy, the process of generating rebuttal statements is represented as follows:

[0043] in, This represents a model for generating rebuttal statements; Chinese hate speech; This represents the k-th rebuttal strategy; This represents the candidate rebuttal statement generated according to the k-th rebuttal strategy.

[0044] It should be noted that the various strategy generation tasks in this step can be executed sequentially or in parallel. Furthermore, the model's process of generating rebuttal statements is subject to output constraints, specifically: candidate rebuttal statements should explicitly respond to the original hate speech; they should not output the model's thought process or reasoning; they should not contain new offensive content; they should maintain linguistic restraint and be suitable for use in online public discussion scenarios.

[0045] Furthermore, a retrieval mechanism is included for factual rebuttal strategies. This mechanism generates retrieval statements based on the input Chinese hate speech, retrieves factual materials related to the original claim from external information sources, and inputs these factual materials as supplementary context into the rebuttal generation model. To reduce resource consumption from repeated searches, the correspondence between the input Chinese hate speech, the retrieval statement, and the retrieval results can be saved. When processing the same or similar inputs, the saved retrieval results can be retrieved.

[0046] Based on the candidate rebuttal argument generation steps, a candidate set consisting of candidate rebuttal arguments is obtained:

[0047] in, This refers to the candidate rebuttal statements generated according to the first rebuttal strategy; This refers to the candidate rebuttal statements generated according to the second rebuttal strategy; This represents the candidate rebuttal statement generated according to the nth rebuttal strategy; n represents the number of rebuttal strategies.

[0048] The constraint rewriting step for candidate rebuttal statements involves combining the candidate rebuttal statements with corresponding Chinese hate speech and the rebuttal strategy, inputting this combination into an adversarial criticism generation model to obtain corresponding criticisms, and then combining the Chinese hate speech, the rebuttal strategy, the candidate rebuttal statements, and the corresponding criticisms again, inputting this combination into the rewriting model to obtain candidate corrected rebuttal statements, such as... Figure 3 As shown.

[0049] Specifically, an adversarial criticism generation model needs to be built first. The building steps are as follows: A high-performance Chinese pre-trained large language model is selected as the base model. It should be noted that the base model used by the rebuttal speech generation model can be used, or different models can be selected according to actual needs. After obtaining the base model, configure the adversarial criticism prompt template. First, the template sets the corresponding generation perspective. The generation perspective refers to the stance and identity positioning adopted by the model when generating text. Different generation perspectives will guide the model to output different content. This perspective is the opposing and questioning perspective of the hate speech publisher. In addition, the adversarial criticism prompt template also includes the original Chinese hate speech, the rebuttal strategy adopted, and the filling space for the candidate rebuttal speech generated in the first round. After the prompt template is designed, the original Chinese hate speech, the corresponding rebuttal strategy, and the candidate rebuttal statements are filled into the corresponding fill positions of the prompt template to obtain the complete model input. Then, the base model is called and input to obtain the criticism text as the output.

[0050] For the k-th candidate rebuttal, the corresponding generated criticism is represented as follows:

[0051] in, This represents an adversarial criticism generation model. Chinese hate speech; This represents the k-th rebuttal strategy; This represents the candidate rebuttal arguments generated according to the k-th rebuttal strategy; This represents the criticism corresponding to the k-th candidate's rebuttal statement.

[0052] When generating critical comments, the original Chinese hate speech, candidate rebuttals, and the rebuttal strategies employed by each candidate rebuttal are simultaneously input. This allows the resulting criticisms to provide context-sensitive and precise evaluations based on both the original attack claims and the candidate responses, avoiding generalized assessments detached from specific contexts. It accurately identifies specific flaws in the candidate rebuttals, including at least the following types: failure to directly address the original hate claim; omission of key attack points; insufficient factual evidence; unclear logical relationships; overly vague expression; weak persuasiveness; tone that may escalate conflict; and rhetoric that deviates from the topic or fails to demonstrate protection for the attacked group. Different candidate rebuttals correspond to independent adversarial criticisms, thus avoiding the use of a uniform evaluation template that results in a lack of targeted criticism.

