A false news detection method based on news comment relationship perception preference alignment

Abstract

This invention relates to the fields of natural language processing and social media content analysis, proposing and publishing a fake news detection method based on news comment relationship-aware preference alignment. The invention includes: First, guiding a teacher's large language model to generate high-quality inference trajectories around the relationship between news and comments through relationship-oriented cues and real label constraints, providing relationship-sensitive supervision signals for the student model. Second, based on the supervised fine-tuning initialization of the student model, a relationship-aware preference dataset is constructed. Correct inference processes and real labels are used as positive samples, and direct preference optimization and joint training are performed by combining the student model's erroneous outputs and multiple types of relationship error negative samples, thereby enhancing the model's ability to distinguish between correct and erroneous relationship inference. Through these improvements, this invention effectively alleviates the problem of insufficient modeling of news and comment relationships in existing methods and significantly improves the accuracy and interpretability of fake news detection tasks.

Description

Technical Field

[0001] This invention relates to a method for detecting fake news based on news comment relationship perception preference alignment, belonging to the fields of natural language processing, generative artificial intelligence and social media content analysis. Background Technology

[0002] Fake news detection aims to determine the authenticity of news based on its content and related information, and is an important research direction in online public opinion governance, information security, and social media content moderation. With the widespread use of social media platforms, news dissemination is no longer limited to the single news text itself. User comments, as immediate feedback and collective expression of news content, often contain valuable supplementary clues for judging the authenticity of news. Therefore, how to effectively utilize comment information to improve the performance of fake news detection has become an important research question in this field.

[0003] In recent years, with the increasing capabilities of Large Language Models (LLMs) in text generation and complex reasoning tasks, researchers have begun to explore their application in fake news detection. Existing methods typically involve designing prompt templates or fine-tuning the LLMs to enable them to output authenticity labels or explanations based on news text and related comments. Some methods also introduce reasoning chain generation mechanisms to enhance the interpretability of the model's output. However, existing techniques still have the following shortcomings. First, while many methods incorporate comment information, they often simply splice or superficially merge comments as supplementary text, failing to fully characterize the interaction between comments and news. Second, even when using LLMs to generate the reasoning process, the model output is easily affected by its own generation bias and comment noise, resulting in insufficient relationship modeling and a lack of noise filtering. Third, traditional supervised fine-tuning mainly relies on positive sample training, which only teaches the model "what should be output," but it is difficult to explicitly tell the model "which reasoning paths, although seemingly reasonable, should not actually be preferred." Therefore, the model's ability to discern relationships on complex samples remains limited.

[0004] To address the aforementioned problems, this invention proposes a fake news detection method based on news-comment relationship-aware preference alignment. This method first utilizes a teacher's large language model to generate high-quality inferences about the relationship between news and comments under the constraint of real labels, providing relationship-sensitive supervision signals for the student model. Then, a lightweight student model is initialized through supervised fine-tuning. Based on this, preference data pairs are further constructed, using not only the student model's erroneous outputs as negative samples but also three additional types of negative samples: relationship-missing, noise-biased, and relationship-inverted. Joint training with direct preference optimization and supervised fine-tuning loss enhances the model's ability to distinguish between correct and incorrect relationship inferences. Through these improvements, this invention effectively alleviates the problem of insufficient modeling of news-comment relationships in existing methods and improves the accuracy and interpretability of fake news detection. Summary of the Invention

[0005] This invention provides a fake news detection method based on news comment relationship-aware preference alignment, in order to alleviate the problems of shallow utilization of comments, insufficient relationship modeling, and difficulty in effectively distinguishing correct and incorrect reasoning paths in existing large language model methods.

[0006] The technical solution of this invention is: a method for detecting fake news based on news comment relationship-aware preference alignment, the method comprising:

[0007] Step 1: Construct a reasoning trajectory generation module for modeling the relationship between news and comments. Input the news text, comment set, and real labels into the teacher's large language model. Generate a positive sample reasoning process consistent with the real labels through relationship-guided prompts for subsequent student model training.

[0008] Step 2: Construct a lightweight student model initialization module. Based on the positive sample inference process and real labels, perform supervised fine-tuning initialization of the student model to enable the student model to have preliminary ability to infer the relationship between news and comments and generate real labels.