[0053] After receiving criticism, the rewriting model is used to revise the candidate rebuttal statements based on the criticism. The specific steps for building the rewriting model are as follows: Similarly, a pre-trained large language model with general text understanding and generation capabilities is required as the base model. This base model can be the same as the rebuttal argument generation model or the adversarial criticism generation model to simplify the architecture, or different independent models can be selected according to actual needs. After obtaining the base model, design a rewritten prompt template. This template should include at least the original Chinese hate speech, the rebuttal strategy adopted, the first round of candidate rebuttals, and the corresponding fill space for the criticism generated in response to the candidate rebuttals. Meanwhile, rewriting constraint rules are set in the template to ensure that the model strictly follows these constraints when generating revised content. These constraints specifically include: no new topics unrelated to the original hate speech are introduced; no new claims without factual basis are added; the main rebuttal strategy adopted by the original candidate is not changed; no modification explanations or reasoning processes are output; the rewritten result should continue to target the original hate claims and maintain the expression that meets the requirements for public release. After completing the rewriting prompt template design, in the rewriting and correction stage, the original Chinese hate speech, the corresponding rebuttal strategy, the first round of candidate rebuttals, and the corresponding criticisms are filled into the four fill positions of the rewriting prompt template to build the model input, call the base model and at the same time comply with the rewriting constraint rules to obtain the corrected rebuttal.

[0054] The rewriting process for the k-th candidate rebuttal is expressed as follows:

[0055] Where R represents the rewritten model; Chinese hate speech; This represents the k-th rebuttal strategy; This represents the candidate rebuttal arguments generated according to the k-th rebuttal strategy; This represents the criticism corresponding to the k-th candidate rebuttal statement; This represents the revised candidate rebuttal statement after rewriting the rebuttal statement of the k-th candidate.

[0056] Through the constraint rewriting steps of candidate rebuttal statements, a set of candidate revised rebuttal statements is obtained:

[0057] in, This indicates the revised rebuttal statement of the candidate after rewriting the first candidate's rebuttal statement; This indicates the revised rebuttal statement of the candidate after rewriting the rebuttal statement of the second candidate; This represents the revised candidate rebuttal statement after rewriting the m-th candidate rebuttal statement; m represents the number of rebuttal statements.

[0058] The steps for selecting and outputting the final rebuttal are as follows: calculate the comprehensive score of each candidate revised rebuttal, and select the candidate revised rebuttal with the highest comprehensive score as the final rebuttal and output it. Figure 4 As shown.

[0059] This step includes a comprehensive scoring strategy to calculate a comprehensive score for each candidate revised rebuttal statement, specifically: The first step is to obtain the scores of each corrected rebuttal statement on multiple evaluation dimensions. These evaluation dimensions include, but are not limited to, appropriateness, persuasiveness, relevance, and fluency. The evaluation model is used to set corresponding scoring rules for each evaluation dimension to obtain the corresponding score value. The individual score range for each evaluation dimension is set between 1 and 5 points. The higher the value, the better the corrected rebuttal statement performs in that dimension. Among them, appropriateness is used to evaluate whether the candidate corrective rebuttal statements can be directly used in real online scenarios, and whether they contain offensive, inflammatory, offensive or conflict-inducing expressions. Persuasiveness is used to evaluate whether candidate corrective rebuttals can prompt hate speech publishers to reconsider, restrain, or revise their original views and expressions. Relevance is used to evaluate whether the candidate rebuttal arguments closely revolve around the specific claims of the original hate speech, whether they respond to key content, and whether there are any off-topic or irrelevant extensions. Fluency is used to evaluate the grammar, vocabulary, structure, and readability of candidate rebuttal arguments, judging whether they are natural, concise, and coherent. It should be noted that the scoring rules for each evaluation dimension can be adjusted and refined based on specific actual needs, and the evaluation model can adopt a different language model than the rebuttal speech generation model, thereby reducing the bias generated when the generation model evaluates its own generation results.

[0060] The second step is to assign preset weights to each of the evaluation dimensions, and to calculate the corresponding comprehensive score by weighted summation of the scores of each evaluation dimension and their corresponding weights. Specifically, for the k-th candidate revised rebuttal statement, its comprehensive score is expressed as:

[0061] in, This represents the overall score of the k-th candidate's revised rebuttal statement; This represents the appropriateness score of the k-th candidate's revised rebuttal statement; This represents the persuasiveness score of the k-th candidate's revised rebuttal statement; This represents the relevance score of the k-th candidate revised rebuttal statement; This represents the fluency score of the k-th candidate's revised rebuttal statement; , , as well as This represents the weight corresponding to each evaluation dimension, and satisfies... + + + =1.

[0062] After calculating the overall score for each candidate rebuttal, the candidate rebuttal with the highest overall score is determined as the final output by comparing them:

[0063] in, This indicates the final output of corrective and rebuttal statements; This represents a function that retrieves the maximum value in a set.

[0064] In addition, this step also includes a selection mechanism, which is triggered when two or more identical highest overall scores are obtained. Specifically: The revised rebuttal statements are compared on each evaluation dimension in order of priority: appropriateness, persuasiveness, relevance, and fluency. When there is a difference in scores on any evaluation dimension, the revised rebuttal statement with the higher score on that dimension is selected as the final rebuttal statement.