[0009] Step 3: Construct a relation-aware preference data construction and joint optimization module. Use the correct reasoning process and real labels generated by the teacher's big language model as positive preference samples. Combine the erroneous output of the student model after supervised fine-tuning with the multi-class relation error negative samples constructed by the big language model to form preference pairs. On this basis, perform direct preference optimization on the student model and train it in conjunction with supervised fine-tuning loss to improve the model's ability to identify and model the relationship between news and comments.

[0010] Step 4: Perform fake news detection and prediction. Input the news text and comment set in the test sample into the trained student model, output the corresponding reasoning process and authenticity label, and obtain the final fake news detection result.

[0011] Furthermore, Step 1 specifically includes the following:

[0012] Step 1.1: This invention models fake news detection as a structured generation task. Let the news text be... The corresponding set of user comments is Through predefined prompt template functions Organize news and commentary into input context Target model Target sequence needs to be generated ,in To represent the reasoning process, Indicates the final predicted label, Let the target sequence be... Then its conditional generation probability is defined as ,in For the i-th generated token, Indicates a historical prefix.

[0013] Step 1.2: Transfer the news text Comments Collection and real labels Input the teacher's large language model and generate an inference process consistent with the real labels using a relation-guided prompt template. Thus, the supervised target sequence is obtained. .

[0014] Step 1.3: The relationship-oriented prompt template is used to guide teachers to organize reasoning around the interaction between news and comments using the large language model. Specifically, it includes the following three sub-processes: 1) Comment noise filtering and evidence screening: filtering emotional expressions, repetitive statements, irrelevant chatter, and invalid interactions in the comment set, retaining key comments relevant to the judgment of authenticity; 2) News-comment relationship analysis: analyzing the questioning, correction, supplementation, agreement, or misleading effects of comments on news content to form an evidence chain around the relationship between news and comments; 3) Reasoning generation under label constraints: explicitly injecting real labels into the prompts. This guides teachers to generate reasoning processes consistent with real labels using large language models, thereby improving the consistency and integrity of reasoning trajectories with gold labels.

[0015] Step 1.4: Obtain the supervised training dataset based on the above process. ,in .

[0016] Furthermore, Step 2 specifically includes the following:

[0017] Step 2.1: Select the lightweight student model as the target model, and use the positive sample inference trajectory dataset generated based on the teacher's large language model. The student model is subjected to efficient parameter fine-tuning using a low-rank fitting method, and the supervised fine-tuning loss is defined as follows:

[0018]

[0019] Step 2.2: By optimizing the supervised fine-tuning loss, the student model learns the relational reasoning trajectory generated by the teacher's large language model, thereby obtaining the initial "reasoning process generation + label prediction" capability, providing an initial model for subsequent relational perception preference optimization.

[0020] Furthermore, Step 3 specifically includes the following:

[0021] Step 3.1, for the input sample The correct reasoning process and real labels generated by the teacher's large language model are used as positive preference samples, i.e. ;

[0022] Step 3.2: Constructing model misjudgment negative samples. The supervised, fine-tuned student model is then compared to the input... The generated result is represented as ,when When this is the case, it is used as a model misjudgment negative sample to reflect the true error distribution of the student model on complex samples.

[0023] Step 3.3: Construct Negative Samples with Missing Relationships. Using a large language model, negative samples are generated under constraints, ignoring key interactions between news and comments while maintaining correct labels. This is represented as follows: This type of negative sample is used to constrain the model to distinguish between suboptimal reasoning that is “correctly labeled but has insufficient relationship modeling” and optimal reasoning that is “reasonable in relation and correctly labeled”.

[0024] Step 3.4: Construct noise-biased negative samples. Using a large language model, negative samples are generated that maintain correct labeling while overly relying on comment sentiment, stance conformity, or local high-frequency noise under constrained cues. This is represented as follows: These negative samples are used to suppress the model's over-reliance on comment noise cues.

[0025] Step 3.5: Construct relation-reversal negative samples. Using a large language model, under constraint cues, the logical relationship between news items and comments is intentionally reversed, generating reverse labels consistent with the erroneous relationship. Let the reverse label be denoted as... ,in Then the negative sample of the relation inversion type is represented as These negative samples contain both relation-level and label-level errors, which are used to enhance the model's ability to identify erroneous relation inference paths.

[0026] Step 3.6: For each input sample Constructing a negative sample set And further construct a preference dataset

[0027]

[0028] Each preference sample is derived from the input. Positive samples and a single negative sample This is used for subsequent direct preference optimization training.

[0029] Step 3.7: Based on the supervised fine-tuning of the student model, utilize the preference dataset. Perform direct preference optimization on the model. Let... For the strategy model to be optimized, To freeze the reference model, then for preference samples... The direct preference optimization loss is defined as

[0030]

[0031] in, It is the Sigmoid activation function. This is a hyperparameter used to control the intensity of preference.

[0032] Step 3.8: To avoid the lightweight student model deviating from the original language distribution due to direct preference optimization alone, supervised fine-tuning loss is introduced during the preference optimization stage for joint training. The joint loss function is expressed as follows: .in, The represents the balance coefficient. Through joint optimization, the model simultaneously possesses generative stability and relation discriminative ability.

[0033] Furthermore, Step 4 specifically includes the following:

[0034] Step 4.1: In the prediction phase, for the news text in the test sample... and a collection of comments Construct input context using predefined prompt templates .

[0035] Step 4.2: Set the input context Input the trained student model and generate the output sequence. ,in, The reasoning process generated for the model, This is a label for the accuracy of the model's predictions.

[0036] Step 4.3, add the authenticity label As the output of fake news detection results, and the reasoning process is also included. This serves as a corresponding interpretable criterion, thereby enabling fake news detection based on modeling the relationship between news and comments.

[0037] The present invention also provides a fake news detection system based on news comment relationship-aware preference alignment, the system comprising: a module for executing the fake news detection method based on news comment relationship-aware preference alignment.

[0038] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the fake news detection method based on news comment relationship-aware preference alignment.

[0039] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the fake news detection method based on news comment relationship-aware preference alignment.

[0040] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the fake news detection method based on news comment relationship-aware preference alignment.

[0041] The beneficial effects of this invention are:

[0042] This invention proposes a fake news detection method based on news comment relationship-aware preference alignment, which enhances the effective utilization of comment information in fake news detection tasks. This method has three main advantages: First, it generates high-quality inference trajectories through relationship-guided hints and label constraints, enhancing the model's ability to model the relationship between news and comments. Second, it utilizes a lightweight student model for supervised fine-tuning initialization, achieving preliminary inference and discrimination capabilities while ensuring deployment efficiency. Finally, it constructs multiple types of erroneous relational negative samples and performs joint training combined with direct preference optimization, enabling the model to better distinguish between correct and incorrect relational inferences. Experimental results show that the proposed method effectively improves the accuracy of fake news detection and enhances the interpretability of the model's output. Attached Figure Description

[0043] Figure 1 This invention provides a framework for a fake news detection method based on news comment relationship perception and preference alignment. Detailed Implementation

[0044] The method proposed in this embodiment was implemented on two public datasets, RumourEval-19 and Weibo-comp.

[0045] RumourEval-19 originates from the SemEval-2019 rumor detection task and includes social media posts and related discussions from Twitter and Reddit. The original annotations categorize samples into three classes: "real," "fake," and "unverified." This invention uses a three-class classification version for experiments, focusing on news text and related comments for model training and evaluation. Weibo-comp is a Chinese fake news detection dataset containing Weibo post text, user comments, and other auxiliary information. This dataset covers multiple news fields, including technology, politics, and military, exhibiting strong thematic diversity. The original data contains three labels: "real," "fake," and "undetermined." This invention retains only the "real" and "fake" samples and uses the corresponding news text and comments for experiments. It should be noted that although this dataset contains additional modal information such as images, this invention only uses news text and its corresponding text comments as model input. Statistical information for the RumourEval-19 and Weibo-comp datasets is shown in Table 1.

[0046] Table 1 Dataset Statistics

[0047] While existing fake news detection methods can determine authenticity using news text and comments, most methods treat comments merely as shallow auxiliary features, simply splicing or fusing them without fully modeling the interaction between news and comments. This makes it difficult for the model to effectively distinguish between questioning, agreement, and noise in the comments. Furthermore, large language models are prone to insufficient relationship modeling or over-reliance on comment noise during the generation and inference process, thus affecting fake news detection performance. To address these issues, this invention proposes a fake news detection method based on news comment relationship-aware preference alignment; as follows... Figure 1 As shown, a fake news detection method based on news comment relationship-aware preference alignment is used. The method includes:

[0048] Step 1: Construct a reasoning trajectory generation module for modeling the relationship between news and comments. Input the news text, comment set, and real labels into the teacher's large language model. Generate a positive sample reasoning process consistent with the real labels through relationship-guided prompts for subsequent student model training.