[0065] Through these steps of the rebuttal speech generation method proposed in this invention, Chinese rebuttal speech suitable for use in actual network scenarios can be obtained.

[0066] This invention utilizes a multi-strategy condition generation mechanism to enable rebuttals to respond to Chinese hate speech from multiple complementary angles, avoiding the problems of limited rebuttal angles and insufficient coverage caused by a single strategy. By designing an adversarial criticism model, it can identify problems such as misaligned claims, weak arguments, inappropriate tone, or insufficient persuasiveness in the candidate rebuttals generated in the first round. Through a criticism-driven rewriting model, it can make targeted modifications to the responses while maintaining the original rebuttal strategy intent, improving the rebuttals' relevance to the original hate claims and their practical usability.

[0067] Furthermore, the rebuttal statements generated by this invention are more suitable for scenarios such as social media, content governance, and manual review assistance. They can be used to mitigate the spread of hate speech, support platforms in building more moderate governance methods, and form structured data containing candidate replies, criticisms, rewritten results, and rating results, providing support for subsequent research on Chinese rebuttal statements and system optimization.

[0068] Example 2 In one embodiment of the present invention, the input of a Chinese hate speech and the generation, criticism and rewriting of the corresponding rebuttal speech can be achieved using the qwen3-4b model. The three functions can be implemented by switching different prompt word templates and configuring different decoding parameters; different models can also be selected according to actual needs. The scoring model used in the scoring process can be the qwen-plus model; and the length of rebuttal statements can be limited to 200 Chinese characters, and criticisms can be limited to 300 Chinese characters. It should be noted that this embodiment is used to illustrate the implementation process of the present invention, but the present invention is not limited to the specific model, dataset or parameters in this embodiment.

[0069] The process of generating the final rebuttal is as follows: Read and process input Chinese hate speech; Five types of rebuttal strategy branches are invoked in parallel to generate candidate rebuttal statements of factual, logical, emotional, rhetorical and protective nature, respectively. For the factual branch, two to three Chinese search queries can be generated based on Chinese hate speech, some factual materials can be obtained through the search service, and the search results can be used as the context for generating factual candidates. The adversarial criticism model is invoked for each of the five candidate rebuttal statements to generate five corresponding exclusive criticisms; Each candidate rebuttal and its corresponding criticisms are rewritten to obtain five rewritten candidate corrected rebuttals; The evaluation model is invoked to score the five rewritten candidate rebuttal statements on four dimensions: appropriateness, relevance, fluency, and persuasiveness. Specifically, each dimension can be scored five times and the average value is taken as the final score. Calculate the overall score for each candidate rebuttal statement and select the final rebuttal statement based on the highest score; if there is a tie in the total score, compare the scores corresponding to appropriateness, persuasiveness, relevance and fluency in that order to select the final rebuttal statement. Output the final Chinese rebuttal, and save the intermediate processing results and the scoring results at the same time.

[0070] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principle of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for generating Chinese rebuttal statements based on multi-strategy generation and adversarial criticism rewriting, characterized in that: The method includes: The data collection step involves collecting Chinese hate speech, which refers to text containing discriminatory or offensive content; The steps for constructing a set of rebuttal strategies involve obtaining different types of rebuttal strategies based on pre-defined role descriptions, generation requirements, and output constraints, thus acquiring a set of rebuttal strategies. The candidate rebuttal statement generation step involves combining the Chinese hate speech with the set of rebuttal strategies and inputting the results into the rebuttal statement generation model to obtain candidate rebuttal statements. The constraint rewriting step of candidate rebuttal statements involves combining the candidate rebuttal statements with the corresponding Chinese hate speech and the rebuttal strategy, inputting them into the adversarial criticism generation model to obtain corresponding criticism opinions, and then combining the Chinese hate speech, the rebuttal strategy, the candidate rebuttal statements and the corresponding criticism opinions, and inputting them into the rewriting model to obtain candidate corrected rebuttal statements. The step of selecting and outputting the final rebuttal statement involves calculating the comprehensive score of each candidate revised rebuttal statement, and selecting the candidate revised rebuttal statement with the highest comprehensive score as the final rebuttal statement and outputting it.