[0049] Furthermore, Step 1 specifically includes the following:

[0050] Step 1.1: This invention models fake news detection as a structured generation task. Let the news text be... The corresponding set of user comments is Through predefined prompt template functions Organize news and commentary into input context Target model Target sequence needs to be generated ,in To represent the reasoning process, Indicates the final predicted label, Let the target sequence be... Then its conditional generation probability is defined as ,in For the i-th generated token, Indicates a historical prefix.

[0051] Step 1.2: Transfer the news text Comments Collection and real labels Input the teacher's large language model and generate an inference process consistent with the real labels using a relation-guided prompt template. Thus, the supervised target sequence is obtained. .

[0052] Step 1.3: The relationship-oriented prompt template is used to guide teachers to organize reasoning around the interaction between news and comments using the large language model. Specifically, it includes the following three sub-processes: 1) Comment noise filtering and evidence screening: filtering emotional expressions, repetitive statements, irrelevant chatter, and invalid interactions in the comment set, retaining key comments relevant to the judgment of authenticity; 2) News-comment relationship analysis: analyzing the questioning, correction, supplementation, agreement, or misleading effects of comments on news content to form an evidence chain around the relationship between news and comments; 3) Reasoning generation under label constraints: explicitly injecting real labels into the prompts. This guides teachers to generate reasoning processes consistent with real labels using large language models, thereby improving the consistency and integrity of reasoning trajectories with gold labels.

[0053] Step 1.4: Obtain the supervised training dataset based on the above process. ,in .

[0054] Step 2: Construct a lightweight student model initialization module. Based on the positive sample inference process and real labels, perform supervised fine-tuning initialization of the student model to enable the student model to have preliminary ability to infer the relationship between news and comments and generate real labels.

[0055] Furthermore, Step 2 specifically includes the following:

[0056] Step 2.1: Select the lightweight student model as the target model, and use the positive sample inference trajectory dataset generated based on the teacher's large language model. The student model is subjected to efficient parameter fine-tuning using a low-rank fitting method, and the supervised fine-tuning loss is defined as follows:

[0057]

[0058] Step 2.2: By optimizing the supervised fine-tuning loss, the student model learns the relational reasoning trajectory generated by the teacher's large language model, thereby obtaining the initial "reasoning process generation + label prediction" capability, providing an initial model for subsequent relational perception preference optimization.

[0059] Step 3: Construct a relation-aware preference data construction and joint optimization module. Use the correct reasoning process and real labels generated by the teacher's big language model as positive preference samples, and combine the erroneous output of the student model after supervised fine-tuning with the multi-class relation error negative samples constructed by the big language model to form preference pairs. On this basis, perform direct preference optimization on the student model and train it in conjunction with supervised fine-tuning loss to improve the model's ability to identify and model the relationship between news and comments.

[0060] Furthermore, Step 3 specifically includes the following:

[0061] Step 3.1, for the input sample The correct reasoning process and real labels generated by the teacher's large language model are used as positive preference samples, i.e. ;

[0062] Step 3.2: Constructing model misjudgment negative samples. The supervised, fine-tuned student model is then compared to the input... The generated result is represented as ,when When this is the case, it is used as a model misjudgment negative sample to reflect the true error distribution of the student model on complex samples.

[0063] Step 3.3: Construct Negative Samples with Missing Relationships. Using a large language model, negative samples are generated under constraints, ignoring key interactions between news and comments while maintaining correct labels. This is represented as follows: This type of negative sample is used to constrain the model to distinguish between suboptimal reasoning that is “correctly labeled but has insufficient relationship modeling” and optimal reasoning that is “reasonable in relation and correctly labeled”.

[0064] Step 3.4: Construct noise-biased negative samples. Using a large language model, negative samples are generated that maintain correct labeling while overly relying on comment sentiment, stance conformity, or local high-frequency noise under constrained cues. This is represented as follows: These negative samples are used to suppress the model's over-reliance on comment noise cues.

[0065] Step 3.5: Construct relation-reversal negative samples. Using a large language model, under constraint cues, the logical relationship between news items and comments is intentionally reversed, generating reverse labels consistent with the erroneous relationship. Let the reverse label be denoted as... ,in Then the negative sample of the relation inversion type is represented as These negative samples contain both relation-level and label-level errors, which are used to enhance the model's ability to identify erroneous relation inference paths.

[0066] Step 3.6: For each input sample Constructing a negative sample set And further construct a preference dataset

[0067]

[0068] Each preference sample is derived from the input. Positive samples and a single negative sample This is used for subsequent direct preference optimization training.

[0069] Step 3.7: Based on the supervised fine-tuning of the student model, utilize the preference dataset. Perform direct preference optimization on the model. Let... For the strategy model to be optimized, To freeze the reference model, then for preference samples... The direct preference optimization loss is defined as

[0070]

[0071] in, It is the Sigmoid activation function. This is a hyperparameter used to control the intensity of preference.

[0072] Step 3.8: To avoid the lightweight student model deviating from the original language distribution due to direct preference optimization alone, supervised fine-tuning loss is introduced during the preference optimization stage for joint training. The joint loss function is expressed as follows: .in, The represents the balance coefficient. Through joint optimization, the model simultaneously possesses generative stability and relation discriminative ability.

[0073] Step 4: Perform fake news detection and prediction. Input the news text and comment set in the test sample into the trained student model, output the corresponding reasoning process and authenticity label, and obtain the final fake news detection result.

[0074] Furthermore, Step 4 specifically includes the following:

[0075] Step 4.1: In the prediction phase, for the news text in the test sample... and a collection of comments Construct input context using predefined prompt templates .

[0076] Step 4.2: Set the input context Input the trained student model and generate the output sequence. ,in, The reasoning process generated for the model, This is a label indicating the accuracy of the model's predictions.

[0077] Step 4.3, add the authenticity label As the output of fake news detection results, and the reasoning process is also included. This serves as a corresponding interpretable criterion, thereby enabling fake news detection based on modeling the relationship between news and comments.

[0078] The present invention also provides a fake news detection system based on news comment relationship-aware preference alignment, the system comprising: a module for executing the fake news detection method based on news comment relationship-aware preference alignment.

[0079] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the fake news detection method based on news comment relationship-aware preference alignment.

[0080] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the fake news detection method based on news comment relationship-aware preference alignment.

[0081] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the fake news detection method based on news comment relationship-aware preference alignment.

[0082] Evaluation metrics: This study uses the F1 score as the main evaluation metric. The F1 score combines the model's precision and recall, and can comprehensively measure the model's predictive ability across different categories.

[0083] To verify the effectiveness of the model proposed in this invention, the method of this invention is compared with the following baseline methods of different labeling frameworks, as follows:

[0084] BERT: BERT is a Transformer-based pre-trained model. Its core method utilizes both left and right contexts for deep bidirectional encoding across all layers. Only an output layer needs to be added for fine-tuning, allowing it to adapt to various natural language processing tasks.

[0085] HSA-BLSTM: This model first learns a hierarchical representation of text using a hierarchical bidirectional long short-term memory model, and then incorporates social contextual information (such as user characteristics, propagation structure, etc.) through an attention mechanism, thereby more robustly identifying rumors.

[0086] Dual-Emo: This model proposes a "dual sentiment feature" to describe the relationship between the publisher's sentiment and the social sentiment evoked by public comments. This feature can be easily incorporated into existing detection models, effectively improving the identification of fake news.

[0087] dEFEND: dEFEND utilizes a sentence-comment collaborative attention network to simultaneously analyze news content and user comments, identifying the most crucial sentences from the original news article and comments, thereby determining the veracity of the news and providing explanatory evidence.

[0088] MTEFN: This method proposes a multi-task learning framework that simultaneously predicts the sentiment category and truthfulness of text for fake news and rumor detection. Experiments show that the multi-task model significantly outperforms the single-task model in both intra-domain and cross-domain scenarios, and sentiment can be used as a general feature across datasets.