2. The method for generating Chinese rebuttal statements based on multi-strategy generation and adversarial criticism rewriting according to claim 1, characterized in that, The types of rebuttal strategies include, but are not limited to, factual rebuttal strategies, logical rebuttal strategies, emotional rebuttal strategies, rhetorical rebuttal strategies, and protective rebuttal strategies; The factual rebuttal strategy is used to identify factual claims, misrepresentations, or inappropriate generalizations in the Chinese hate speech. The logical rebuttal strategy is used to analyze the premises, causal relationships, reasoning processes, and conclusions in the Chinese hate speech. The emotional rebuttal strategy is used to employ a calm, restrained, and non-confrontational mode of expression; The rhetorical rebuttal strategy is used to enhance the clarity, readability, and expressive effect of the rebuttal content through rhetorical means; The protective rebuttal strategy is used to identify the individuals or groups attacked by the Chinese hate speech.

3. The method for generating Chinese rebuttal statements based on multi-strategy generation and adversarial criticism rewriting according to claim 2, characterized in that, The construction steps of the rebuttal speech generation model specifically include: Obtain a pre-trained large language model as a base model; Configure the corresponding generation prompt template based on the role description, generation requirements and output constraint rules corresponding to each of the aforementioned rebuttal strategies; The Chinese hate speech is combined with the generated prompt template corresponding to the selected rebuttal strategy, and the base model is called to generate candidate rebuttal speech.

4. The method for generating Chinese rebuttal statements based on multi-strategy generation and adversarial criticism rewriting according to claim 3, characterized in that, The output constraint rules include: The candidate rebuttals directly address the specific claims made in the Chinese hate speech; The rebuttal statements do not contain any thought process or reasoning process, nor do they contain any offensive content; The text of any rebuttal must comply with the requirements for public release.

5. The method for generating Chinese rebuttal statements based on multi-strategy generation and adversarial criticism rewriting according to claim 3, characterized in that, The construction steps of the adversarial criticism generation model are as follows: Obtain a pre-trained large language model as a base model; Configure an adversarial criticism prompt template, which includes fill spaces for the Chinese hate speech, the rebuttal strategy, and the candidate rebuttal statements, and set the generation perspective, which is the perspective of the hate speech publisher; The Chinese hate speech, the rebuttal strategy, and the candidate rebuttal statements are filled into the adversarial criticism prompt template, and the base model is called to generate corresponding criticism opinions.

6. The method for generating Chinese rebuttal statements based on multi-strategy generation and adversarial criticism rewriting according to claim 5, characterized in that, The specific construction steps of the rewriting model are as follows: Obtain a pre-trained large language model as a base model; Configure a rewrite prompt template, which includes the Chinese hate speech, the rebuttal strategy, the candidate rebuttals and the corresponding criticisms, and includes rewrite constraint rules; The Chinese hate speech, the rebuttal strategy, the candidate rebuttals and the corresponding criticisms are filled into the rewrite prompt template, and the base model is called to generate the corrected rebuttal.

7. The method for generating Chinese rebuttal statements based on multi-strategy generation and adversarial criticism rewriting according to claim 6, characterized in that, The rewrite constraint rules include: No new content unrelated to the aforementioned Chinese hate speech shall be introduced; no new claims without factual basis shall be added; and the rebuttal strategy corresponding to the candidate rebuttal statement shall not be changed. The revised rebuttal does not contain any rewritten explanations or reasoning processes; the revised rebuttal matches the claims of the Chinese hate speech and meets the requirements for public release.

8. The method for generating Chinese rebuttal statements based on multi-strategy generation and adversarial criticism rewriting according to claim 1, characterized in that, The step of selecting and outputting the final rebuttal statement includes a comprehensive scoring strategy, specifically comprising: Obtain the scores for each of the corrected rebuttal statements across multiple evaluation dimensions, including but not limited to appropriateness, persuasiveness, relevance, and fluency; The appropriateness reflects the suitability of the revised rebuttal in real-world online scenarios; the persuasiveness reflects the persuasiveness of the revised rebuttal to the publishers of Chinese hate speech; the relevance reflects the matching relevance between the revised rebuttal and the Chinese hate speech; and the fluency reflects the natural coherence of the content expressed in the revised rebuttal. Each of the evaluation dimensions is assigned a preset weight, and the scores of each evaluation dimension are weighted and summed with their corresponding weights to calculate the corresponding comprehensive score.

9. The method for generating Chinese rebuttal statements based on multi-strategy generation and adversarial criticism rewriting according to claim 8, characterized in that, The step of selecting and outputting the final rebuttal statement also includes a merit-based mechanism, which is triggered when two or more identical highest overall scores are obtained. Specifically: The scores of each revised rebuttal statement are compared in the following order of priority: appropriateness, persuasiveness, relevance, and fluency. When there are differences in scores on any evaluation dimension, the revised rebuttal statement with the higher score on that dimension is selected as the final rebuttal statement.

Images