[0089] While LLM itself performs poorly as a detector, ARG (Alternative Reasoning Model) can generate multi-faceted analytical reasons (such as textual descriptions and common-sense reasoning), serving as an "excellent advisor" to assist Small Language Models (SLMs). The ARG network, through three modules—news-reason interaction, LLM judgment prediction, and reason usefulness evaluation—enables SLMs to adaptively select and integrate reasons provided by LLMs.

[0090] ARG-D: ARG-D is obtained by distilling the ARG version to obtain the lightweight model ARG-D that does not require LLM inference, so as to support low-cost deployment.

[0091] SARC: Enhances fake news detection from the perspective of group behavior patterns by combining comment text and sentiment features with differentiable deep clustering to mine the potential user roles behind the comments.

[0092] Table 2 Comparison of the proposed method with different baseline models

[0093] The proposed method outperformed the comparative methods on both datasets. Specifically, on the RumorEval-19 dataset, the proposed method achieved a Macro-F1 score of 0.445, significantly higher than all baseline methods, indicating better comprehensive discrimination ability in class imbalance scenarios. On the Weibo-comp dataset, the proposed method achieved scores of 0.967, 0.972, 0.961, and 0.966 in Accuracy, Precision, Recall, and F1-score, respectively, all the best results among the comparative methods, demonstrating that the proposed method can more accurately complete the fake news detection task. The superior performance of the proposed method is due to two factors: firstly, the proposed method does not simply utilize comments as shallow auxiliary features, but rather models the relationship between news and comments, enabling the model to more effectively explore the auxiliary role of comments in judging authenticity; secondly, the proposed method further constructs multi-type relation error negative samples and combines them with preference alignment for joint optimization, enabling the model to better distinguish between correct and incorrect relation reasoning, thereby improving detection accuracy and result stability. In summary, the experimental results show that the fake news detection method based on news comment relationship perception preference alignment proposed in this invention has good effectiveness and superiority.

[0094] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A method for detecting fake news based on news comment relationship-aware preference alignment, characterized in that, The method includes the following steps: Step 1: Construct a reasoning trajectory generation module for modeling the relationship between news and comments. Input the news text, comment set, and real labels into the teacher's large language model. Generate a positive sample reasoning process consistent with the real labels through relationship-guided prompts for subsequent student model training. Step 2: Construct a lightweight student model initialization module. Based on the positive sample inference process and real labels, perform supervised fine-tuning initialization of the student model to enable the student model to have preliminary ability to infer the relationship between news and comments and generate real labels. Step 3: Construct a relation-aware preference data construction and joint optimization module. Use the correct reasoning process and real labels generated by the teacher's big language model as positive preference samples. Combine the erroneous output of the student model after supervised fine-tuning with the multi-class relation error negative samples constructed by the big language model to form preference pairs. On this basis, perform direct preference optimization on the student model and train it in conjunction with supervised fine-tuning loss to improve the model's ability to identify and model the relationship between news and comments. Step 4: Perform fake news detection and prediction. Input the news text and comment set in the test sample into the trained student model, output the corresponding reasoning process and authenticity label, and obtain the final fake news detection result.

2. The fake news detection method based on news comment relationship-aware preference alignment according to claim 1, characterized in that: Step 1 specifically includes the following: Step 1.1: Model fake news detection as a structured generation task: Let the news text be... The corresponding set of user comments is Through predefined prompt template functions Organize news and commentary into input context Target model Target sequence needs to be generated ,in To represent the reasoning process, Indicates the final predicted label, ; Let the target sequence be Then its conditional generation probability is defined as ,in For the i-th generated token, Indicates a historical prefix; Step 1.2: Transfer the news text Comments Collection and real labels Input the teacher's large language model and generate an inference process consistent with the real labels using a relation-guided prompt template. Thus, the supervised target sequence is obtained. ; Step 1.3: The relationship-oriented prompt template is used to guide teachers to organize reasoning around the interaction between news and comments using the large language model. Specifically, it includes the following three sub-processes: 1) Comment noise filtering and evidence screening: filtering emotional expressions, repetitive statements, irrelevant chatter, and invalid interactions in the comment set, retaining key comments relevant to the judgment of authenticity; 2) News-comment relationship analysis: analyzing the questioning, correction, supplementation, agreement, or misleading effects of comments on news content to form an evidence chain around the relationship between news and comments; 3) Reasoning generation under label constraints: explicitly injecting real labels into the prompts. This guides teachers to generate reasoning processes consistent with real labels using large language models, thereby improving the consistency and integrity of reasoning trajectories with gold labels. Step 1.4: Obtain the supervised training dataset based on the above process. ,in .

3. The fake news detection method based on news comment relationship-aware preference alignment according to claim 1, characterized in that: Step 2 specifically includes the following: Step 2.1: Select the lightweight student model as the target model, and use the positive sample inference trajectory dataset generated based on the teacher's large language model. The student model is subjected to efficient parameter fine-tuning using a low-rank fitting method. The supervised fine-tuning loss is defined as: ； Step 2.2: By optimizing the supervised fine-tuning loss, the student model learns the relational reasoning trajectory generated by the teacher's large language model, thereby obtaining the initial reasoning process generation and label prediction capabilities, providing an initial model for subsequent relational perception preference optimization.

4. The fake news detection method based on news comment relationship-aware preference alignment according to claim 1, characterized in that: Step 3 specifically includes the following: Step 3.1, for the input sample The correct reasoning process and real labels generated by the teacher's large language model are used as positive preference samples, i.e. ; Step 3.2: Constructing model misclassification negative samples. The supervised, fine-tuned student model is then compared to the input... The generated result is represented as ,when When this is the case, it is used as a model misjudgment negative sample to reflect the true error distribution of the student model on complex samples; Step 3.3: Construct Negative Samples with Missing Relationships. Using a large language model, negative samples are generated under constraints, ignoring key interactions between news and comments but maintaining correct labels. This is represented as follows: This type of negative sample is used to constrain the model to distinguish between suboptimal reasoning that is "correctly labeled but with insufficient relation modeling" and optimal reasoning that is "reasonable in relation and correctly labeled". Step 3.4: Construct noise-biased negative samples. Using a large language model, negative samples are generated that maintain correct labeling while overly relying on comment sentiment, stance conformity, or local high-frequency noise under constrained cues. This is represented as follows: These negative samples are used to suppress the model's over-reliance on comment noise cues; Step 3.5: Construct relation-inverted negative samples. Using a large language model, under constraint cues, the logical relationship between news items and comments is deliberately reversed, generating inverse labels consistent with the erroneous relationship; denoted as [the inverse label is missing here]. ,in Then the negative sample of the relation inversion type is represented as This type of negative sample contains both relation-level and label-level errors, which is used to enhance the model's ability to identify erroneous relation inference paths. Step 3.6: For each input sample Constructing a negative sample set And further construct a preference dataset: ； Each preference sample is derived from the input. Positive samples and a single negative sample This is used for subsequent direct preference optimization training; Step 3.7: Based on the supervised fine-tuning of the student model, utilize the preference dataset. Perform direct preference optimization on the model; let For the strategy model to be optimized, To freeze the reference model, then for preference samples... The direct preference optimization loss is defined as ； in, It is the Sigmoid activation function. The hyperparameters used to control preference intensity; Step 3.8: To avoid the lightweight student model deviating from the original language distribution due to direct preference optimization alone, supervised fine-tuning loss is introduced during the preference optimization stage for joint training. The joint loss function is expressed as follows: .in, The represents the balance coefficient. Through joint optimization, the model simultaneously possesses generative stability and relation discriminative ability.

5. The fake news detection method based on news comment relationship-aware preference alignment according to claim 1, characterized in that: Step 4 specifically includes the following: Step 4.1: In the prediction phase, for the news text in the test sample... and a collection of comments Construct input context using predefined prompt templates ; Step 4.2: Set the input context Input the trained student model and generate the output sequence. ,in, The reasoning process generated for the model, Assign a true label to the model's predictions; Step 4.3, add the authenticity label As the output of fake news detection results, and the reasoning process is also included. This serves as a corresponding interpretable criterion, thereby enabling fake news detection based on news-comment relationship modeling.

6. A fake news detection system based on news comment relationship perception and preference alignment, characterized in that, The system includes a module for performing the fake news detection method based on news comment relationship-aware preference alignment as described in any one of claims 1 to 5.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the fake news detection method based on news comment relationship awareness preference alignment as described in any one of claims 1 to 5.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the fake news detection method based on news comment relationship awareness preference alignment as described in any one of claims 1 to 5.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the fake news detection method based on news comment relationship awareness preference alignment as described in any one of claims 1 to 5.

Images